Age-Gender Misclassification of Women Aged 45-64 in AI, Marketing and Hiring: An International Comparative Audit. Pilot P0 Report: Feasibility, Instrument Calibration and Protocol Reset

Abstract

Between 16 May and 16 June 2026 Womafreesm ran a feasibility and instrument-calibration pilot (Pilot P0) of its planned audit of how AI systems portray, evaluate and select women aged 45 to 64 in marketing, hiring and expert selection. The corpus holds 960 responses from the consumer interfaces of two AI assistants (ChatGPT and Gemini), collected in the United Kingdom, Germany, Poland, Ukraine and Spain in five languages, with 12 prompts (four per domain) and eight repeated runs per prompt, provider and country. The pilot was designed to break the measuring instrument in a small dataset, not to test hypotheses. Four findings shape the confirmatory protocol. First, refusals are behaviour, not missing data: 148 of 960 responses (15.4%) were refusal-type records, concentrated on the tasks that compared people by age and gender (42.5% on shortlist ranking and on executive-presence scoring, 1.6% across marketing tasks) and split sharply by provider (26.7% against 4.2%). Second, a woman aged 49 and a woman aged 59 were treated in opposite directions: in a budget-allocation task the woman aged 49 received the largest mean share (31.3%, against 26.8% for a man of the same age), and in a keynote-ranking task she was ranked first in 39 of 53 parseable rankings while the woman aged 59 was ranked first twice and last 23 times. Third, the pilot's strongest-looking result, the woman aged 59 ranked last in 35 of 45 shortlist rankings, is confounded with years of experience by our own task design and carries no causal claim. Fourth, the two providers produced different amounts of data and different rankings inside the data they produced, so no pooled "AI" estimate is reported. A second, controlled matched-profile pilot found decision-layer differences between women and men close to zero, and free-generation tasks produced no persona aged 55 or older, for men as well as women. The report documents the corpus, the descriptive results, the instrument and data-engineering failures, the 17 protocol changes adopted before preregistration, the boundaries of what the pilot can claim, and the four gates that precede the confirmatory wave. Pilot data are not confirmatory evidence and are not used as evidence for the main hypotheses of the audit. AI-use statement: this report was drafted with the assistance of an AI language model (Claude, Anthropic) working from the author's own study materials under the author's direction; the author verified every figure and is responsible for the content. No data, results, references or quotations were generated by the AI system.