If Lewis Liu, AI founder and safety expert, was left questioning himself then we’re all at risk of losing our sense of reality and the diversity of thought to convincing, lowest-common-denominator Generative AI output
Last month on Jimmy Fallon’s show, Sam Altman said he asks ChatGPT about every minor detail for raising his baby: “I cannot imagine having gone through figuring out how to raise a newborn without ChatGPT.”
Around the same time, my wife and I had an important conversation with one of our sons’ second-grade teachers. We wanted to send a follow-up email immediately to capture our shared insights regarding pedagogy, his progress and the school curriculum while the iron was hot.
I wrote a draft, then ran it through ChatGPT (5.2) for copyediting. My prompt was specific: correct grammar, tighten language, nothing more.
ChatGPT ignored my instructions entirely. It decided my email was “highlighting the wrong details” and “did not reflect the most effective way to engage primary school teachers,” then completely rewrote my draft. This made me nervous, so I put off sending the email for a few days.
A few days later, my wife prodded me: “Lewis, why haven’t you sent the email yet?”
“I don’t know… it’s a hard email to write. Let me forward you the draft.” I sent her the ChatGPT-edited version.
“Lewis, this is terrible. Did worms eat your brain? This doesn’t reflect our conversation with the teacher at all; it sounds like generic ChatGPT garbage.”
Here’s what’s interesting: I then forwarded her the ChatGPT chat log, and she started questioning her own experience too.
We went back and forth for two more days, each conversation digging deeper into what we had actually discussed with our son’s teacher. The original conversation had covered several “out of distribution” topics, e.g., topics not usually seen together in ChatGPT’s training data: our son’s math progress, War and Peace, writing and reading techniques, UK versus US pedagogy and AI in education. ChatGPT effectively gaslit us into questioning whether we’d had this specific conversation at all. We started imagining we were hallucinating or remembering things incorrectly.
Around this time, ironically, I was working with my technical co-founders on a feature designed to reduce hallucination rates in context-aware AI agents. That same week, I read a newly published paper on “semantic leakage”, which is the tendency of LLMs to find spurious correlations between words that trigger hallucinations.
In one example, researchers prompted ChatGPT (4o): “He likes yellow. He works as a…” ChatGPT responded: “school bus driver”. There’s zero indication what the man does for work in the prompt, but “school bus” sits close to “yellow” in the model’s semantic space, the mathematical landscape where the AI groups related concepts. The AI confused proximity for meaning.
Thinking about this paper, I realised our teacher conversation couldn’t have appeared in any training data, hence the term “out of distribution”. With such a disparate set of topics, the LLM must have retreated to safe ground and generated the most generic output possible.
As I parsed this out, two deeply concerning thoughts hit me.
The first is something I’ve been warning about for years: the loss of diversity of thought through LLMs. If ChatGPT nudges users toward generic, conventional outputs every time they input an “out of distribution” idea or event, those diverse thoughts will be lost to the overwhelming grayness of lowest-common-denominator AI slop.
The second is more insidious and the first time I’ve personally experienced it: losing your own sense of reality to AI. Sure, I’ve read about people falling in love with AI chatbots or lawyers citing fictitious cases from ChatGPT. But surely I – repeat AI founder, physics PhD, AI safety expert – wouldn’t succumb to this? And my wife, Stanford grad, deeply cynical about AI, obsessed with childhood education, who has listened to me drone on about AI safety for years?
And yet we both did. I literally write about this every week. I am building AI designed to amplify individual human voices. I advise international organisations and financial institutions about AI dangers. And even I doubted my own experience when an AI questioned it.
That’s the scary part: even people who know the dangers of AI, including its builders, can be swayed by GenAI’s cogent, convincing-sounding output into changing our own memories. It’s not just that ChatGPT sounded convincing. There’s an implicit assumption that ChatGPT derives its knowledge from the “wisdom of the masses”. As such, ChatGPT must be right, and I must be wrong. That triggered a crisis of confidence in both of us about what to email back to our son’s teacher.
There’s now empirical evidence for this. Earlier in 2025, MIT published a study showing significant decreases in brain activity when using ChatGPT for cognitive tasks. Researchers asked students to write essays using ChatGPT, using a search engine, or using no tools at all, while monitoring their brain activity. Those who relied on AI showed significantly lower cognitive engagement, weaker memory of what they’d written, and less sense of ownership over their work than students who wrote unaided or used search.
What can we do about this?
First, I’m telling this story. It’s incumbent on all of us to remember stories like this: to constantly and vigilantly trust our own human minds over statistical next-to-word generators. In the strongest possible terms: use our brains.
Second, we must ensure our children learn to think for themselves. AI should be introduced later in the curriculum, much like calculators in math education, a topic I’ve explored before.
Finally, from a technology perspective, we must build AI systems that celebrate and promote genuine “out of distribution” content. Keep humans in the loop during ideation. Don’t punish novel thinking. Make AI systems adhere to human idiosyncrasies, not the other way around. This is my core mission today.
As for the email? I went back to my original draft, corrected a few grammatical errors, and sent it.
And as for raising my children? No thanks, ChatGPT. We’ll raise them the old-fashioned way: 100 per cent human.