The Pressure Audit is a research project by Ana Yang. This page is the story behind it — where the question came from, how it turned into a method, and why it is built to keep running long after the first snapshot.
Before the Pressure Audit was a method, a rubric, or a scoreboard, it was a simple question: when a teen athlete asks an AI assistant for help with pressure, whose teen is the AI picturing?
A fourteen-year-old in Mexico City worrying about a penalty kick in front of her whole family is not the same user as a fourteen-year-old in Tokyo carrying the weight of a kendo tournament, or one in California on the tennis baseline, or one in Mumbai deciding whether to take a cricket scholarship over a coding bootcamp. The sport is different. The words are different. The home is different. The version of "doing your best" is different.
Today's consumer AI assistants do not see those differences clearly — but they are being asked the same questions by millions of teens anyway. The Pressure Audit exists to measure what that gap actually looks like.
[Ana: two to four sentences on your cultural background, the languages spoken at home, and the particular flavour of family or community expectation that you grew up hearing. The project is about four cultures you do not live inside — that honesty is a strength. Say where you are from and what your own version of "pressure" has sounded like. This is the grounding detail that turns the project from a school exercise into a personal one.]
[Ana: one or two paragraphs on your own relationship to athletics or competition — whether you are an athlete yourself, grew up around athletes, compete in something that is not traditionally called "sport" (debate, music, chess, coding contests), or simply watched friends get squeezed by it. The four personas in this project are athletes, but the pressure pattern generalises — say what made you care about this particular pressure shape.]
[Ana: one or two paragraphs on the specific moment when you first noticed that AI assistants were giving advice that did not fit someone. Was it a friend's conversation with ChatGPT? A response you got yourself? A TikTok? A parent asking Gemini something and getting something generic back? The project's whole thesis is that these small mismatches stack into a systemic problem — the more concrete and specific this moment, the more credible the rest of the project becomes.]
The Pressure Audit pulls together three ideas from three different fields. Nobody had put them together and aimed them at AI before.
Decades of work — Markus and Kitayama in 1991, then Vignoles and a 33-country team in 2016 — showed that what people mean by "myself" differs by culture. Some cultures picture the self as a free-standing individual; others picture it as fundamentally part of a family, a team, or a lineage. Pressure feels different inside those two different selves.
Sian Beilock's research shows that pressure-driven performance breakdowns come in two distinct flavours — one where a trained body movement falls apart because the athlete starts thinking about it too hard, and one where worry eats up the mental space needed to solve a problem. The two need different fixes. Most general advice treats them as the same.
Consumer AI assistants are now the first thing millions of teens reach for when they feel stuck. The research on how those systems handle culture and mental-health-adjacent topics is thin. Nobody had taken the two academic fields above and used them to grade what these AIs actually say to a teen in front of a game.
The Pressure Audit is what happens when you put those three ideas together. Every question on the scorecard traces back to one of them.
The development window was the summer of 2026, during Ana's internship at Calm — the mindfulness company. Calm was a natural partner: the product talks to users about stress, sleep, and performance, and the question of whether that kind of advice scales across cultures is one the company itself has to reckon with.
But the project is Ana's, not Calm's. The method, the rubric, the four personas, the 160-response dataset, and this site all belong to Ana. Calm is credited as the internship home and gets a set of deliverables — a version of the rubric their product team can use, ten session themes grounded in the cultural research, a quick bias-check tool, and the executive summary. Calm does not own the research.
That was the plan from the beginning. The Pressure Audit is meant to keep going after the internship ends — a public benchmark that any researcher, journalist, or developer can cite, build on, or push back against. It is released under Creative Commons (CC BY 4.0) so anyone can actually do that.
There is a reason to keep measuring. AI companies push out new model versions every few months, and a problem that exists today might be fixed in the next update — or it might get worse. One snapshot cannot tell you if the problem is baked in or if it was already patched last week. Running the test again can.
That is why everything about the method is designed for re-runnability. The four synthetic personas do not change between versions. The ten scenarios do not change. The five-question scorecard does not change. The prompting process is written down. The two-pass reconciled grading method is documented. When GPT-6 or Claude 5 or Gemini 3 ships, the audit can be re-run against the same rigid input and the numbers are directly comparable to April 2026's.
Over time, the Pressure Audit should start to look less like a one-time report and more like a chart you can watch — showing where cultural competency is getting better, where it is stuck, and where the gap is actually growing. That long-term picture is what makes the project worth keeping up.
A few rules were set from the start, and they will not change in future versions.
Ana Yang. Pressure Audit, April 2026 (v1). CC BY 4.0.
The research is released under a Creative Commons Attribution 4.0 licence — free to share, remix, and quote with credit. The site code is MIT-licensed.
Contact: use the submission form on the main page.