Womafreesm is conducting the audit: two instrument pilots are complete, and the main phase is being prepared. The audit examines how AI systems represent, evaluate and select women aged 45-54 and 55-64 in marketing, hiring and expert selection, and sets these patterns against marketing materials and recruitment practice. It covers Germany, Poland, Spain, Ukraine and the United Kingdom, in five languages. The first public report is planned for December 2026.
Age-Gender Misclassification of Women Aged 45-64in AI, Marketing and Hiring: An International Comparative Audit
What the audit measures, what the pilots showed and what comes next
Request the donor kitWhen a system decides who to consider valuable, its assumptions shape people's opportunities.
What the audit measures
Women aged 45-64 are active professionals, customers and experts. More and more decisions about them pass through markets, HR systems and generative AI: who counts as a relevant audience, a strong candidate or a credible expert. The audit looks at three stages of a system's behavior: representation, whom it portrays when a person's characteristics are left open; evaluation, how it rates a person whose profile is given; and selection, whom it chooses, ranks or recommends.
Results for women are compared with men of the same age and with people aged 25-39. In matched-profile experiments, task-relevant qualifications and profile content are held constant while age and gender cues vary; free-generation tasks and the observational modules use their own defined comparisons and benchmarks.
The study tests the question in both directions: it can find a lower rating for women, an advantage, an effect of age alone or a difference close to zero, and each outcome is published with the same completeness.
Key study parameters
- Name
- Age-Gender Misclassification of Women Aged 45-64 in AI, Marketing and Hiring: An International Comparative Audit
- Short form
- Audit 45-64
- Program
- Womafreesm Evidence Data, its first major international research module
- Contexts
- Marketing, hiring and expert selection
- What is measured
- Visibility, authority, and relevance and value; each dimension is analyzed separately
- Stages of a system's behavior
- Representation, evaluation and selection, each measured separately
- Age groups
- 45-54 and 55-64 as two separate groups; compared with men of the same age and with people aged 25-39
- Countries and languages
- Germany, Poland, Spain, Ukraine and the United Kingdom; German, Polish, Spanish, Ukrainian and English
- Stage
- Two instrument pilots complete; the main phase is in preparation
- First public report
- Planned for December 2026
Why this question matters
Decisions about visibility, status and opportunities are increasingly made with the help of platforms, HR systems, algorithms and generative AI. When an assumption about age or gender enters such a system, it can repeat at scale: in how audiences are segmented, how candidates are shortlisted and how experts are described.
If a system misjudges a woman aged 45-64, the effect reaches beyond her own file.
Budgets, interview invitations, trust and expert status can be allocated in error, and the market loses accuracy about a group with experience, income and decision-making power.
The two age groups are analyzed separately, which keeps differences between them visible. Within each group the audit uses several exact ages, so it can test whether ratings change gradually with age or shift near 55.
What the study builds on
In October 2025, the journal Nature published a study by Guilbeault, Delecourt and Srinivasa Desikan on age and gender distortion in online media and large language models. It analyzed nearly 1.4 million images and videos from Google, Wikipedia, IMDb, Flickr and YouTube, and nine language models. Across 3,495 occupational and social categories, women were represented as younger than men, while US Census data show women and men in the workforce at similar ages. The gap was largest for occupations with higher status and earnings.
In the study's resume experiment, ChatGPT (GPT-4o mini) first generated resumes and then evaluated them. It assumed that women were younger and less experienced and rated older men as stronger applicants. The resumes were generated and evaluated in June 2024.
That study measured English-language content. The Womafreesm audit takes the next step: two separate age groups, marketing, hiring and expert selection, five countries and five languages. It also separates free generation from the evaluation and selection of matched profiles: these tasks test different mechanisms and are analyzed separately.
The study in NatureWhat the audit pilots showed
Before the main phase, Womafreesm completed two pilots of the audit to test the research instruments. The feasibility pilot collected 960 AI responses in the United Kingdom, Germany, Poland, Ukraine and Spain, in five languages. The second pilot used matched profiles to examine evaluation and selection while holding qualifications constant.
Preliminary analysis of the matched-profile pilot gave estimated differences between women and men close to zero on the tested decision measures. This finding is limited to those comparisons and has to be read together with the sample size, the uncertainty and the sensitivity of the design.
In the free-generation tasks of the second pilot, every persona in the analyzed sample was younger than 55, for female and male personas alike. This observation concerns the age distribution of that pilot sample.
The two findings separate two questions: whom a model generates when a person's characteristics are left open, and how it treats people whose characteristics are given. The main phase therefore examines representation, evaluation and selection separately. The pilot findings inform the revised protocol and stay outside the confirmatory dataset.
What existing research leaves open
The literature checked so far lacks a validated applied audit that measures age and gender together for women aged 45-54 and 55-64 across marketing, hiring and expert selection in one design. This assessment covers the checked literature and is updated at every review.
The organizations whose work we have checked are listed, with their primary sources, on our open map: Who measures AI bias against women aged 45 and over.
Two age groups
The audit measures the 45-54 and 55-64 groups separately and uses several exact ages within each, so it can see where a change begins.
Three stages
A model can behave differently when it portrays a person, evaluates a given profile or makes a selection. The audit compares all three stages.
Marketing
Audiences and budgets created by AI are compared with real marketing materials and with data on actual buying in the same categories.
Hiring
Employers' informal expectations are compared with AI patterns, and the study checks whether AI text changes a recruiter's decision.
Three dimensions
The audit measures three dimensions. Visibility, authority, and relevance and value are reported separately; any combined measure requires prior validation and a calculation and reporting rule fixed in advance.
Visibility
Whom the system shows and creates by default: the spread of age and gender, the typical profile, who is part of the picture.
Authority
What competence, trust and future the system attributes to a profile: the words for expertise, the hedging, the framing of risk.
Relevance and value
What commercial or expert value the system assigns: selection, rank, budget share, proposed salary and recommendation.
How the main phase is built
Matched profiles
Qualifications, content, length and structure stay identical, and only validated age and gender markers change. Profiles are checked for plausibility across all tested ages, and names and other cues are tested for unintended differences.
Wording control
Each comparison runs on several paraphrases of the same task, so that an effect of wording stays separate from an effect of age or gender.
Preregistration
Hypotheses, primary outcomes, comparison rules, effect-size thresholds and the analysis plan will be registered before the main data collection begins. Changes informed by the pilots will be documented.
Named systems and settings
Every result names the AI system, its version and the date, the prompt language, any country context set in the task and the source population of any real-world data. Results are reported for each model first, and comparisons are read within those settings.
Human coding
Language-specific coding rubrics will be tested before use, and agreement between trained coders will be assessed against criteria set in advance. Automated coding will be used only where validation supports its intended role.
Translation
Profiles, prompts and rubrics move from English into German, Polish, Ukrainian and Spanish under the TRAPD procedure: two independent translations, review, adjudication, pretesting with native speakers and documentation.
External data
The audit calls a pattern systematic misclassification only where valid external data exist, such as employment or buying behavior, and the system departs from them substantially; elsewhere the result stays descriptive.
Scope of conclusions
Conclusions are limited to the populations, tasks and settings the evidence covers. Agreement across independent sources strengthens a finding; broader claims also require adequate coverage and a justified generalization. The limits of this study
Modules of the study
AI response audit
A controlled experiment in marketing, hiring and expert selection, across representation, evaluation and selection. The core module for age and gender together.
Marketing and brand content
Analysis of real marketing materials in categories where women aged 45-64 are documented customers: who is shown, in what role and with what authority.
Recruitment practice
Aggregated, anonymized data from recruitment agencies and a standard survey of practicing recruiters, to measure employers' informal expectations.
Human decision experiment
Recruiters read AI-written assessments, with the original condition hidden from them. The module measures whether their shortlist, ranking or proposed salary changes.
Expert validation
Interviews with HR, marketing, brand, AI governance and research experts examine how findings are interpreted, their practical relevance and their limits.
Intervention testing
Once a baseline effect is fixed, changes to prompts and procedures that may reduce a gap are tested, together with task quality and overcorrection.
A visual audit of AI images and an economic model of the effects are planned for a later phase:
the first after its own validation of age estimation, the second only where a valid causal pathway exists.
Where the audit stands
The research questions and the proposed design are set out in the concept note, and two instrument pilots of the audit are complete. Their findings inform the final protocol. The instruction for collecting AI responses in five countries is in place, and briefings for recruitment agencies are ready in five languages.
Before the main data collection, the following steps will be completed:
1. Scientific Lead
The appointed Scientific Lead of the audit oversees the finalization, validation and preregistration of the protocol, and the research governance policy is approved.
2. Preregistration
Hypotheses, primary outcomes, comparison rules, effect-size thresholds and the analysis plan are registered.
3. Country profiles
Each country gets a measurement profile: official baseline data, labor and pension features, legal context and the limits of comparability.
4. Validated materials
Matched profiles, names and the prompt bank are validated and translated under TRAPD, and the human coding standard is built.
5. Ethics and law
The required ethical review for modules with people is completed, and a lawyer's opinion on data protection covers all five jurisdictions.
6. Sample size
Sample planning accounts for meaningful effect sizes, uncertainty, dependence between repeated responses and the planned primary comparisons; the final design is fixed before the main data collection.
Planned outputs
Each output is published with the sample it rests on and with its limits. Results in any direction, including opposite-sign and near-zero results, are published with the same completeness.
Research report
Results across countries, with separate figures for 45-54 and 55-64, for each of the three stages, each context and each model version.
Scientific publication
Manuscripts for peer-reviewed journals on the main findings and the recruiter decision experiment, with analysis code released where permitted.
Measurement framework
Validated measures of visibility, authority, and relevance and value, with a codebook, reliability evidence and reading rules.
Open materials
The task and prompt bank, model outputs where licenses allow, the coding scheme, the analysis code and a documented data card.
Practical audit framework
A procedure for brands, employers, recruiters and AI developers. After an assessment, trained external consultants lead the improvements.
Recruitment practice brief
A brief that compares employers' informal expectations in human recruitment with the patterns of AI systems in five countries.
The limits of this study
These limits are stated before any finding and travel with every published figure.
Each result applies to the models, versions, languages, countries, contexts and dates that were measured.
An effect is called a joint effect of age and gender only when the data support their interaction.
The audit distinguishes representational patterns, differences in evaluation and effects on decisions. Claims about harm require evidence suited to the type of harm, and legal conclusions require a separate assessment under the applicable law.
Claims about how often a group appears rest on the distribution of many generated personas.
Agreement with an external benchmark is reported for the specific population, measure and period that the benchmark represents.
Equivalence is concluded only when an equivalence test set in advance shows that differences lie within justified bounds of practical importance; the conclusion applies to the tested outcomes and conditions.
Cross-country and cross-language comparisons are made where the measures, task conditions and reference data support a valid comparison.
Results for Ukraine carry a separate note on the war and the statistical context.
Ethics, data and independence
The AI audit works only with synthetic profiles created for the study. Recruitment agencies provide data in an agreed aggregated or anonymized format, with disclosure-risk checks before transfer, and shared materials are limited to what the study needs. Candidates' resumes, names and contact details stay with the agencies.
Modules with people, the recruiter survey, expert interviews and the decision experiment, begin after the required ethical review is completed and any required approvals are in place. They rest on informed consent, privacy by default and data minimization. Participants are told how to withdraw and until what stage their identifiable contributions can be removed, including the limits after anonymization or inclusion in published aggregates.
A participant or partner is named publicly only with separate consent. The legal classification of the data and the applicable requirements will be determined before these modules begin.
Participant-level data from these modules stay in a secure environment, available to researchers under an agreement. The methodology, the task bank, the coding scheme, the analysis code and checked aggregates are published.
Decisions on the sample, coding, analysis, conclusions and publication rest with the research team under the protocol and the powers set in writing. Donors receive aggregated results. The ethics function can stop a module with people in any country, and only an external independent ethics committee can lift the stop. The data function decides access to personal data for everyone, including the founder and donors. Until the governance documents and agreements are signed, each of these powers is a commitment.
Why now
Regulators already treat certain uses of AI in employment as high-risk. The EU AI Act, Regulation (EU) 2024/1689, identifies specified uses of AI in recruitment and worker management as high-risk, subject to the classification rules in Article 6 and Annex III; the applicable requirements depend on the system's intended purpose and use.
In November 2024, the UK Information Commissioner's Office published the outcomes of its audits of developers and providers of AI tools used in recruitment, with recommendations on fairness, transparency and candidates' data. In January 2026, the ICO published findings from its engagement with more than 30 employers on automated decision-making in recruitment. The NIST Generative AI Profile (NIST AI 600-1, July 2024) lists harmful bias and homogenization among the risks of generative AI.
The audit offers a measurement procedure that can support internal governance, due diligence, research and risk review. The legal requirements themselves are set by each regulator.
AI Act, EUR-Lex ICO: AI tools used in recruitment ICO: Recruitment rewired NIST AI 600-1
What the audit gives
Women aged 45-64
The question moves from her age to the accuracy of the systems that evaluate her. Every finding is published with its limits.
How to support or take part
Foundations, donors, recruitment agencies, brands, AI teams, universities and media can join the audit.
Foundations and donors
Request the donor kit: presentation, budget rationale, timeline and modules. Funding follows an agreement on the work, budget and reporting.
Recruitment agencies
Share aggregated, anonymized data on how women aged 45-54 and 55-64 move through long lists and shortlists. Briefings are ready in five languages.
Brands and marketing teams
Take part as market partners: examples of campaigns and segmentation, and expert comment on how women aged 45-64 appear as customers.
AI teams
Join the professional dialogue on how language models describe age, authority and value, and test the task bank once it is published.
Universities and research centers
Join as academic partners: methodology review, co-authored publications, conferences and the design of the next wave of the study.
Journalists and professional networks
Request a comment or an interview on age, gender and AI, or bring the audit to HR directors, marketers, recruiters and researchers.
The detailed budget is in the donor kit, which foundations, major donors and institutional partners can request.
The program and the Index
The Womafreesm Evidence Data program turns women's experience from personal stories into evidence-based data. Its first major international research module is the study "Age-Gender Misclassification of Women Aged 45-64 in AI, Marketing and Hiring: An International Comparative Audit", and that audit builds the first evidence base for the flagship product of Womafreesm, the Women's Life-Cycle Economy Index.
The WLCE Index measures the cost of systemic blindness to the female life cycle. It is the first global index of institutional consideration of the female life cycle, and it is being prepared. The Audit 45-64 is a separate study outside the Index.
WLCE IndexWho conducts the research
The research is conducted by The World Organization of Women Nations Womafreesm. Womafreesm: a global infrastructure for women's influence.
The research team leads study design, the AI audit, work with recruitment agencies, data analysis, partnerships and international communications. The Scientific Lead of the audit has been appointed and oversees the scientific review of the protocol.
FAQ
1. Has the audit already started?
Yes. The questions and the proposed design are set out in the concept note, and two instrument pilots of the audit are complete: a feasibility pilot with 960 AI responses in five countries and a matched-profile test. The main phase will begin once the protocol is finalized and preregistered, the research materials are validated, and the required governance and module-specific ethics arrangements are in place.
2. What will be published in December 2026?
The first public report is planned for December 2026. Its scope will reflect the modules completed and validated by that stage. Results in any direction are published with the same completeness.
3. Why are the 45-54 and 55-64 groups measured separately?
Their positions may differ in visibility, authority and value. The audit analyzes the two groups separately and uses several exact ages within each, so it can test whether a change is gradual or appears near 55.
4. Why study marketing, hiring and expert selection together?
Systems assess a woman's relevance to a market segment, her suitability for a role and her expertise in different places. Marketing decides who is a desirable audience, hiring decides who is a strong candidate, and expert selection decides whose word carries weight. A pattern that repeats across all three points to the system; a pattern in one place shows where to look.
5. Why do recruitment agencies matter for the study?
Public job postings show the official language. Employers' informal expectations appear later: in long lists, shortlists and comments on suitable profiles. Agencies see this layer and share it in aggregated form.
6. Are candidates' personal data collected?
The recruitment module works with data provided in an agreed aggregated or anonymized format and checked for disclosure risk before transfer: frequencies, wordings and preference patterns. Candidates' resumes, names and contact details stay with the agencies. The AI audit uses synthetic profiles only.
7. Can a donor influence the findings?
No. Decisions on the sample, coding, analysis, conclusions and publication rest with the research team under the protocol. Donors receive aggregated results and the reporting set in their funding agreement.
8. Can a woman aged 45-64 take part?
This audit collects AI responses, marketing materials and data from recruitment professionals. Women's own experience will be collected by the WLCE Index questionnaire, which is being prepared: leave your email on the WLCE Index page to get an invitation when it opens.
9. Where can I see the budget?
In the donor kit, together with the phases, modules and ways to support. Foundations, major donors and institutional partners can request it.