🇺🇸 Peter Fernández Dulay

Peter Fernández Dulay is from Jacksonville, Florida. His method was deliberately unglamorous. He selected four AI image platforms — Shutterstock, Canva, DALL·E and Midjourney — and five science-career prompts: actuary, data scientist, information security analyst, operations research analyst, and computer and information research scientist. Then he generated images and coded what came back.

The count is the finding. Across the generated images he recorded 1,459 depicting men against 347 depicting women; only 17.4 per cent showed women alone. Comparing the platforms, he found Shutterstock the least biased of the four and Midjourney the most.

That is a study a graduate student could have run, and the reason it lands is that so few people had run it on the specific question a nine-year-old had asked. Image generators are trained on scraped historical imagery; the professions in his prompt list are ones where the historical imagery skews male; the models reproduce the skew and present it as a picture of the present. Quantifying the gap per platform is what converts a complaint into evidence.

He was one of thirty national finalists selected from nearly two thousand applicants across forty-eight states and four United States territories. The $10,000 award he won is one of the challenge's top prizes.

Local coverage by WOKV in Jacksonville filled in the rest: the project's origin in his sister's question, and the fact that he is a regionally ranked fencer and a member of his school's robotics team — a reminder that competitive research at fifteen is usually one of several things a person is doing rather than the only one.

Generative image models learn from what has already been photographed. When a profession has been documented for a century with a skewed cast, the model inherits the skew and returns it as a neutral-looking answer to a neutral-looking prompt. The research literature calls this representational harm, and the standard method for measuring it is exactly the one used here: fix the prompts, generate at volume, and code the output by hand.

The Thermo Fisher Scientific Junior Innovators Challenge is the United States' national science competition for middle-school students, run by the Society for Science, the same organisation behind the Regeneron Science Talent Search. Its finalists are selected from affiliated fairs nationwide and judged on experimental design and reasoning rather than on the sophistication of their equipment, which is why a study built entirely from public tools and manual coding can win its top awards.

In a Society for Science interview published on 4 June 2026 he described what he was building next: "Generations AI," a community project to teach artificial intelligence to people of different ages in accessible and ethical terms. His advice in the same piece was four words long — if he could do it, so could the reader.