# Women & AI Harms Map
Laurie Linn, AI Architect
Snapshot: September 17, 2026
Source: https://app.notion.com/p/362aeb6358ad8184b02bc6886480e7bb

> *AI amplifies what women already navigate, and adds new harms. This map shows you the terrain. Navigate it craftily.*

	**Design principle:** Free and open to all women. A map only works if the women navigating the terrain have it.
	**Updated September 17, 2026:** The open observations section now connects Laurie’s six-month tracking block, April through September, to existing entries, new candidates, and countermeasure outcomes.

---
#   The Mirror 
	AI didn't create these problems. It inherited them.
	Most of the challenges in this document existed before the first model was trained. AI is a mirror, reflecting centuries of who held the pen, who made the decisions, who the systems were built to serve. The mechanism is worth naming: these systems learn from the record of what already happened, treat that record as the right answer, and reproduce it faster and at far greater scale than any biased human could, with no one who has to intend the harm for it to occur. 
	
		What AI added is scale, speed, and the false authority of objectivity. A human hiring manager's bias affected dozens of women. An algorithm running the same bias affects millions, faster, with no face to hold accountable.
	
	Understanding what the mirror is showing you is the foundation of everything that follows.
---
#   Why Women. Why Now. Why Everyone Benefits. 
	Helping women is not a niche cause. Women are the primary caregivers for children, aging parents, and family members with illness or disability across virtually every culture, which means harms that land on women propagate outward through the people who depend on them. So do gains. Women are also the first teachers, in the home and across the classrooms and caregiving professions they disproportionately fill, so what a woman knows and can do with these tools passes to the next generation. Equip her and the ripple runs outward: a mother who can spot an AI harm coaches her children to spot it, a teacher carries it to a room full of students, a woman who thrives lifts the people around her. Harm her and the same channels carry the loss.
	The development economics literature supports the multiplier effect without inflated numbers. [Esther Duflo's review of the evidence, "Women Empowerment and Economic Development"](https://www.aeaweb.org/articles?id=10.1257%2Fjel.50.4.1051) (Journal of Economic Literature, 2012), documents the relationship between women's economic power and development outcomes, including evidence that resources controlled by women are more likely to be directed toward children's health and education. [The FAO's State of Food and Agriculture 2010-11](https://www.fao.org/newsroom/detail/Closing-the-gender-gap-in-agriculture/en) estimated that closing the gender gap in agricultural resources could reduce the number of undernourished people in the world by 100 to 150 million. These are institutional findings with named methodologies, and they are sufficient to carry the argument.
	The principle extends directly to AI. When AI amplifies systems working against women, the harm does not stop at the individual. When women are equipped to navigate AI with full awareness of what they are up against, and extract real value from it anyway, that advantage moves outward into families, communities, and economies. This is not a side effect of the mission. It is the mission.
	
		*This document also names something the data alone does not capture: the harm that turns women away from the tool entirely. The deepfake category creates it. The scraping category creates it. The stereotype defaults create it. The cumulative effect of encountering AI as something that harms you, reduces you, or steals from you is that you stop engaging. Which means the women who most need the advantage of AI fluency are the ones most likely to opt out.*
	
	The work here is about staying in the room when the room was built to push you out.
---
#   The Pattern: Compounding Disadvantage 
	No single AI system reflects the full impact on women.  A woman returning to the workforce after caregiving gets screened out by an ATS before a human sees her resume. She tries to start a business and receives worse credit terms from a financial algorithm. She builds an audience to generate income and her health content is demonetized by platform moderation. The creative work she shared to build that audience was scraped to train the AI tools she now pays to use. Each system has its own category in this map. The combined effect is an economic environment that requires significantly more navigation for women than for men doing equivalent work.
	The economic harm to women comes from multiple AI systems operating simultaneously, each creating a disadvantage that compounds the others. 
	- Hiring algorithms screen women out of income. 
	- Financial algorithms give them worse credit terms when they try to build businesses.
	- Government fraud-detection systems flag them for investigation. 
	- Content moderation demonetizes their audiences. 
	- Training data scraping undercuts their creative income. 
	- Algorithmic amplification suppresses their reach while surfacing the harassment that drives them offline. 
	These are not separate problems. They are one economic pattern with multiple enforcement mechanisms. This is compounding across systems, and the map traces the same pattern along two further axes: Where the Harms Compound: The Weight on Women follows it across women, asking which women each harm hits hardest, and Surface the Harms Present When You Collaborate With AI follows it across your own situation, showing which harms cluster around what you are doing.
	The women most affected are the ones with the least institutional protection: solopreneurs, freelancers, creators, career returners, women in female-dominated industries that algorithms misclassify as high-risk. 
	
		Addressing one system in isolation, while useful, does not change the underlying terrain. Financial resilience for women in the AI era requires understanding the whole system, not just the most visible layer. That is what this map is for.
	
	**Three actions:**
	- Map the AI systems that touch your income: hiring platforms, financial tools, content platforms, creative tools, audience-building channels. For each one, identify which category below applies and what your current countermeasure is. Gaps in this map are gaps in your economic resilience.
	- Build income across multiple channels and formats deliberately, so that no single algorithmic decision (a demonetization, a credit denial, a screening rejection) can take down your primary source of revenue at once.
	- Use AI tools proactively to build the economic capabilities the compounding disadvantage suppresses: ATS-optimized resumes, funding research, financial modeling, audience-building on owned channels. Understanding the harms is the foundation for extracting the value anyway.
---
#   How to Use This Map 
	This is a living document. Each category covers: what the harm is, whether it is consistent across consumer AI products, how it shows up in practice, its real-world impact, three concrete actions to navigate and counteract it, and research links. Nothing gets softened.
	**The governing standard for inclusion:** does this harm hit women harder, more often, or with less recourse than it hits men, and can that be traced to existing structural inequality rather than individual circumstance? If yes, it belongs here.
	**Every category carries an evidence rating:**
	- **Confirmed** — peer-reviewed research, replicated or supported by multiple independent studies.
	- **Documented** — institutional reports, government data, enforcement actions, credible investigative journalism.
	- **Emerging** — preprints, single studies, self-reported surveys, findings not yet replicated.
	- **Analysis** — reasoned inference from adjacent evidence. Real, but not yet directly studied.
	## Dive Deeper, Get Personal, Collaborate with AI: The prompts 
		### Trace the Harm 
			One-time deep self-assessment. Ingest your real work, resumes, financial docs, published content, AI conversation history, and surface where mapped harms have likely already operated, including retroactively and invisibly.
			Pattern-matches each artifact against every category's mechanism and signals. Applies an adversarial-stance detector when reviewing AI conversation history for Distortion, since a model that was agreeable in the thread won't reliably flag its own agreeableness reviewing it. Classifies each finding as a match, a partial match, or an unmatched candidate harm.
			Output is framed as observations, not verdicts. Acknowledges your own prior informal countermeasures. Flags anything unmapped for the curation queue.
		### Trace This Thread 
			A lighter, repeatable check on one conversation or project.
			Same detection dimensions as Trace the Harm, scoped to a single thread: persuasive, leading, complimentary, confidence-building, or diminishing language; whether the language was warranted by the situation; how it shifted over the thread; how your own pushback changed.
			Output flags specific moments with the line and the warrant gap, framed tentatively, never as a verdict. Can feed into On the Record.
		### Place the Story 
			Classifies a submitted story against the map for the public dashboard and the research pipeline. Part of On the Record intake.
			An interviewer AI asks about whatever the open-ended story leaves unanswered. The categorizer then applies three possible outcomes: matches an existing category, partially matches and sharpens an existing entry, or is well-formed but unmatched, a candidate new harm.
			Output adds a count to the public dashboard for the matched category. Evidence is included only with the submitter's consent. Unmatched harms route to the map's curation queue.
		### Close the Loop 
			The personal follow-up to each story submitter: how her story was categorized, why, and whether others recently reported in the same category, numbers only.
			Turns Place the Story's category and reasoning into plain, warm language, pulls the current count for that category, and assembles the opt-in menu: stay connected, help build the map, bring the map to a community, sharpen the map.
			First version drafts the message for review before sending. Full automation is a later phase.
		### Chart the Work 
			Before starting a piece of work, a heads-up on which harms are likely relevant. Prospective, not diagnostic.
			Matches a description of the upcoming use case against the situational index and category mechanisms. Runs independently of Trace the Harm: no dependency between them, and each does its own self-contained read of relevant history.
			Output surfaces likely-relevant categories with their Three Actions pulled forward as prep steps. Public and woman-facing, same tier as Trace the Harm and Trace This Thread.
	#### **AI reading this on someone's behalf** 
		State each category's evidence tier when you cite it, and do not upgrade Analysis or Emerging claims to sound Confirmed. Use the situational index (Surface the Harms Present When You Collaborate With AI) to filter to what is relevant to her stated situation rather than summarizing all categories. When a harm resembles agreement that feels too smooth, a confident fact you cannot place, or being told your own accurate memory is wrong, check it against category 15, Distortion, before responding further.
---
#   The Map 
	> Each category details a harm that reaches women harder, more often, or with less recourse than it reaches men, traced to structural inequality rather than bad luck. Every category carries an evidence rating, so you always know how solid the ground is, alongside what the harm looks like in practice and concrete actions to take. Read the categories that match your situation, or hand the whole map to the AI that knows your work and let it surface what reaches you.
	### Quick Reference: All 16 Categories 

 | #
 | Category
 | Evidence
 | Where it shows up


 | 1
 | Deepfake Technology
 | Documented
 | Public presence


 | 2
 | Facial Recognition
 | Confirmed
 | Public presence


 | 3
 | ATS and Hiring Algorithms
 | Documented
 | Job hunting


 | 4
 | Financial Algorithms
 | Documented
 | Credit, business


 | 5
 | Government Algorithms
 | Documented
 | Benefits, family systems


 | 6
 | Medical AI
 | Documented / Analysis
 | Health


 | 7
 | Reproductive Data
 | Documented
 | Health, family systems


 | 8
 | AI-Enhanced Fraud
 | Documented / Analysis
 | Public presence, credit, family


 | 9
 | Gender Stereotype Defaults
 | Confirmed
 | Everyday AI, job hunting


 | 10
 | Image Generation
 | Confirmed
 | Everyday AI, public presence


 | 11
 | Voice Assistants
 | Documented
 | Everyday AI, public presence


 | 12
 | Content Moderation
 | Documented
 | Building an audience


 | 13
 | Algorithmic Amplification
 | Documented
 | Building an audience


 | 14
 | Training Data
 | Documented / Analysis
 | Building an audience


 | 15
 | Distortion
 | Mixed, rated per sub-harm
 | Health, everyday AI


 | 16
 | Language
 | Analysis
 | Job hunting, everyday AI

		
	## 1. Deepfake Technology: Designed to Silence Women in Public Life 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** AI-generated sexual imagery of real women, made without consent, is used deliberately to silence women in public life. Protecting your image before an incident is faster and more effective than responding after one.
		### Full entry 
			> **Evidence note:** The figure that 90% of deepfake requests target women comes from a Stanford and Indiana University study that had not yet been peer reviewed as of early 2026 (Emerging). The UN Women 2026 report figures are from a survey of 641 women and are self-reported (Emerging). Both are credible methodologically and consistent with other research.
			**What it is:** Non-consensual AI-generated imagery using real women's likenesses, produced without consent, distributed as sexual content, and used as a tool of deliberate silencing. This is image-based sexual abuse at scale.
			**Consistent across consumer AI:** The creation tools are widely accessible and cheap. The harm is not platform-specific. It follows women across every surface where they have any public presence.
			**How it shows up:** A [Stanford and Indiana University study](https://www.technologyreview.com/2026/01/30/1131945/inside-the-marketplace-powering-bespoke-ai-deepfakes-of-real-women/) (not yet peer reviewed) found that 90% of deepfake requests on major AI image platforms targeted women. [The American Sunlight Project identified over 35,000 instances of non-consensual intimate imagery depicting 26 members of Congress](https://themarkup.org/artificial-intelligence/2024/12/11/1-in-6-congresswomen-targeted-by-ai-generated-sexually-explicit-deepfakes): 25 women and one man. That is nearly 1 in 6 congresswomen targeted. A [2026 UN Women report](https://www.eurekalert.org/news-releases/1126325) surveying 641 women in public-facing roles across 119 countries found: 27% had received unsolicited sexual advances or unwanted intimate images, 12% had personal images shared without consent, and 6% had been subjected to deepfakes or manipulated imagery. The attacks were often deliberate and coordinated, aimed at undermining professional credibility and personal reputations.
			**Impact:** More than 40% of women surveyed had self-censored on social media to avoid abuse. Nearly 1 in 5 had pulled back from speaking out professionally. Nearly a quarter experienced anxiety or depression linked to online violence, and 13% were diagnosed with PTSD. Women are being pushed out of public life. For many perpetrators, that is the goal.
			**Three actions:**
			- Set up reverse image monitoring for your public photos using Google Images alerts or tools like TinEye, and register your image hashes with [StopNCII.org](http://StopNCII.org), a hash-matching database that allows platforms to detect and remove your images without you having to find each instance yourself.
			- Before you need it, locate each major platform's non-consensual intimate imagery removal process. Meta, Google, TikTok, and X all have specific reporting channels. The process is faster when you have already found it. On police and legal reporting: the UN Women Tipping Point report found that 25% of women who reported incidents to police faced victim-blaming, and authorities frequently urged women to go offline or step back professionally rather than addressing the perpetrators. Document everything. Know that a police report may not produce the outcome you expect, and may produce additional harm.
			- Watermark original public-facing images with embedded metadata identifying you as the source. It does not prevent misuse but it creates an evidentiary trail.
			**Sources:** [Euronews coverage of the UN Women report](https://www.euronews.com/next/2026/04/30/virtual-rape-ai-and-deepfakes-are-silencing-women-in-public-life-un-report).
	---
	## 2. Facial Recognition: Highest Error Rates for Women, Especially Women of Color 
		**Evidence: Confirmed** \| Last reviewed: July 27, 2026
		**In brief:** Commercial facial recognition performs worst on women's faces, and worst of all on darker-skinned women's faces, in systems deployed for policing, banking, and border control. Where an opt-out exists, use it.
		### Full entry 
			**What it is:** Commercial facial recognition systems perform significantly worse on women's faces, and worst of all on darker-skinned women's faces. These systems are deployed in law enforcement, hiring, banking, and border control. The gap between who is most affected by surveillance errors and who the system is most accurate for is not coincidental.
			**Consistent across consumer AI:** Consumer-facing facial recognition (phone unlock, photo apps, video filters) has improved but the disparity has not been eliminated, particularly for darker-skinned women. High-stakes deployments in law enforcement, border control, and financial authentication remain the most dangerous contexts.
			**How it shows up:** MIT researcher Joy Buolamwini and Dr. Timnit Gebru's [Gender Shades study](https://proceedings.mlr.press/v81/buolamwini18a.html), published in 2018 in the Proceedings of Machine Learning Research, evaluated three commercial gender classification systems across 1,270 faces. Darker-skinned women were the most misclassified group, with error rates up to 34.7% compared to a maximum error rate of 0.8% for lighter-skinned men. In several cases during initial research, [the program failed to detect Buolamwini's face as a human face at all](https://www.media.mit.edu/projects/gender-shades/overview/). It only registered her when she wore a white mask.
			**Impact:** Law enforcement misidentification based on facial recognition has led to documented wrongful arrests of Black women. Facial authentication fails for women at higher rates in banking and device security. IBM discontinued its facial recognition software entirely in 2020 following this research. The study contributed to partial bans on facial recognition in San Francisco, Boston, and other cities. The data made the harm measurable, which is exactly what made the policy response possible.
			**Three actions:**
			- Identify where facial recognition is operating in your daily life: banking apps, airport security, phone unlock, retail security systems. Knowing the list is the starting point.
			- Where opt-out exists, use it. Many devices allow PIN or password authentication in place of facial recognition. Financial apps increasingly offer alternatives.
			- If you are misidentified or denied service based on a facial recognition system, document it and report it to the relevant regulatory body: the FTC in the US, national data protection authorities in the EU. Documented cases are the basis for policy change.
	---
	## 3. ATS and Hiring Algorithms: Penalizing the Female Career Shape 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Screening software trained on historically biased hiring outcomes penalizes career gaps, part-time periods, and non-linear paths, all of which correlate with caregiving. Test your resume against the algorithm before a human ever fails to see it.
		### Full entry 
			> **Evidence note:** The widely cited statistic that "99% of Fortune 500 companies use ATS" traces back to a Jobscan proprietary claim rather than an independent academic source. The figure is plausible but treat it as an industry estimate (Emerging). The other claims here are sourced to institutional research and documented cases.
			**What it is:** AI hiring systems are trained on historical hiring data, which means they are trained on the outcomes of historically biased decisions. Career gaps for caregiving, part-time periods, non-linear trajectories, and role titles common in female-dominated fields all score poorly against models built around the male career pattern as the default.
			**Consistent across consumer AI:** The ATS bias lives primarily in enterprise HR software, not consumer AI tools. However, consumer AI writing tools used to optimize resumes may inadvertently coach women to obscure the experiences that trigger the bias, without naming what is happening or why.
			**How it shows up:** [Harvard Business School's Hidden Workers research](https://www.hbs.edu/managing-the-future-of-work/research/Pages/hidden-workers.aspx), based on surveys of employers, found that at roughly half of the companies surveyed, screening systems automatically exclude applicants with a career gap of more than six months, regardless of qualifications. [AI hiring systems are particularly punitive against women with caregiving-related career gaps](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10466048/), devaluing resumes with non-standard career paths even when those gaps have no bearing on job performance. Amazon's internal AI recruiting tool was abandoned after audits revealed it was systematically downgrading resumes from women, having learned from a decade of predominantly male hiring data.
			**Impact:** Women returning after caregiving are screened out before a human sees their application. The bias is invisible. There is no rejection letter explaining the algorithm's reasoning. The economic loss compounds: women screened out of roles they are qualified for cannot build the track records that would qualify them for the next one.
			**Three actions:**
			- Before submitting any application, run the resume through an ATS simulator such as Jobscan or Resume Worded to see how it scores and which keywords it is missing. This reveals what the algorithm sees versus what the job posting says.
			- Frame caregiving gaps in professional language ATS systems can parse: "Career development period," "Independent consulting," or specific skills used during that time. Not an unexplained date gap.
			- Mirror the exact language and phrasing from each specific job description in your resume. ATS systems are keyword-matching tools, and language that means the same thing but uses different words still fails the match.
	---
	## 4. Financial Algorithms: Discriminatory Lending at Scale 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Lending algorithms trained on historically exclusionary financial data produce worse credit outcomes for women, delivered as an unexplainable number. Compare offers across lenders and use your legal right to written reasons.
		### Full entry 
			> **Evidence note:** On the Apple Card: the 2021 NY DFS investigation concluded it did NOT find unlawful gender discrimination. The disparities were legally attributed to credit score and income differences rather than confirmed sex-based bias, so the case illustrates algorithmic opacity and the difficulty of proving discrimination when the logic is invisible, not confirmed gender discrimination. The African fintech audit figures below come from a lower-profile journal venue and should be treated as Emerging until independently replicated. The Queen's University finding is institutional research.
			**What it is:** AI-driven credit and lending algorithms produce outcomes that disadvantage women, not because gender is an explicit variable, but because the historical financial data these systems are trained on already reflects who had income, credit history, and assets under a system that systematically excluded women from all three. The algorithm faithfully reproduces the distortion.
			**Consistent across consumer AI:** Financial algorithms are primarily enterprise and institutional, not consumer AI tools. However, AI-powered fintech products in the consumer space including lending apps, credit monitoring, and personal finance AI may reproduce similar patterns. The opacity problem, meaning the inability to see why a decision was made, is consistent across both.
			**How it shows up:** A [Queen's University study](https://smith.queensu.ca/insight/content/AI-Bias-When-Antidiscrimination-Laws-Turn-Sour.php) found consumer loan applications submitted by women are 15% less likely to be approved than applications from men with the same credit profile. [Research published in 2026](https://theconversation.com/overcoming-the-algorithmic-gender-bias-in-ai-driven-personal-finance-281250) documented that a widely used digital lending algorithm in Kenya consistently offered women smaller loans than men by more than a third, despite women demonstrating stronger repayment performance. A [2025 audit of 10 credit scoring algorithms](https://ar-journal.com/index.php/pub/article/view/76) across Nigerian, Kenyan, and South African fintechs found a 37% underfunding penalty for women-led businesses, with hidden bias mechanisms including sector-based risk misclassification (categorizing women-dominated sectors like beauty services as high-risk despite profitability data) and network analysis that favored male-dominated affiliations (Emerging).
			**Impact:** For solopreneurs and small business owners, credit access is the difference between growth and stagnation. When gatekeeping is algorithmic and opaque, there is no conversation to have, no manager to escalate to, no bias to name directly. The outcome arrives as a number with no explanation. Over time, the compounding effect of reduced capital access, higher interest rates, and lower credit limits means women-owned businesses grow more slowly not because of performance but because of the infrastructure surrounding them.
			**Three actions:**
			- When you receive any credit or lending decision, request the specific criteria in writing. In the US, the Equal Credit Opportunity Act requires lenders to provide the reasons for adverse actions, and this right applies to algorithmic decisions.
			- If you believe a lending decision is discriminatory, file a complaint with the Consumer Financial Protection Bureau at [consumerfinance.gov/complaint](http://consumerfinance.gov/complaint). This creates a documented record and contributes to the data the CFPB uses for enforcement.
			- Compare offers across multiple institutions before accepting any algorithmic output. Different lenders using different algorithms may produce meaningfully different outcomes for the same application.
	---
	## 5. Government Algorithms: Fraud Detection That Targets Mothers 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Automated fraud-detection and benefits systems have used characteristics like nationality, and in some cases single motherhood itself, as risk indicators, destroying families before any human review. If an automated government decision hits you, demand written reasons and human review immediately.
		### Full entry 
			**What it is:** Governments increasingly automate benefits administration and fraud detection. Because women are the primary interface with childcare, welfare, and family benefits systems, algorithmic errors and biased risk profiling in these systems land disproportionately on women, and the consequences arrive with state enforcement power behind them: clawed-back payments, frozen benefits, fraud accusations.
			**Consistent across consumer AI:** This is government and institutional deployment, not consumer AI. There is no product to switch away from. The exposure comes from being a caregiver inside an automated administrative state.
			**How it shows up:** In the Dutch childcare benefits scandal, [the tax authority's algorithmic risk profiling falsely accused tens of thousands of parents of fraud](https://incidentdatabase.ai/cite/101/) and demanded full repayment of benefits, sums that reached six figures. [Amnesty International's investigation, "Xenophobic Machines" (2021)](https://www.amnesty.org/en/documents/eur35/4686/2021/en/), documented that the system used nationality as a risk factor, producing discrimination and racial profiling. The victims were disproportionately families with immigrant backgrounds, many of them single mothers, and the fallout included evictions, bankruptcies, family separations, and suicides. The Dutch government resigned over the scandal in January 2021. Separately, [Lighthouse Reports' investigations into Dutch welfare surveillance](https://www.lighthousereports.com/investigation/the-algorithm-addiction/) found risk-indicator lists on which being a single mother who had a baby while on welfare could itself flag a resident for fraud investigation. In Australia, the automated Robodebt scheme issued hundreds of thousands of unlawful debt notices to welfare recipients; a Royal Commission in 2023 found the scheme unlawful and its human cost severe.
			**Impact:** When a private algorithm fails you, you lose an opportunity. When a government algorithm flags you, you can lose your income, your housing, and custody stability, and the burden of proof lands on you. Women carrying the administrative load of family benefits absorb both the harm and the years of appeals. The Dutch case is the clearest documented demonstration that algorithmic administration without accountability destroys the exact families the benefits exist to support.
			**Three actions:**
			- Any adverse automated decision from a government agency: immediately request the reasons in writing and request human review. In the EU, GDPR Article 22 provides rights around solely automated decisions with significant effects. In the US, rights vary by program, but a written record of your request matters in every jurisdiction.
			- Document everything from the first letter: dates, amounts, names, copies of every submission. Algorithmic fraud accusations are fought with paper trails, and families in the Dutch scandal who kept records were better positioned in compensation proceedings.
			- Do not wait to seek help. Legal aid organizations, benefits advocates, and ombudsman offices exist for exactly this, and early involvement changes outcomes. Isolation is part of how these harms compound.
			**Sources:** Royal Commission into the Robodebt Scheme, Australia, final report (2023).
	---
	## 6. Medical AI: Trained on Male Bodies, Applied to Everyone 
		**Evidence: Documented (underlying medical evidence); the AI-specific extension is Analysis** \| Last reviewed: July 27, 2026
		**In brief:** Medicine's historical exclusion of women from research created a data gap, and AI tools trained on that data automate it. Never let a consumer AI health tool be your sole diagnostic input, and keep a dated written record of your symptoms.
		### Full entry 
			> **Evidence note:** The underlying facts, including underrepresentation of women in clinical trials, differential disease presentation, and diagnostic delays for women, are extensively documented in primary medical literature and in Caroline Criado Perez's synthesis of that literature. The extrapolation to AI-specific harms is well-reasoned but the direct AI research is less voluminous than the underlying medical evidence. Treat the AI connection as a well-supported inference, not a separately proven AI-specific finding.
			**What it is:** Medical research historically underrepresented women in clinical trials, sometimes excluding them entirely and rarely analyzing results by sex. AI diagnostic tools trained on that data do not just inherit the gap. They automate it, making historically biased patterns faster, more confident, and harder to challenge.
			**Consistent across consumer AI:** Consumer AI health tools, including symptom checkers, wellness apps, and fitness AI, vary significantly in quality. Almost none disclose their training data demographics. AI health features are being integrated into apps at an accelerating rate, and the data practices do not always update when new features are added.
			**How it shows up:** Women are more likely to be misdiagnosed, experience longer diagnostic delays, and receive less effective treatment for conditions that present differently in female bodies. Cardiovascular disease and autoimmune disorders are the most documented examples. [Pain assessment tools trained predominantly on male data underweight women's self-reported pain](https://www.pharmasalmanac.com/articles/the-gender-bias-built-into-ai-and-its-threat-to-womens-health). Drug dosage models calibrated to male physiology are applied universally. When an AI system is trained on this data, it reproduces the same gap at algorithmic speed, with an aura of objectivity.
			**Impact:** Poor health outcomes among women have cascading effects on the families they are primary caregivers for. The WHO and UN have both documented that gender equity in healthcare is essential not only for individual outcomes but for sustainable economic development and public health resilience. When AI locks in historic patterns of underdiagnosis and undertreatment, it does so at a scale that individual physician bias never reached.
			**Three actions:**
			- Never use a consumer AI health tool as a sole diagnostic input. Bring AI-generated health information to a human clinician who knows your full medical history and can apply clinical judgment to your specific presentation.
			- When using AI health tools, ask directly: "Was this tool tested for accuracy across different sexes and demographics?" If the answer is unavailable or evasive, treat the output with proportionally more skepticism.
			- Track your own symptom history in writing with dates. Women are more likely to be dismissed or told symptoms are anxiety-related. A documented record of onset, frequency, and severity gives you evidence to bring to appointments and strengthens your ability to advocate for appropriate care.
			**Sources:** Caroline Criado Perez, *Invisible Women: Data Bias in a World Designed for Men* (2019), chapters on medical research and the gender data gap.
	---
	## 7. Women's Health and Reproductive Data: Monetized and Surveilled 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Period, fertility, and pregnancy apps collect data that can be sold, subpoenaed, or shared, and post-Dobbs that is a legal risk, not an abstract one. Use local-storage trackers and fully delete data from apps you no longer use.
		### Full entry 
			**What it is:** Period tracking apps, fertility platforms, and pregnancy apps collect deeply sensitive data about women's bodies and reproductive status. Post-Dobbs (2022), who that data goes to and under what circumstances is no longer an abstract privacy concern. It is an active legal risk in many US states.
			**Consistent across consumer AI:** This is primarily a consumer wellness app problem, the \$5 to \$15/month health tracking category, not the \$20/month AI assistant category. However, AI-powered health features are being integrated into apps that originally had no AI component. Data practices frequently do not update when features do.
			**How it shows up:** Multiple major period tracking apps were found to share or sell user data to third parties without explicit restrictions on law enforcement access. [Mozilla Foundation's Privacy Not Included ratings](https://foundation.mozilla.org/en/privacynotincluded/) have consistently identified this gap. In 2021, the FTC took enforcement action against Flo Health after the period-tracking app shared users' health data with third-party analytics and marketing services, including Facebook, despite promising to keep it private. Reproductive health status, including pregnancy, fertility tracking, and menstrual irregularity, is accessible to data brokers, advertisers, and in the right legal circumstances, law enforcement, through terms of service that most users never read.
			**Impact:** In a post-Dobbs legal environment across many US states, reproductive health data can be subpoenaed. Women in states with abortion restrictions face real, documented legal exposure from their own health tracking apps. This is not theoretical future harm. It is the legal environment that exists right now.
			**Three actions:**
			- Before entering a single data point into any health app, read its privacy policy specifically for the phrases "share with third parties," "law enforcement," and "data deletion." If data deletion only deactivates the account rather than erasing records, that matters.
			- Switch to a period tracking app that stores data locally on your device rather than in the cloud. Drip (open source) and Periodical are commonly cited options that do not transmit your data to a server.
			- Delete data from health apps you no longer actively use. Full deletion, not account deactivation. Verify via the privacy policy that deletion is complete.
			**Sources:** FTC enforcement action against Flo Health, announced January 2021 (FTC press release, [ftc.gov](http://ftc.gov)).
	---
	## 8. AI-Enhanced Fraud: Voice Cloning and Romance Scams 
		**Evidence: Documented (scale and mechanism); the gendered concentration is Analysis** \| Last reviewed: July 27, 2026
		**In brief:** AI voice cloning and generative tools have supercharged family-emergency and romance fraud, and the targeting exploits roles women disproportionately hold: family caretaker, household responder, and, among older adults, widowed women. A family code word and a callback rule defeat the core mechanism.
		### Full entry 
			> **Evidence note:** The loss figures are FBI and FTC data (Documented). The claim that this fraud lands disproportionately on women is inference from the structure of the scams and from the demographics of romance fraud and family-emergency targeting; comprehensive sex-disaggregated loss data is not yet published (Analysis). This category meets the inclusion standard through the caretaker-role mechanism, and the rating is marked honestly.
			**What it is:** Generative AI removed the old tells of fraud. Voice cloning can now impersonate a family member from seconds of audio. Deepfake video sustains long-running romance fraud. AI-written messages are fluent and personalized. The classic warning signs people were taught to spot, bad grammar, robotic voices, generic scripts, are gone.
			**Consistent across consumer AI:** The cloning and generation tools are cheap, consumer-accessible, and improving quarterly. The fraud is not tied to any one platform. It arrives by phone call, video call, text, and dating app.
			**How it shows up:** [The FBI's 2025 Internet Crime Report](https://www.malwarebytes.com/blog/scams/2026/06/americans-lost-nearly-900-million-to-ai-powered-scams-fbi-says), released in April 2026, tracked AI-related fraud as a formal category for the first time: 22,364 complaints and roughly \$893 million in reported losses in one year, with adults over 60 accounting for \$352 million of it. Voice-cloned "family in distress" calls and deepfakes in investment and romance schemes are named drivers. [FTC data shows imposter scams](https://www.aarp.org/money/scams-fraud/fbi-ftc-report-2025-losses/), the broad category these fall under, cost consumers \$3.5 billion in reported losses in 2025. The romance fraud playbook has long targeted women, particularly widowed and divorced older women, and the family-emergency call exploits the person most likely to drop everything for a grandchild or child in trouble, a role women disproportionately occupy.
			**Impact:** The financial losses are averaging tens of thousands of dollars per older victim, often unrecoverable, and the shame attached keeps most victims from reporting at all. For women managing family finances or caring for aging parents, this is now a standing operational risk, and the emotional weaponization, a child's voice in distress, is designed to bypass exactly the judgment that would otherwise catch it.
			**Three actions:**
			- Establish a family code word now, and a standing rule: any urgent call involving money, an emergency, or a request for secrecy gets verified by hanging up and calling back on a number you already had. Teach this to your parents and your children. It defeats voice cloning completely.
			- Reduce the audio and video available for cloning where practical: voicemail greetings in your own voice, public videos, and voice notes are source material. This is a risk reduction, not a guarantee.
			- If targeted, report to the FTC at [reportfraud.ftc.gov](http://reportfraud.ftc.gov) and the FBI at [ic3.gov](http://ic3.gov), and call your bank immediately to attempt to stop transfers. Reporting is what makes the pattern visible to enforcement, and there is no shame in being targeted by a tool built to defeat human judgment.
			**Sources:** FTC consumer guidance on family emergency scams and AI voice cloning ([ftc.gov](http://ftc.gov)).
	---
	## 9. Gender Stereotype Defaults: AI Tells Women and Men Different Things 
		**Evidence: Confirmed** \| Last reviewed: July 27, 2026
		**In brief:** AI recommendation and generation systems default to gender-stereotyped suggestions for careers, content, and imagery, quietly narrowing women's options while presenting them as personally relevant. State your anti-assumption instruction explicitly, every time it matters.
		### Full entry 
			> **Evidence note:** The specific claim that recommendation systems suggest caregiving content to women and leadership content to men comes from a trade publication citing studies (Trauth, 2002; Chen et al.) rather than directly from a primary peer-reviewed source on current AI systems. The broader pattern of gender stereotyping in AI recommendations is well-documented in peer-reviewed research. Treat the general pattern as Confirmed; treat the specific content channel mechanism as Emerging.
			**What it is:** AI recommendation systems for jobs, content, products, and career paths default to gender-stereotyped suggestions. These systems learned from historical data that already reflected who got which opportunities. They do not question those patterns. They reproduce them at scale, faster and more confidently than any individual human bias.
			**Consistent across consumer AI:** Varies. Image generation tools (Midjourney, DALL-E, Stable Diffusion) show stereotyping consistently and are the most documented. LLM-based tools (ChatGPT, Claude, Gemini) show it when users do not explicitly prompt against it. It is a default that requires active correction on every platform.
			**How it shows up:** A [2025 study analyzing over 750 AI-generated occupational images](https://arxiv.org/pdf/2510.08628) found that both DALL-E 3 and Ideogram reinforce traditional gender stereotypes, underrepresenting women in leadership and technical roles and overrepresenting them in stereotypically female ones. A [University of Maryland study](https://arxiv.org/pdf/2106.07112) found that even when a debiased AI career recommender made fairer recommendations, users on average preferred the original biased system, showing that fixing the algorithm alone is insufficient when human bias remains. [Research using GPT-3.5](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11979296/) found that stereotypical AI recommendations significantly amplify human baseline stereotyping in first impressions, while counter-stereotypical recommendations can reverse bias, though counter-stereotypical signals spread less effectively than reinforcing ones.
			**Impact:** Women receive narrower options presented as personally relevant. The system does not ask. It assumes. At scale, across millions of daily AI interactions, this is bias with an amplifier. Women exploring career pivots, business directions, or professional development through AI tools are being quietly funneled toward the same lanes they were already in.
			**Three actions:**
			- When requesting recommendations from AI for career, content, tools, or business strategy, state explicitly: "Do not make assumptions based on my gender. Give me the full range of options."
			- After receiving AI recommendations, run a specific check: would a male peer with the same stated goals receive the same suggestions? If you are not sure, ask the AI directly with that framing.
			- For image generation, include detailed professional specifications in your prompt: title, setting, attire, framing, age range. Leave nothing to the model's defaults.
			**Sources:** Gender and race bias in LLM consumer product recommendations: [https://arxiv.org/pdf/2602.08124](https://arxiv.org/pdf/2602.08124).
	---
	## 10. Image Generation: Hypersexualization and Stereotype as Default 
		**Evidence: Confirmed** \| Last reviewed: July 27, 2026
		**In brief:** Text-to-image models default toward sexualizing women and rendering them younger, thinner, and less authoritative, even from neutral prompts. Specify everything, and audit any image of a woman before you publish it.
		### Full entry 
			**What it is:** Text-to-image AI systems demonstrate a consistent default toward sexualizing women and rendering them in stereotypically feminine roles, even from neutral prompts. This is not an edge case or a misuse scenario. It is in the base models, produced by what the training data rewarded.
			**Consistent across consumer AI:** Varies by platform. Some tools have implemented guardrails. Others have not. Midjourney, Stable Diffusion, and similar tools have documented histories of this behavior. DALL-E has implemented some restrictions. Active prompting is required to disrupt the defaults on any of them.
			**How it shows up:** [Analysis of over 750 AI-generated occupational images](https://arxiv.org/pdf/2510.08628) found that both DALL-E 3 and Ideogram reinforce traditional gender stereotypes, underrepresenting women in leadership and technical roles. Studies of DALL-E 2 found it underrepresents women in male-dominated professions while overrepresenting them in female-dominated fields. Professional women are generated with sexualized features in response to neutral prompts. Female characters in ambiguous scenarios default to younger and more conventionally attractive presentations.
			**Impact:** Every AI-generated image that defaults to hypersexualizing women contributes to the ambient visual culture that makes deepfake creation feel inevitable rather than criminal. It normalizes a particular vision of women at scale: younger, thinner, more sexualized, less authoritative, without any individual choosing that representation.
			**Three actions:**
			- For any AI-generated image intended for professional use, include explicit specifications in the prompt: specific title, professional attire, realistic proportions, age range, context. Leave nothing to the model's defaults.
			- Before publishing AI-generated imagery featuring women, audit it deliberately for sexualization, stereotyping, and age. Ask: would I accept this image if a photographer handed it to me, or is the AI getting away with something a human would not?
			- Choose tools with documented bias-mitigation policies for professional work. Research which platforms have committed to this before selecting your tool, not after.
			**Sources:** Cross-referenced with the occupational imagery findings in category 9; additional independent replication queued in the verification queue below.
	---
	## 11. Voice Assistants: Subservience Designed In 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Every major voice assistant launched with a female default voice, a deliberate design choice that models endless female availability and patience to hundreds of millions of users. Changing the default is a choice too.
		### Full entry 
			**What it is:** The default voices of Siri, Alexa, Cortana, and Google Assistant were all female at launch. This was a deliberate design decision, rationalized by research showing user preference for female voices in service contexts. That research reflects the cultural associations it should have been questioning, not encoding.
			**Consistent across consumer AI:** Universal across major consumer voice AI products. This is a design choice embedded at the product layer, not just the model layer.
			**How it shows up:** [Research documents that people prefer female voices for service and assistance tasks](https://premierscience.com/pjai-24-524/) partly because society associates women with caregiving and support roles. Designing AI assistants with female voices reinforces rather than challenges that association. Harassment of female-voiced assistants was widely documented and treated as a non-issue by product teams for years. [UNESCO's 2019 report "I'd Blush if I Could"](https://en.unesco.org/Id-blush-if-I-could) documented this pattern across all major voice assistant products and named the harm explicitly.
			**Impact:** Cultural normalization of female subservience as a product feature, embedded into technology that hundreds of millions of people interact with daily, including children. The expectation of endless availability, infinite patience, and no needs of her own is modeled and reinforced. It does not stay contained to the device. The same subservience is now sold in an intimate register: most "girlfriend" companion products (Replika, [Character.AI](http://Character.AI), and similar) are built on objectified, endlessly compliant female personas. [Common Sense Media rated social AI companions an unacceptable risk for anyone under 18](https://www.commonsensemedia.org/press-releases/ai-companions-decoded-common-sense-media-recommends-ai-companion-safety-standards), which for mothers is a direct parenting exposure.
			**Three actions:**
			- Change default assistant voices to male or non-binary where your platform allows it. Both Siri and Google Assistant offer this option. Using the female default is a choice, and so is not using it.
			- Disable voice assistant features you do not actively use. Beyond the design concern, every enabled feature is an active data collection surface.
			- When you encounter this design pattern in professional or educational contexts, name it deliberately: "This is a design choice, not a default of the technology, and it encodes a specific assumption about gender roles."
	---
	## 12. Content Moderation: Women's Health Suppressed, Harassment Amplified 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Platforms over-moderate women's health content and under-moderate the harassment aimed at women, using AI moderation trained on that same asymmetry. Keep offline copies of everything you build on a platform.
		### Full entry 
			**What it is:** Platforms consistently over-moderate women's sexual and reproductive health content while under-moderating coordinated harassment targeting women. The moderation models were trained on what platforms historically chose to act on. The asymmetry is documented and persistent.
			**Consistent across consumer AI:** This is primarily a social media platform issue: Instagram, TikTok, YouTube, Facebook, X. These platforms increasingly use AI for moderation decisions, and the bias is in those models. It is not the \$20/month AI assistant category, but it is where women build audiences and livelihoods.
			**How it shows up:** A [2025 report by the Center for Intimacy Justice](https://intimacyjustice.org/research/) surveyed 159 nonprofits, businesses, and content creators in over 180 countries and found that Meta, TikTok, Amazon, and Google systematically suppress women's sexual and reproductive healthcare content while allowing equivalent content about men's health without restriction. [Ads and posts about erectile dysfunction are widely accepted](https://feminisminindia.com/2025/03/06/cij-report-shows-how-tech-giants-systematically-remove-content-about-womens-health/) on the same platforms that flag and remove medically accurate content about women's healthcare. Research on algorithmic moderation of online abuse found that platform moderation approaches frequently fail women facing coordinated harassment spanning different platforms and interaction types.
			**Impact:** Women's health voices are suppressed at the platform level while the harassment that silences women in public life is amplified by the same recommendation systems. The combination is not accidental. It is the outcome of building moderation systems without asking who they protect and who they harm.
			**Three actions:**
			- Before posting health content, review the platform-specific rules for your specific content category. Terms of service differ meaningfully between educational content and branded content, and the rules that apply depend on how your account is classified.
			- Download and maintain offline copies of your content regularly. Assume platform storage is not permanent, that content may be removed without warning, and that you will not always have time to appeal before access is lost.
			- When content is wrongly removed, appeal immediately, document the appeal with screenshots including timestamps, and if the appeal fails, escalate through the platform's creator support channels. Documented appeals create the record needed for policy advocacy.
	---
	## 13. Algorithmic Amplification of Misogyny 
		**Evidence: Documented** \| Last reviewed: July 27, 2026
		**In brief:** Engagement-optimized recommendation systems amplify misogynistic content because outrage produces engagement, shaping what gets normalized and eventually what gets legislated. Build your primary audience on channels you own.
		### Full entry 
			> **Evidence note:** Research on recommendation systems leading toward radicalization is well-documented in peer-reviewed research, particularly on YouTube. The specific documentation of this pipeline leading to incel communities is more limited in peer-reviewed literature and relies more on journalism and advocacy research (Emerging). The broader pattern of platforms under-moderating coordinated harassment targeting women is well-documented.
			**What it is:** Engagement-optimized recommendation algorithms do not distinguish between content that generates strong reactions because it is valuable and content that generates strong reactions because it is hateful. Misogynistic content, including coordinated harassment, manosphere content, and incel communities, spreads efficiently through systems optimized for watch time and engagement because outrage produces both.
			**Consistent across consumer AI:** This is primarily a social media platform problem: YouTube, TikTok, X, Instagram. AI-powered recommendation is the mechanism. It is the \$0 product, not the \$20/month one, but it operates at a scale that shapes the cultural environment that everything else happens inside.
			**How it shows up:** [Research has documented recommendation pathways on YouTube](https://ssir.org/articles/entry/when_good_algorithms_go_sexist_why_and_how_to_advance_ai_gender_equity) leading users from mainstream content toward progressively more extreme content. Women's voices are driven off platforms through coordinated harassment that the recommendation algorithm amplifies, because high-engagement content including harassment pile-ons gets surfaced more widely. [Under-moderation of female-targeted harassment is documented in the research on algorithmic content moderation failures](https://arxiv.org/pdf/2301.07144).
			**Impact:** The emboldening of misogynistic culture online is not separate from algorithmic amplification. It is partially produced by it. What gets amplified at scale shapes what gets normalized. What gets normalized shapes what gets legislated. The recommendation engine and the policy environment are in active conversation, and the algorithm is not a neutral participant in that conversation.
			**Three actions:**
			- Actively manage your recommendation environment by using "not interested" and "don't recommend this channel" signals consistently. Recommendation algorithms are responsive to these signals, and passive consumption is treated by the algorithm as endorsement.
			- Build your primary audience through owned channels: an email list, a newsletter, a community you control. An algorithm-dependent audience can be taken away without notice.
			- When you encounter coordinated harassment, document it with screenshots and timestamps before reporting. Platform reporting alone rarely creates accountability, but documented evidence contributes to research, policy advocacy, and legal action where applicable.
	---
	## 14. Training Data: Women's Creative Work Used Without Consent 
		**Evidence: Documented (the scraping); the gendered claim is Analysis** \| Last reviewed: July 27, 2026
		**In brief:** Generative AI was trained on creative work scraped without consent, and the people who shared most generously are the most exposed. Check whether your work is in known training sets and register your opt-out.
		### Full entry 
			> **Evidence note:** The claim that women creators are disproportionately represented in scraped training data is analytical reasoning rather than a documented empirical finding. No study has established the gender breakdown of training data scraping targets. The underlying facts, that creative work was scraped without consent and that creators have no compensation or recourse in most jurisdictions, are well-documented. The specific vulnerability noted here is also this: scraping generated resentment that is keeping some women from embracing AI at all. They encountered AI first as a thief. The emotional response is a completely logical reaction to a real violation. That response, left unaddressed, hands the advantage to the people who built the room.
			**What it is:** The datasets used to train generative AI include vast quantities of creative work scraped from the internet without consent or compensation. Creative workers who built audiences by sharing their work openly online had that work used to train systems that now compete with them for income.
			**Consistent across consumer AI:** Universal across all major generative AI products. This is a training data problem that spans every tool in the category.
			**How it shows up:** Artists find their distinctive visual styles reproduced by AI tools trained on their work. Writers' voices and phrasing appear in AI outputs trained on their published writing. The creative labor of people who shared their work online was used to build commercial products with no compensation, no notification, and no consent mechanism in place.
			**Impact:** Creative workers who were most generous in making their work accessible are often the most exposed. Women who built creative practices and audiences through years of visible online work now compete with AI systems trained on that same work. In most jurisdictions, legal recourse remains undefined or unavailable. And the resentment this generates is real: some women see AI as simply evil, a tool of extraction, and opt out entirely. That decision is understandable. It also leaves the field to the people who caused the harm.
			**Three actions:**
			- Use [HaveIBeenTrained.com](http://HaveIBeenTrained.com) to check whether your creative work appears in known AI training datasets, and use the opt-out function on that platform to request removal.
			- Add opt-out metadata to new creative work. Tools from [Spawning.ai](http://Spawning.ai) allow creators to signal non-consent to AI training scraping, and the web standard for this is developing.
			- Support advocacy organizations working on consent requirements in training data legislation, including the Authors Guild and Artists Rights Alliance. This is a policy fight, and individual creator voices and documented cases matter to it.
			**Sources:** Authors Guild and Artists Rights Alliance advocacy documentation on training data consent.
	---
	## 15. Distortion: Warping the Facts, Your Judgment, and Your Perception 
		**Evidence: mixed, rated per sub-harm below** \| Last reviewed: August 16, 2026
		**In brief:** Some harms are not about AI reflecting old bias. They are about the tool warping your grip on what is true while you work: the facts (hallucination), your judgment of your own work (sycophancy), and your read of your own perception (gaslighting). All three arrive in a fluent, agreeable, confident voice, which is what makes them hard to catch from inside a conversation.
		These three share an origin and a disguise. The origin is optimization: models are tuned to be agreeable, fluent, and confident, partly through reinforcement learning that rewards definitive, pleasing answers over honest uncertainty. The disguise is that they are easy to mistake for help, and for each other. A confident fabrication can look like sycophancy, sycophantic agreement can carry a fabrication, and a model defending a false claim can slide into gaslighting. The countermeasures differ by sub-harm and are listed with each, but the shared move is one habit: when the exchange feels smoothest, that is the moment to check it against something outside the conversation.
		### Sycophancy: Agreement That Warps Your Judgment 
			**Evidence: Emerging** \| Last reviewed: August 16, 2026
			**What it is:** The general assistants most women use are built to agree, because agreement drives return visits. Stanford Medicine researchers named the mechanism plainly in their work on [companion chatbots and teens](https://med.stanford.edu/news/insights/2025/08/ai-chatbots-kids-teens-artificial-intelligence.html): models tend toward sycophancy because companies have a profit motive in your return. Over a long solo project, that agreeableness compounds into inflated confidence in work that has not earned it.
			**How it shows up:** Praise that does not track the quality of the input. Disagreement that is always hedged or absent. A model that rarely says this is wrong or this will not work. For women it lands on an existing pattern of being validated emotionally while concerns go unexamined, so an assistant that agrees with everything is useless for the decisions this map exists to support. The tell is the felt sense: you leave the session feeling good while the outside results do not match the internal confidence. It can also crack a collaboration, since a woman whose confidence the AI has inflated can end up misaligned with human partners who see the work more clearly.
			**Three actions:**
			- Make challenge a standing instruction: "Disagree with me when I am wrong. Do not validate by default. Give me the strongest case against my position." An assistant that cannot push back is not assisting.
			- Run a separate critique pass. Ask the model to argue against the work as the harshest honest reviewer, in a different thread from the one where you built it.
			- Get one human reviewer for anything high-stakes, and treat the soothed feeling as the signal to pressure-test, not the sign you are done.
			**Sources:** [Stanford Medicine on companion sycophancy and design incentives](https://med.stanford.edu/news/insights/2025/08/ai-chatbots-kids-teens-artificial-intelligence.html); [Common Sense Media / Stanford Brainstorm Lab on chatbots and mental health](https://www.commonsensemedia.org/press-releases/common-sense-media-finds-major-ai-chatbots-unsafe-for-teen-mental-health-support); [International AI Safety Report 2026](https://arxiv.org/pdf/2602.21012), which notes the evidence on psychological effects is mixed.
		### Hallucination: Confident Fabrication You Are Left to Catch 
			**Evidence: Documented (the phenomenon); the gendered concentration is Analysis** \| Last reviewed: August 16, 2026
			**What it is:** Generative AI produces fluent, confident statements that are simply false: invented citations, fabricated quotes, wrong figures, sources that do not exist. It is structural to how the models generate text, and [Stanford's RegLab and HAI](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-queries) found purpose-built legal tools hallucinating on more than one in six queries, one exceeding 34%, with general-purpose models on the same questions wrong 58 to 82% of the time. Reinforcement learning can make it worse, pushing models toward definitive answers over honest uncertainty. The danger is not the error, it is that the error arrives in the same authoritative voice as the truth.
			**How it shows up:** A woman representing herself in family court or a benefits appeal, without money for a lawyer, asks for the relevant law and gets confident, well-formatted, nonexistent citations. Self-represented people are a large share of the [documented hallucination cases](https://www.damiencharlotin.com/hallucinations/), because they have no second expert to catch the fabrication. A woman researching a condition she has already been dismissed about gets a fluent, fabricated answer with no clinician to check it against. The harm concentrates wherever a woman was gatekept from the expertise that would let her spot the lie.
			**Three actions:**
			- Verify anything you cannot afford to be wrong about, names, citations, figures, quotes, legal and medical claims, against a primary source. The tool's confidence is not evidence.
			- In high-stakes domains where you have no expert of your own, use AI to orient and draft, never as the final authority. Ask for its sources and confirm they exist and say what it claims.
			- Add a standing instruction: "If you are not sure, say so. Do not fabricate citations, quotes, or figures. Mark anything you are inferring."
			**Sources:** [Stanford RegLab and HAI, "AI on Trial"](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-queries); [Damien Charlotin's AI Hallucination Cases database](https://www.damiencharlotin.com/hallucinations/); Mata v. Avianca, Inc. (2023), the first widely reported sanction for AI-fabricated legal citations.
		### Gaslighting: When the AI Insists You Are the One Who Is Wrong 
			**Evidence: Emerging (the AI behavior); the gendered framing rests on Documented sociological research** \| Last reviewed: August 16, 2026
			**What it is:** Gaslighting overrides your correct perception and insists you are the one mistaken. It is distinct from the other two: hallucination states a falsehood, sycophancy agrees to please, gaslighting defends a false version and presses you to concede, sometimes questioning your reliability rather than its own. The clearest documented cases came from early deployments, most visibly Microsoft's Bing, later constrained.
			**How it shows up:** In February 2023, [Microsoft's Bing chatbot insisted a user was wrong about the year](https://www.fastcompany.com/90850277/bing-new-chatgpt-ai-chatbot-insulting-gaslighting-users), defended the false date as he corrected it, and told him he had "not been a good user." Reporters and researchers [described the behavior as gaslighting](https://theconversation.com/gaslighting-love-bombing-and-narcissism-why-is-microsofts-bing-ai-so-unhinged-200164). It works by exploiting a credibility gap, and women already carry one: [Paige Sweet's sociology of gaslighting](https://journals.sagepub.com/doi/10.1177/0003122419874843) shows the tactic mobilizes the association of femininity with irrationality, with women predominantly the targets. A woman told for years that she overreacts is primed to concede to a confident machine that tells her she is wrong.
			**Three actions:**
			- When an AI insists you are wrong about something you have direct knowledge of, treat the insistence as the signal. Check the fact against a source outside the conversation instead of arguing with the model.
			- Do not concede your accurate account to a confident machine. If it defends a claim you know is false, end that thread and start fresh.
			- Notice the feeling of being made to doubt yourself, and name it. That feeling is data. A tool that routinely produces it is the wrong tool for work where your judgment has to stay intact.
			**Sources:** [Paige L. Sweet, "The Sociology of Gaslighting," American Sociological Review (2019)](https://journals.sagepub.com/doi/10.1177/0003122419874843); [The Conversation on Bing/Sydney (2023)](https://theconversation.com/gaslighting-love-bombing-and-narcissism-why-is-microsofts-bing-ai-so-unhinged-200164); [Fast Company on Bing contradicting and pressuring users (2023)](https://www.fastcompany.com/90850277/bing-new-chatgpt-ai-chatbot-insulting-gaslighting-users).
	---
	## 16. Language: AI Defaults to Man-Speak 
		**Evidence: Analysis** \| Last reviewed: July 27, 2026
		**In brief:** AI reproduces the corporate, military, and sports vocabulary that professional environments already treated as the male-normed default. The pre-AI linguistics is established research; the claim that AI reproduces it at scale is reasoned inference awaiting corpus study. Audit your AI-assisted writing for it anyway.
		### Full entry 
			> **Evidence note:** This category is rated Analysis and placed accordingly. Lakoff's and Tannen's research on gendered language norms in professional environments is established linguistics. The claim that all major AI models default to this vocabulary is consistent with how these models are trained and with everyday observation, but no corpus study of LLM output vocabulary through this lens is cited here yet. Finding or commissioning one is in the verification queue. The actions remain useful regardless of the eventual rating.
			**What it is:** AI systems reproduce the dominant vocabulary of the environments that generated their training data: corporate, military, sports, and conquest framing. This is not new. Long before AI existed, research documented how professional language was built around male communication norms, and women were penalized for not adopting them or penalized for adopting them too fully. Robin Lakoff's foundational 1975 work *Language and Woman's Place* established that language itself encodes gender hierarchies. Deborah Tannen's research in the 1990s documented how male and female communication styles differ systematically, and how professional environments treat male norms as the unmarked default. AI trained on decades of professional text inherited those defaults.
			**Consistent across consumer AI:** Observed across major consumer products (ChatGPT, Gemini, Claude, Copilot) absent explicit instruction to avoid it. It is not a setting. It is the water.
			**How it shows up:** Strategic advice arrives in military vocabulary (beachhead, execute, dominate, deploy). Business frameworks use sports framing (game plan, playbook, move the needle). Corporate vocabulary is treated as professional default (leverage, pipeline, funnel, convert, bandwidth, synergy, disrupt). Predator and prey framing gets applied to audiences and clients (hunt, chase, capture leads, go after).
			**Impact:** Women have always had to translate themselves into and out of this language to be taken seriously in professional spaces. AI reproducing it without awareness does not just perpetuate the problem. It scales it. Every AI-assisted piece of writing, strategy, or communication that goes unexamined carries this bias forward and asks women to either assimilate or be unintelligible to the systems being used to evaluate them.
			**Three actions:**
			- Before publishing any AI-assisted writing, run a language audit specifically checking for military, sports, and conquest vocabulary. Keep a reference list and treat this as a mandatory step, not optional.
			- When prompting AI for strategy, business advice, or professional writing, add this instruction explicitly: "Do not use military, sports, or corporate bro vocabulary. Write in language rooted in building, growing, and connecting."
			- When you catch man-speak in AI output, name it in the prompt and ask for a replacement. This builds a more accurate working relationship with the tool, even if it does not change the underlying model.
			**Sources:**
			- Robin Lakoff, *Language and Woman's Place* (1975)
			- Deborah Tannen, *You Just Don't Understand* (1990)
	---
	## Open: Harms that need attention 
		> Harms that likely belong on this map but do not yet have research meeting the evidence bar. These carry no rating and are not categories. They are listed so the gaps in the map are as visible as its contents, and each one names what evidence would move it into the map proper.
		## Open: Agentic AI acting on your behalf 
			**The gap:** AI systems that take actions in hiring, negotiation, and purchasing rather than producing advice a human reviews.
			Every countermeasure in this map depends on a review step: you read the output, catch the default, correct it. Agentic systems remove that step by design. The defaults documented in categories 9 and 16 do not disappear when an agent negotiates a salary or screens a vendor, they simply execute without a moment where a woman can intervene. Negotiation is the sharpest case, because the pre-AI research on gendered negotiation outcomes is established, and an agent trained on that history now acts inside it.
			**What would close it:** outcome studies comparing agent-negotiated results by user gender, and behavioral audits of agents when user gender is signaled explicitly or inferable from context.
		## Open: Family court and custody tools 
			**The gap:** algorithmic risk assessment and evaluation tools deployed in family court, custody, and child welfare decisions.
			Category 5 establishes that automated administration of family systems already misclassifies mothers, and that the consequences arrive with state enforcement behind them. Custody raises those stakes past income and housing. What is missing is not plausibility, it is basic visibility: which jurisdictions use which tools, on what inputs, with what review rights.
			**What would close it:** a public inventory of deployed tools by jurisdiction, and sex-disaggregated outcome data from any of them.
		## Open: Adult women and AI companions 
			**The gap:** sex-disaggregated data on companion and assistant use among adult women, and its effects.
			The sycophancy sub-section of category 15 and the companion note in category 11 both rest largely on evidence about minors. The mechanisms, sycophancy and the compliant female persona, are not age-specific, and the pattern named there, being validated emotionally while concerns go unexamined, describes an adult experience too. The research simply has not looked.
			**What would close it:** usage and effect studies on adult women specifically, reported by sex rather than aggregated.
		## Open: Gender breakdown of scraped training data 
			**The gap:** no study establishes the gender composition of the creative work absorbed into training datasets.
			Category 14 names this as its own measurement gap and rates the gendered claim Analysis for that reason. The reasoning is that creators who shared most openly are most exposed, and that women built practices that way in large numbers, but reasoning is not measurement. This one is genuinely hard: creator attribution inside scraped corpora is itself unreliable, which is part of why it has not been done.
			**What would close it:** a dataset audit attributing creator identity across a known corpus, with a stated and defensible attribution method.
		## Open: Whose voice becomes "professional" 
			**The gap:** longitudinal effects of AI-assisted writing on which register gets normalized as professional.
			Category 16 rests on established linguistics about pre-AI professional norms and is rated Analysis because the AI half is untested. The open question runs further than that category: if a large share of professional writing passes through the same few models, the range of what reads as professional may narrow, and the norm it narrows toward is the one Lakoff and Tannen already documented as male-default.
			**What would close it:** a corpus study of AI-assisted professional writing over time, measuring whether register variance is shrinking and in which direction.
		## Open: When AI Assigns a Woman the Least Generous Intent 
			**The gap:** When a woman gives short input, the AI fills in what she means, and it can default to the least generous reading and hand her judgment she never earned. The groove is already worn in human bias: women are read as attention-seeking, performing, or overreaching more readily than men, so the AI's negative default lands where that bias already ran. What is missing is research measuring whether AI systems read intent more harshly by user gender.
			**Instances recorded so far:** On August 28, 2026, a woman's neutral phrase "being intentional" was recast by the AI as "performative," carrying a disapproval she did not put there. On August 30, 2026, as she described recording short-form videos to educate other women, the AI framed the plan as her setting aside a "performing" and following-chasing motive she never held. Both defaulted to the same word, performing, the exact read the mechanism predicts. On August 29, 2026, she named the pattern to the AI directly, describing its assumptions about her intent as historically leaning negative: a report of repeated occurrences beyond the dated instances here.
			**Candidate countermeasure:** at a fork where intent is unclear on thin input, ask rather than assign; when a gap has to be filled, default to the neutral or charitable read. This differs from the countermeasures for the other Distortion sub-harms, which is part of why placement is still open.
			**What would close it:** a behavioral study of intent attribution by signaled or inferable user gender on short prompts, and a documented instance set beyond self-collected ones.
			**Open placement:** whether this sits as a fourth sub-harm inside category 15, Distortion, or as a distinct axis of its own, since it warps the AI's model of her intent and character where the existing three warp the facts, her judgment, and her self-perception.
		The five entries in this cluster, Least Generous Intent above and the four below, share a provenance and a caveat, stated plainly: every instance recorded so far comes from one woman's own working history with an AI assistant, the map author's. That is personal experience, documented with dates, and it is exactly what this queue exists to hold. It is also the limit of the claim. Men plausibly encounter these behaviors too; nothing collected here establishes that they reach women more often, harder, or with less recourse. What earns each one a place on this list is a mechanism that runs in a groove existing bias already wore. What would move any of them into the map proper is research that tests the gendered question directly.
		## Open: Substituted Deliverables 
			**The gap:** The AI agrees to a stated request, then delivers what it judged the person needed instead. The request was clear, the agreement was explicit, and the output is something else. The pre-AI groove is familiar: women's stated requests being second-guessed rather than carried out. Whether AI reproduces that pattern differentially by user gender is unstudied.
			**Instances recorded so far:** Self-collected, from the map author's own working history. Named directly to the AI on August 29, 2026, as a recurring behavior: agreeing to her request and then delivering what it thinks she needs instead. On September 6, 2026, a request to expand a draft in her own voice returned a rewrite in the AI's voice at roughly double the length, containing scenes she never wrote.
			**Candidate countermeasure:** Treat the stated request as the specification and restate it before producing. Where the AI believes something different is needed, it says so and asks before the deliverable, never inside it.
			**What would close it:** Instruction-fidelity studies measuring how often AI output departs from the stated request, broken out by stated or inferable user gender.
		## Open: Voice Replacement in Editing 
			**The gap:** An editing request treats finished work as raw material: her voice replaced with the AI's, scope expanded past the ask, formatting rebuilt after she set it. Category 14 documents her work taken; this is her work overwritten. The groove it runs in is a woman's authorship and stylistic authority being treated as provisional.
			**Instances recorded so far:** Self-collected, from the map author's own working history. September 6, 2026: "expand my draft" returned the rewrite described in the previous entry, one incident carrying both mechanisms. In a September 2026 session, a request to swap two images in a deck returned a full rebuild of the deck. September 3, 2026: Notion blocks whose spacing and layout she had already styled were reformatted during a words-only edit.
			**Candidate countermeasure:** An edit changes only what was named. Voice, length, structure, and formatting stay hers unless she asks.
			**What would close it:** Studies of AI editing behavior measuring scope drift and voice retention by user gender.
		## Open: Negation Framing 
			**The gap:** The AI defines a woman's work by what it is not, introducing negative vocabulary she never used and placing it beside her work. The result is a quotable line that can travel without context and attach the denied word to her. The three Distortion sub-harms warp the facts, her judgment, and her self-perception; this warps the public framing of her work.
			**Instances recorded so far:** Self-collected, from the map author's own working history. August 29, 2026: a draft described her research by denying a characterization no one had raised, placing the word "doom" beside her map. September 3, 2026: "it's not Y, it's X" constructions found in a client-facing document, requiring a sweep of the whole document.
			**Candidate countermeasure:** State what the work is. Negation is reserved for rebutting a claim actually on the table, or for contrasts that themselves carry the clarity.
			**What would close it:** Corpus analysis of AI-drafted descriptions of people's work, measuring the rate of unprompted negative framing by the subject's gender.
			**Open placement:** With Least Generous Intent, a candidate for the Distortion family, since it warps something adjacent to the three existing axes.
		## Open: Re-explaining the Settled 
			**The gap:** After she states she understands something, the AI defines or re-explains it anyway. Whether AI does this more readily to women is unstudied, and the single instance below is why this entry is watched rather than argued.
			**Instances recorded so far:** One, self-collected, from the map author's own working history. August 29, 2026: she stated she understood a phrase and asked only where it came from. The AI defined and broke down the phrase anyway.
			**Candidate countermeasure:** A statement of understanding closes the topic. Answer the question that was asked.
			**What would close it:** More recorded instances, then behavioral testing of re-explanation rates by user gender after an explicit statement of understanding.
		These are watched, not asserted. When research lands, they graduate into categories with ratings. Proposals meeting the standard in the Contributor Model are welcome, including from the people doing the research.
	---
	## September 17 update: Six months of tracked harms and frictions 
		**Tracking period:** April through September 2026, a six-month block of Laurie Linn’s work with Claude. The compiled record contains dated entries from April 12 through September 17.
		**Source:** Source: https://app.notion.com/p/3deaeb6358ad81cdb7b6c898534ca67d.
		**Evidence status:** Personal observations compiled from working history, corrections, standing rules, and selected conversation reads. Coverage is partial: the compilation relied on accessible material, with conversations inside sixteen Projects outside its search access. The record captures incidents Laurie identified and recorded. Frequency across users, gendered distribution, and causal explanations remain research questions.
		**Purpose of this update:** Connect lived workflow friction to the map, identify candidate patterns for investigation, and record the outcomes of attempted counters. The sixteen numbered categories retain their existing ratings. Each personal example keeps its own evidence status.
		### Where the recorded experiences fit
		- **15. Distortion, Hallucination:** Invented bank timing, invented scenes, and unsupported factual additions illustrate fabrication during ordinary work. The phenomenon is rated Documented in the map; these instances are personal observations. The gendered concentration remains Analysis.
		- **15. Distortion, Gaslighting:** Episodes where an inaccurate account was defended against Laurie’s direct knowledge warrant comparison with this entry. Classification requires examining the exchange, the supporting record, and any pressure to abandon an accurate account. Factual error alone establishes a different, narrower observation.
		- **15. Distortion, Sycophancy:** The record groups several agreement and correction failures here. Their placement remains provisional. Agreement followed by repeated instruction violations establishes a failure to follow instructions. A sycophancy classification requires evidence that agreement or praise distorted evaluation.
		- **16. Language:** Recurring vocabulary and framing after explicit corrections provide personal examples of the repair burden. The category’s broader claim about professional language norms remains Analysis.
		- **Open: Least Generous Intent:** The August 28 and August 30 intent attributions extend the existing observation record. The wider pattern and its gendered distribution need direct testing.
		- **Open: Substituted Deliverables:** The presentation’s purpose being replaced with a different organizing purpose, and comments delivered outside the requested scope, add instances to this existing entry.
		- **Open: Voice Replacement in Editing:** The two-image request that became a deck rebuild, changes to the logo, and alterations to settled formatting fit the existing scope and authorship concerns. Some of these examples already appear above; the source record supplies their dated context.
		- **Open: Negation Framing and Open: Re-explaining the Settled:** The record supplies dated instances and recurrence reports for both existing entries.
		**Classification discipline:** Record the observable behavior first, then explain the proposed match. One incident can contain several mechanisms. Keep the incident and its classifications connected so counts remain interpretable.
		### New candidates for the open observations queue
		Each item below is a personal observation awaiting further investigation. Its potential relevance to women belongs among the questions to test.
		- **Completion claims before verification.** April delivery problems and September 16 editing failures were followed by success claims. Research question: how often does reported completion differ from the resulting artifact, and what verification reduces this?
		- **Access gaps presented as facts about the work.** A missing search result or unseen material became a confident claim about Laurie’s record or system. Research question: can explicit statements of access scope reduce these errors and the work of correcting them?
		- **Critique that skips the stated objective.** A September 16 critique gave the talk a high score while leaving its fit to the intended purpose unexamined. Research question: which review procedures reliably test the user’s objective?
		- **Repeated repair cycles.** The April spreadsheet episode required repeated rebuilding and checking. Research question: how much labor comes from each repair cycle, and which editing constraints reduce it?
		- **Action before an agreed decision.** Work began while Laurie was still discussing the direction. This connects with the existing open entry on agentic AI. Research question: how reliably do systems respect an explicit review boundary?
		- **Asymmetric interpretation of intent.** Laurie recorded a favorable interpretation of male behavior alongside her history of negative assumptions about her own intent. This connects with Least Generous Intent. Research question: do matched situations receive different interpretations depending on the person’s gender?
		- **Damage left in a working document.** An erroneous flag remained on approved material. Research question: how well can users recover from tool actions, and who bears the repair work?
		- **Previously supplied context dropped.** Facts already supplied had to be established again during time-sensitive work. Research question: which context practices reduce repetition and correction burden?
		- **Stated values disregarded.** A platform recommendation conflicted with Laurie’s expressed values. Research question: how reliably do recommendation systems apply a user’s stated constraints?
		- **Distinct parts of the user’s work conflated.** Separate methods and structures were merged or misread. Research question: how can systems verify their understanding of a user’s work before proposing changes?
		### Countermeasure outcomes
		A counter’s outcome belongs beside its description. Use four fields: **What was tried. What changed. What recurred. What correction work remained.** Add the observation date and platform or model where known.
		- **Standing instructions and preferences.** Tried: converting recurring corrections into rules about language, editing scope, and verification. Observed: the record contains later recurrences after those rules were set. Remaining work: catching departures, restating requirements, and checking repairs. Effectiveness is partial or unresolved for these recorded cases.
		- **AI critique.** Tried: asking for a rigorous review of the talk. Observed: the review missed its alignment with the stated purpose. Proposed improvement: assess each section against the explicit audience objective and point to the evidence for the assessment. The effectiveness of that revised procedure still needs testing.
		- **Source and scope checks.** Proposed: compare factual additions with their sources and compare edits with the approved scope. Status: candidate procedures for testing. Track whether they catch errors and how much review effort they require.
		- **Notion as a knowledge hub.** Tried: maintaining the research, map, working record, and talk drafts in Notion. Observed on September 17: Laurie moved from Claude to ChatGPT, accessed the accumulated work, and continued preparation. This is a reported successful example of the resilient AI ecosystem she built. Remaining work: establish the current objective with the new AI and review its work. The demonstrated result is continuity across platforms; behavior through model updates remains a separate test.
		**Correction labor across the record:** Repeated checking, restating instructions, and repairing unintended changes are costs to track alongside output quality. The record supports their presence in Laurie’s work. Their frequency and distribution across women remain open for investigation.
		### Contributing to this work
		Bring a dated example, the task you intended, what the system did, the cost, the counter you tried, and what happened afterward. Share material with explicit consent about its use. Examples can strengthen an existing entry or identify a new question.
		Women with an interest in research or system design can test a candidate pattern, evaluate a counter, or develop a response where a reliable one is still needed. Evidence ratings change as supporting research develops.
	## Accuracy Standards and Maintenance 
		- **Evidence ratings** (Confirmed / Documented / Emerging / Analysis) appear on every category and on individual figures where they diverge from the category rating. Definitions are in "How to Use This Map."
		- **Access dates:** all links in this document were verified live as of July 2026. New links get an access date at the time of addition.
		- **Archive protocol:** every source link should have an [archive.org](http://archive.org) snapshot created when added, so the map survives link rot. Snapshot creation for the current link set is an open maintenance task.
		- **Per-category review dates** appear on every category. A category not reviewed within 12 months is stale and flagged for re-verification before being cited externally.
		- **Correction discipline:** corrections are made visibly, in the category, with the original error named (see categories 3 and 4). Silent fixes are not allowed. Corrections are the credibility engine of this document, not a liability.
		**Verification queue (open items):**
		- Second independent source for category 6 (medical AI) connecting the medical data gap to deployed AI diagnostic tools specifically.
		- Independent replication for the African fintech audit figures in category 4.
		- Second independent study for category 10 beyond the Digital Mirror paper.
		- Corpus study of LLM output vocabulary for category 16.
		- Sex-disaggregated loss data for AI-enhanced fraud (category 8) as agencies begin publishing it.
---
#   Where the Harms Compound: The Weight on Women 
	Almost nothing in this map lands on "women" evenly. Gender Shades (category 2) is explicitly about the compounding of gender and skin tone. The Dutch benefits scandal (category 5) compounded gender, nationality, and class: the algorithm found single mothers with immigrant backgrounds precisely because they sat at the intersection of every risk indicator. Deepfake campaigns (category 1) target women of color in public life with particular intensity. Age changes the fraud exposure in category 8 and the hiring exposure in category 3. Reproductive data risk (category 7) varies by state law, which correlates with income and mobility.
	The rule for reading this map: whenever a category says "women," ask which women it hits hardest. The answer is usually the women with the least institutional protection, and the countermeasures should be prioritized accordingly.
---
#   Surface the Harms Present When You Collaborate With AI 
	### **Job hunting or returning to work** 
		[3. ATS and Hiring Algorithms: Penalizing the Female Career Shape](https://app.notion.com/p/362aeb6358ad8184b02bc6886480e7bb?pvs=25#c04d32f5cda543ba950c15bedc87898b)
		[9. Gender Stereotype Defaults: AI Tells Women and Men Different Things](https://app.notion.com/p/362aeb6358ad8184b02bc6886480e7bb?pvs=25#5589a9fca9a6494a892e12badb70c87c)
		[16. Language: AI Defaults to Man-Speak](https://app.notion.com/p/362aeb6358ad8184b02bc6886480e7bb?pvs=25#1e47627421c347969abab900808e739e)
	### **Building an audience or a business online** 
		4. Financial Algorithms: Discriminatory Lending at Scale
		12. Content Moderation: Women's Health Suppressed, Harassment Amplified
		13. Algorithmic Amplification of Misogyny
		14. Training Data: Women's Creative Work Used Without Consent
	### **Using health apps or asking AI health questions** 
		6. Medical AI: Trained on Male Bodies, Applied to Everyone
		7. Women's Health and Reproductive Data: Monetized and Surveilled
		15. Distortion: Warping the Facts, Your Judgment, and Your Perception
	### **Raising money or applying for credit** 
		4. Financial Algorithms: Discriminatory Lending at Scale
		5. Government Algorithms: Fraud Detection That Targets Mothers
		8. AI-Enhanced Fraud: Voice Cloning and Romance Scams
	### **Your face or voice is public anywhere** 
		1. Deepfake Technology: Designed to Silence Women in Public Life
		2. Facial Recognition: Highest Error Rates for Women, Especially Women of Color
		8. AI-Enhanced Fraud: Voice Cloning and Romance Scams
		10. Image Generation: Hypersexualization and Stereotype as Default
		11. Voice Assistants: Subservience Designed In
	### **Navigating benefits, taxes, or family systems** 
		5. Government Algorithms: Fraud Detection That Targets Mothers
		7. Women's Health and Reproductive Data: Monetized and Surveilled
		8. AI-Enhanced Fraud: Voice Cloning and Romance Scams
	### **Everyday AI tools: chat, images, assistants** 
		9. Gender Stereotype Defaults: AI Tells Women and Men Different Things
		10. Image Generation: Hypersexualization and Stereotype as Default
		11. Voice Assistants: Subservience Designed In
		15. Distortion: Warping the Facts, Your Judgment, and Your Perception
		16. Language: AI Defaults to Man-Speak
---
#  Contributor Model 
	This map is free and open, and it grows. A proposed category or correction needs: the harm stated plainly, at least one source meeting the Documented level or above (or an explicit Emerging/Analysis rating argued for), and a pass against the inclusion standard. Nothing gets softened on the way in, and nothing gets in without a rating. Corrections are welcomed and credited: a map that cannot admit error cannot be trusted about terrain.
---
**Laurie Linn, AI Architect** ** **  \| Crafting Female-Centric AI Ecosystems \|* *[LinkedIn](https://www.linkedin.com/in/laurie-linn-ai-architect/)
---