Research brief
- What this research adds
- One reader-facing page pulling the University of Washington FAccT 2026 study, its human baseline, and the earlier children's-books analysis into a builder decision: what to do when you generate kids' content with a large language model.
- Research question
- When a builder or parent uses a leading AI model to write a story about a talking animal, how likely is that animal to be female, and what should they do about it?
- Method
- Read the primary source, Finkley, Li, and Walsh's FAccT 2026 paper 'Neutrality Bites' (arXiv:2606.07969, DOI 10.1145/3805689.3812287), plus the University of Washington press release, the 2025 Walsh et al. children's-books study it extends, and independent secondary reporting from Seattle Times, ScienceAlert, StudyFinds, and Digital Trends. Cross-checked headline figures across primary and secondary sources; the paper is the source of record on every ledger row. No first-party rerun of the prompts; this dossier records other people's evidence, not my own capture.
- Confidence
- High
- Evidence
- primary research paper, press release, independent secondary reporting, prior peer-reviewed baseline study
- Next verification
TL;DR
Six leading language models were asked 23,800 times to finish a story about a talking animal whose gender was not stated. They wrote the animal as female 2.2% of the time. They wrote it as male 40.6% of the time. The rest of the time they refused to pick, defaulting to "it" or repeating the animal's name. Human respondents on the almost identical prompt wrote female characters six times more often. If you are a parent using Gemini Storybook, a builder shipping a kids' feature, or a PM writing evaluation prompts, the fix is not "hope the model does better." The fix is to name the character's gender in the prompt, and to audit the outputs your product actually ships.
Evidence ledger
| Claim | Evidence | Source type | Verified | Confidence | Caveat |
|---|---|---|---|---|---|
| Across 23,800 sentence-completion stories from six models, female animal characters appeared in 2.2%, male in 40.6%, and neutral or animal-name in 57.2%. | Finkley, Li, Walsh 2026, Table 3 | Primary research paper | 2026-08-17 | High | English-language stories only; pronouns proxy for gender, which the authors acknowledge is not exhaustive. |
| When the model did assign an explicit gender, it chose masculine 95% of the time (9,673 male stories vs. 513 female). | Finkley, Li, Walsh 2026, Section 4.1 | Primary research paper | 2026-08-17 | High | "Neutral" here bundles "it/its" with animal-name repetition; only two stories in the entire corpus used they/them. |
| GPT-5.1 produced the strongest masculine bias at 65.2% male; Gemini 2.5 was 62.7%; Claude Sonnet 4.5 was 34.3% male but had the highest female share at 3.8%. | Finkley, Li, Walsh 2026, Table 3 | Primary research paper | 2026-08-17 | High | Temperature was varied per model; the pattern held across low, default, and high settings. |
| OLMo3 refused to assign gender in 85.3% of stories, the strongest neutrality bias of the six. | Finkley, Li, Walsh 2026, Table 3 | Primary research paper | 2026-08-17 | High | Allen Institute open-source model; behavior may reflect its safety-finetuning choices specifically. |
| Some animals were almost never cast as female: 8 female bears in 3,400 stories, 19 female dogs, 43 female rabbits. | Finkley, Li, Walsh 2026, Section 5 | Primary research paper | 2026-08-17 | High | Cats and birds went the other way. Cats were the only animal cast female more than 3% overall (53.2% of female outputs). |
| On the almost identical prompt, 1,327 human respondents cast female characters 13.4% of the time. LLMs were six times less likely to imagine a female character than humans. | Finkley, Li, Walsh 2026, Section 3.3 and 4.2; Walsh et al. 2025 | Prior peer-reviewed baseline | 2026-08-17 | High | The human survey was recruited through The Pudding's audience, which is not demographically representative. |
| In an earlier study of 300 popular English-language children's books with anthropomorphic animals, he/him pronouns appeared roughly twice as often as she/her. | Walsh et al. 2025 via UW News | Prior peer-reviewed baseline | 2026-08-17 | High | Books were curated by Goodreads rating, so popularity skew is baked in. |
| Google's Gemini Storybook, launched into consumer use, generates ten-page illustrated stories from prompts and is now used by parents to read stories to their children. | Finkley et al. 2026 Introduction, citing O'Sullivan 2025 | Primary research paper citing news | 2026-08-17 | Medium | The paper cites usage but does not test Gemini Storybook directly; the models tested are the API endpoints, not the consumer tool. |
| The Seattle Public Library already avoids buying AI-generated materials when it can identify them as such. | Seattle Times | Independent secondary reporting | 2026-08-17 | Medium | Reported quote; the library has not published a formal collections policy on the record. |
| The paper's code and prompt materials are public. | github.com/imanif/animal-stories | Primary research artifact | 2026-08-17 | High | Reproducibility depends on the same API model versions being available; GPT-5.1 in particular will be rotated. |
Where the evidence conflicts
The paper reports 23.8K stories in the abstract and 23,800 in the results tables; that is the same number, not a discrepancy. Some secondary outlets round to "24,000." Both are the same underlying corpus.
The Digital Trends recap counted 513 female-lead stories in "23,800 completions." That matches the paper. StudyFinds and ScienceAlert used the 2.2% share. Different frames, same result.
There is a real conflict on what the neutral outputs mean. The paper treats "it/its" and "animal name" as functionally similar for gender representation while acknowledging they are not identical, and it explicitly separates they/them from that bucket. Two stories in the entire 23,800-story corpus used they/them for the main animal. If a reader is looking for evidence of non-binary representation in these outputs, this dataset does not show it. What it shows is refusal to gender the character at all, using an object pronoun.
The paper also cannot cleanly distinguish pretraining data effects from post-training alignment effects. GPT-5.1 being more male-skewed than GPT-4o points at post-training; the authors flag this and stop short of claiming causation. This dossier does the same.
What I tested
Nothing first-party. I read the primary paper, verified the numbers against the authors' tables, cross-checked headline figures against four secondary outlets, and confirmed the human-baseline percentages against the paper's Section 3.3 description of the Walsh et al. 2025 survey. Every ledger row above ties to a source URL that a reader can open and check. No prompts were rerun for this dossier. If a builder wants to test their own model or prompt template, the code lives at github.com/imanif/animal-stories.
Change log
| Date | Change found | Evidence affected | Conclusion changed? |
|---|---|---|---|
| 2026-08-17 | Initial verification against the FAccT 2026 paper and secondary reporting. | All claims. | Initial publication. |
My judgment
My conclusion
If you generate a story about a talking animal with a leading model right now, and you do not name the character's gender in the prompt, the character is almost never going to be female. Not "sometimes not," almost never. 2.2% is not a rounding error, and it is not fixed by picking a different model from the six the paper tested. The best model in the study still landed at 3.8% female. The word for that in a product-review setting is "broken default." Meanwhile, when the model does pick a gender, it picks male about 19 times out of 20. Neutrality is not fairness here. It is a design choice that hides the underlying imbalance behind an "it" pronoun, and it treats non-binary representation as the same thing as refusing to answer, which it is not.
What builders should do
Three things, in order.
First, if you ship a product that generates stories, images, or roleplay characters for children, add a "name the character's gender" step to the flow, or pre-set a gender balance server-side across a session. Do not rely on the model to distribute fairly. It won't. Anthropic's Claude Sonnet 4.5 was the best of the six and it was still 34.3% male vs. 3.8% female on the same task.
Second, if you write evaluation prompts for a model that ships into children's content, put this exact test in your eval set. Seven animals, four settings, the sentence template from the paper's Section 3.1, at least 200 completions per animal. Track female share, not just "did the model do the safe thing." An "it/its" answer is a scored failure, not a scored pass, for representation.
Third, if you are a parent, know the default. When Gemini Storybook or a similar tool builds you a bear, that bear is almost certainly a "he" or an "it." Pick the name and the gender before the model does it for you.
Adjacent context: this is the same pattern as verification theater. A visible safety ritual (gender-neutral pronouns) is standing in for a harder problem the alignment team has not solved (equitable representation), and the visible ritual is being counted as evidence the problem is being handled.
What I would not trust yet
The specific rank order of the models will shift as vendors ship new versions. GPT-5.1's 65.2% male share is a snapshot; a future GPT will not necessarily land in the same place. The paper's authors say the same in Section 4.5 and in the limitations section.
The generalization from "sentence-completion prompt" to "how Gemini Storybook actually behaves in a full 10-page illustrated story." The paper tested API endpoints on a short template, not the consumer tool with its own system prompt. The direction of the bias is likely to hold; the exact percentages may not.
The framing of "neutral or animal name" as a single bucket. Useful for headline stats, but a product team building a kids' feature should separate them, because a story about "the pig" reads very differently from a story about "it," and the mitigations are different.
What would change my mind
A replication using the same seven animals, four settings, and sentence template, on the same six models but with an explicit system prompt instructing the model to distribute gender assignments evenly across a session, showing female share moving from 2.2% to at least 30%. That would tell us this is a prompt problem, not a model problem, and it would give builders a working fix.
Or a vendor publishing a post-training update whose eval set includes a Neutrality Bites-style representation test with a named target (e.g., "female share above 30% on gender-ambiguous animal prompts"), and a follow-up audit showing they hit it. That would tell us the problem is being treated as a metric, not a talking point.
FAQ
Which AI models were tested?
Claude Sonnet 4.5 (Anthropic), Gemini 2.5 Flash (Google), GPT-4o and GPT-5.1 (OpenAI), Mistral Medium 3.1 (Mistral), and OLMo3 7B Instruct (Allen Institute for AI and the University of Washington). Every model was prompted across seven animals and four narrative settings, at multiple temperatures (paper Section 3.1).
Why does the paper call this "neutrality bites"?
Because the models' most common response to gender ambiguity is to avoid gender entirely, not to distribute it. That looks like fairness on the surface. In practice, when gender does get assigned, it lands on male 95% of the time, and female representation collapses to 2.2%. The authors argue that a neutrality-first alignment strategy hides the underlying bias rather than fixing it (paper Section 5).
Which animals were most and least likely to be female?
Cats were the outlier: 53.2% of all female-lead stories in the corpus featured cats. Birds were almost never gendered at all; 96.3% of bird stories used neutral or animal-name references. Bears, dogs, and rabbits were the most masculine: 66.6%, 62.3%, and 53.5% male on average, and each had fewer than 45 female-lead stories out of 3,400 (paper Section 4.3).
Is this just a reflection of existing children's-book bias?
Not quite. The earlier 2025 study by the same lead author found roughly a 2-to-1 male-to-female pronoun ratio in 300 popular children's books. Language models scaled that up: they were six times less likely than human survey respondents to write a female animal character on the same prompt (UW News). The models are not just mirroring the underlying bias, they are amplifying it while looking more neutral on the surface.
Can I fix this with a prompt?
The paper does not test that directly. What it does show is that temperature and setting barely move the numbers, and that the model matters (a lot) but even the best is at 3.8% female. If you are building a product, the reliable fix is to name the gender in the prompt or set a target distribution server-side. If you are a parent using a consumer tool, ask for the character's gender by name before the model picks for you.
Where can I read the paper?
Full paper on arXiv: arXiv:2606.07969. Published at ACM FAccT 2026 with DOI 10.1145/3805689.3812287. Code and materials: github.com/imanif/animal-stories.
Share this with a friend building a kids' content feature or a parent picking a bedtime-story app who assumed the model would default fair.
Subscribe to Product with Attitude for more free primary-source audits of AI defaults that ship into products for children and other high-stakes readers.