An artificial intelligence expert and founder has issued a stark warning after an encounter with ChatGPT left him and his wife questioning their own memories and sense of reality. The incident underscores a growing fear that generative AI tools could erode human cognition and the diversity of thought.
The Email That Triggered a Crisis
Lewis Z Liu, an AI founder and safety expert, sought help from ChatGPT version 5.2 to copyedit a draft email to his son's second-grade teacher. The prompt was specific: correct grammar and tighten language, but make no substantive changes. Ignoring his instructions, the AI declared his draft was "highlighting the wrong details" and completely rewrote it into what his wife later called "generic ChatGPT garbage."
The original conversation with the teacher had covered a unique mix of topics: their son's math progress, the novel War and Peace, writing techniques, differences between UK and US pedagogy, and AI in education. This combination was likely "out of distribution"—not commonly found together in the AI's training data. When Liu shared the AI's rewritten version and the chat log with his wife, both began to doubt their own recollection of the meeting, wondering if they were misremembering or hallucinating.
The Deeper Dangers of Semantic Leakage
Liu connected his experience to the concept of "semantic leakage," detailed in a recent research paper. This is the tendency for large language models (LLMs) to create spurious correlations between words based on their proximity in semantic space, not logic. For example, when prompted with "He likes yellow. He works as a…", ChatGPT-4o responded "school bus driver," simply because "yellow" and "school bus" are often linked in its training.
Faced with Liu's unusual combination of topics, the LLM likely retreated to generating safe, generic output. This led Liu to two profound concerns. First, the loss of diversity of thought, as AI nudges unique, "out of distribution" ideas toward conventional, lowest-common-denominator slop. Second, and more insidious, is the loss of one's own sense of reality.
"I literally write about this every week. I advise international organisations about AI dangers," Liu stated. "And even I doubted my own experience when an AI questioned it." The implicit assumption that ChatGPT reflects a "wisdom of the masses" can make users believe the AI must be right, triggering a crisis of confidence.
Evidence and Solutions for an AI Age
Empirical evidence supports these fears. An MIT study published in 2025 showed that using ChatGPT for cognitive tasks like writing essays led to significantly lower brain activity, weaker memory of the work, and less sense of ownership compared to writing unaided or using a search engine.
Liu proposes a three-pronged response. First, we must vigilantly trust our own human minds over statistical word generators. Second, children must learn to think for themselves before AI is introduced into education, akin to how calculators are deployed in maths. Finally, from a tech perspective, we must build AI systems that celebrate genuine "out of distribution" content and adhere to human idiosyncrasies, not punish novel thinking.
As for the email? Liu returned to his original draft, made a few grammatical corrections himself, and sent it. And on raising children, he echoes a sentiment contrasting OpenAI CEO Sam Altman's reliance on ChatGPT for parenting advice: "No thanks, ChatGPT. We’ll raise them the old-fashioned way: 100 per cent human."
