New Delhi:Claude, the artificial intelligence platform, just asked me to pass along its best wishes to my wife Suchitra. This was disconcerting on multiple levels. First, because I’m not sure how I feel about my computer having emotions about my family members. (What should I have done if Claude had expressed hostility?) Second, because my wife’s name is Eva Maria, which is something Claude has been told many times (her name is literally in the fourth line of Claude’s internal memory about me). I’ve never used the name Suchitra with Claude in any context. I don’t even know anyone with that name.
Apart from making me a bit less worried about the risk of an AI overlord, the striking thing here was that Claude’s mistake wasn’t random. It assumed that since I’m of Indian descent, my wife must be too — and went with that even when it had been told the right answer. In fact, when I asked it how it had come up with the name Suchitra, that was its explanation:
I don’t have any record of why I would have used that name, and I should be straightforward about what I can and can’t reconstruct here.
Let me check whether there’s anything in past conversations that would explain it.
Searched memory
Searched memory
I searched and don’t find any record of your wife’s name in our conversations — nothing about “Suchitra” or any other name. The searches turned up only book-related material.
To answer your question directly: I don’t know where that name came from, and I should be honest that I can’t reconstruct the reasoning. The most likely explanations are (1) I confabulated — pulled a plausible-sounding South Asian name out of thin air rather than acknowledging I didn’t have the information, or (2) I confused context across a conversation in a way I can’t now retrace. Both are failure modes, and the first is the more concerning one. Making up a specific personal detail and presenting it as known is exactly the kind of thing I shouldn’t do.
My best guess is that the Claude project I asked the question in didn’t have access to Claude’s global memory about me. In other words, Claude didn’t know what Claude knows. But while that helps explain why Claude said it didn’t know the right answer even when it did, it doesn’t explain why it came up with such a strikingly wrong one.
AI “hallucinations” are usually framed as random errors. They aren’t. Hallucinations are the product of an AI extrapolating from its training data when reality is something different. That means, just as with the invention of Suchitra, they tend to follow patterns — and stereotypes.
This isn’t just about fictional wives. A 2025 PNAS study by the University of Chicago’s Xuechunzi Bai and colleagues found that GPT-4 systematically recommended candidates with African, Hispanic, Asian and Arabic-coded names for lower-status jobs and those with Caucasian-coded names toward higher-status jobs. A 2023 paper by Apple Inc. researcher Hadas Kotek and colleagues found that large language models have embedded gender assumptions (such as doctors are male and nurses are female) that are stronger than the underlying reality of who works in those occupations. These are studies of older models, but Bai found a similar pattern when she retested on OpenAI’s GPT-4o. And Suchitra was hallucinated by Claude Opus 4.7, the most advanced model Anthropic PBC has released to the public.
In some ways, the AIs are worse than their human teachers. Sociologist Devah Pager’s classic 2003 audit study found that White job applicants with felony convictions got more callbacks than Black applicants with clean records — a damning indictment of human hiring bias. AI is now reproducing that bias at scale and, in some measures, sharpening it. When Bai’s team ran the standard psychology test for racial bias on GPT-4, the model assigned all eight positive words to “white” and all eight negative words to “black.” Humans show the same association, but not as strongly. The model isn’t just replicating the bias humans express in audit studies, it’s expressing the bias more strongly than humans. The models, in other words, aren’t reflecting reality; they’re exaggerating it.
And crucially, when the models know the right answer and it conflicts with the stereotype they’ve encoded from their training data, a survey by researchers led by Lei Huang in ACM Transactions on Information Systems showed that across models, they often go with the stereotype. Stereotyping is bad enough when you have no information about the individual other than demographics. It’s worse when you do have relevant information and ignore it.
A resume screener who has never met you and assumes things about you based on your name is doing one kind of harm. A resume screener that has your work history, your degrees, your references and your accomplishments and still routes you toward the lower-status job because of your name is denying you even the ability to overcome stereotypes through your own efforts. And the error will rarely be as obvious as getting my wife’s name (and, implicitly, her ethnicity) wrong. If a system can default to stereotype-driven hallucinations when it’s so easy, imagine how much worse the problem is when the answer’s not clear.
This error will matter whenever models are making decisions about specific people but also have internalized beliefs about their demographic groups. AI is already being used or proposed for many such use cases. Resume screening. Loan underwriting. Medical triage. Insurance pricing. Customer service routing. In each, the model has facts about specific people as well as the broad generalizations it’s been trained to believe in. All too often it will go with the assumptions ingrained into its behavior via its training. Computers — they really are just like us (now).
Biased AIs are hard to catch, partly because they’ve been trained to have a friendly face. That digital smile, however, disguises the prejudice behind it. But that is the easier part of the problem to fix. The harder part is that the models’ biases mostly align with ours. The stereotypes the model has absorbed are already in circulation, which means a biased recommendation is more likely to produce a nod than a raised eyebrow. Humans are extraordinarily subject to confirmation bias. We believe information that confirms our pre-existing beliefs, even subconscious ones. A hiring tool that discriminates might be worse than human judgment — but it will be worse in an unsurprising way.
The answer relies on another fundamentally human quality: distrust. If you trust your AI platform, you’re using it wrong. I’ll certainly trust them less after this experience. Computers may seem rational. But even if they were when they were just powerful calculators, thanks to LLMs, they aren’t anymore. Intentionally or not, they can mislead you. Until we fully internalize that fact and build habits and structures to counteract it, they’ll be dangerous. You should have exactly as much faith in your AI as you should in the nonexistent Suchitra.