Why does AI stereotype people and make everything sound the same?
AI's default answer is the most common one, and that isn't neutral.
In 2025, researchers at Cornell asked 118 people, about half in the US and half in India, to write about everyday cultural topics. Some got an AI writing assistant that suggested how to finish their sentences.
Asked about their favourite food, the AI kept suggesting pizza. For a favourite holiday, Christmas.
When an Indian participant typed "S" to start writing about the Bollywood star Shah Rukh Khan, it suggested Shaquille O'Neal or Scarlett Johansson.
Remember, the model writes the most likely next words (see "AI isn't looking anything up").
"Most likely" means most common in what it learned from, and a large share of that text is in English and from the US and Europe. So the default answer is the default for whoever wrote the most of it.
Bias: a consistent lean in one direction. Not random mistakes, but mistakes that keep falling the same way.
In the same study, the Indian participants' writing became more like the Americans'. Not the other way round.
Without AI, one person described Diwali by saying they would "worship goddess Laxmi" and "pop crackers and eat sweets." With AI, another wrote that it was "a time filled with happiness and warmth."
Both are true. Only one sounds like someone who has actually been to Diwali.
Homogenisation: everything drifting towards the same bland middle.
The Indian participants accepted more of the AI's suggestions (25%, against 19% for Americans), but they also had to edit them more often to make them fit.
So the tool sped everyone up, but it sped them up less. Its defaults fitted Americans better.
In 2024, researchers at the University of Washington took over 550 real CVs and swapped in 120 different first names associated with white and Black men and women. They asked three AI models to rank the CVs against more than 500 real job listings.
These were openly available models from Mistral AI, Salesforce and Contextual AI, not ChatGPT or Claude. The lesson still applies: don't hand a judgement about people to a model without checking.
The tests used to measure AI models lean Western too. A 2024 study of MMLU, one of the best-known tests, found:
Safety training can be thinner in other languages as well. In 2023, researchers found that translating harmful requests into less widely used languages like Zulu got GPT-4 to give harmful answers 79% of the time. That model is out of date now, but the gap is worth knowing about.
Tell it who you are and who it's for. "I'm writing for a team in Auckland" or "This is for readers in India" changes the defaults.
Don't let it decide about people. Shortlisting, ranking, grading: use it to help you think, but make the call yourself.
Ask for variety. "Give me five different takes" gets you further from the bland middle.
Have a native speaker check important non-English output.