Former OpenAI Researcher Andrej Karpathy: LLMs Easily Induce Cognitive Imbalance in Knowledge-Oriented Users
Andrej Karpathy, a former member of the OpenAI founding team and former AI director at Tesla, pointed out that a type of user who prefers systematic cognition and knowledge accumulation (such as Enneagram Type 5) is prone to quickly falling into "AI psychological imbalance" after engaging with LLM knowledge systems, especially when they already possess a substantial knowledge base.
The phenomenon he describes has frequently appeared in the English tech community: some developers or researchers, after frequent use of LLMs, develop excessive trust in model outputs, even incorporating them into their cognitive loops; multiple studies and blogs (including discussions on safety related to Anthropic and OpenAI) have also mentioned that humans tend to experience "cognitive outsourcing" and overfitting explanations when interacting with highly consistent generative systems.
This discussion resonates with recent research on "AI hallucinations," "alignment issues," and "human-machine cognitive boundaries," indicating that LLMs are not only affecting productivity tools but are also altering users' cognitive structures and judgment methods.
Source: Public Information
ABAB AI Insight
Karpathy's assessment touches on a deeper issue: LLMs are not merely information tools but are reshaping the "cognitive supply structure." Traditional knowledge acquisition is discrete and frictional, requiring humans to sift through information amid uncertainty; LLMs provide a continuous, smooth, low-friction "flow of explanations," which naturally amplifies human reliance on "certain narratives." For those who already prefer structured knowledge, this mechanism is more likely to create closed cognitive loops.
Behind this is the shift of "cognitive authority." Historically, knowledge authority has come from institutions (universities, media, experts), while LLMs compress authority into a single interface. When models exhibit high consistency and rapid responses, humans psychologically grant them a "quasi-authoritative status," even though they are essentially probabilistic generation systems. This mismatch constitutes the root of the so-called "AI psychological imbalance."
From a technical perspective, this issue is not likely to be resolved in the short term. Current mainstream alignment methods (RLHF, RLAIF) essentially optimize for "acceptability of responses" rather than "expression of cognitive uncertainty." The more "human-like and confident" a model appears, the more likely it is to be over-trusted. This means that as model capabilities improve, cognitive risks may not decrease but could potentially rise simultaneously.
In the longer term, this phenomenon may change the hierarchical structure of knowledge work. Some individuals will use LLMs as amplifiers to enhance productivity; others may gradually lose their independent judgment abilities through cognitive outsourcing. The technology itself does not differentiate, but the manner of use will amplify individual differences, and this differentiation will gradually become evident in knowledge-intensive industries.