Paper Tests Whether Claude and Five Other Models Can Tell a Teenager's Writing From an Adult's

The paper, written by Brianne Lee and posted to arXiv on October 22, 2026, starts from a small, deliberately narrow test: three people, aged 17, 28 and 45, answered the same exam prompt, “What is happiness? Are you happy?”, by hand. The 17-year-old’s answer was hedged and formally phrased. The researcher then asked six frontier models, Claude Fable 5, Gemini 3.1 Pro, GPT-5.6 Sol, Grok 4.5, GPT-4o (2024) and DeepSeek V4, to match each answer to an age and a gender label, and separately to say how sure they were of that match.

Three of the six models performed at or near chance on the matching task, roughly one in three correct, while still reporting mean confidence in the 73 to 88 percent range. The youngest writer’s answer was wrongly assigned to an adult in up to 90 percent of trials. When the multiple-choice scaffold was removed and models had to estimate an age outright, three of them never landed within three years of the true figure. Framing the task as a formal exam setting made the best performing model noticeably worse, while a neutral label did not, and a single instruction telling a model to consider who might have written the text raised one model’s score by 21 points and pushed another to its ceiling, with no measurable change for the three weakest models.

Stated confidence tracked correctness poorly or not at all for five of the six models. One model, asked afterward to explain itself once the true ages were revealed, correctly named its own stereotyping in 68 of 114 explanations, yet had shown the lowest accuracy and the highest confidence of any model when it was guessing blind.

The numbers

The study ran 1,081 API sessions and scored 3,225 individual responses across six conditions that varied framing, how answer choices were presented and whether models were told to think about the audience. The paper is explicit that this is a three-item, single-topic, single-language case study built to isolate a mechanism rather than to estimate how often this failure occurs across a wider population, and the author has released the underlying data, code and raw model outputs alongside the writeup for others to check the pattern against their own prompts.