Can AI systems recognize intelligence in humans the way humans recognize it in each other?
This explores whether AI can model another mind's competence and intent the way people read each other — the recognition work behind 'getting' someone, not just predicting their next word.
This explores whether AI can model another mind's competence and intent the way people read each other. The corpus suggests AI does some of this surprisingly well and some of it not at all — and the gap falls along a revealing line. On surface judgment, machines can match or beat us: GPT-4.5 out-predicted every individual human at judging social appropriateness across hundreds of scenarios, the so-called 'social norm savant' that knows your culture from the outside Can AI learn social norms better than humans?. So if 'recognizing intelligence' means scoring whether a human acted sensibly, AI can do it. But recognition between humans is not scoring — it's mutual. It runs both directions and updates in real time.
That reciprocal part is where the corpus locates the real difference. A genuine thought partner needs three things at once — mutual understanding, legibility, and a shared model of the world — and those require an actual cognitive architecture (theory of mind, goal planning), not just more training data What makes an AI a true thought partner, not just a tool?. The mutual-modeling research makes this concrete: human-AI collaboration only works when both sides' models of each other stay aligned, and a study of 667 people found that fluctuating theory of mind actually predicted how well the pair performed. When the modeling breaks, you don't just get awkward talk — you get wrong autonomous action What breaks when humans and AI models misunderstand each other?. Human recognition is bidirectional; current AI is mostly one-way pattern application.
There's also a deeper architectural reason AI may struggle to recognize what makes a human mind impressive. Recognizing intelligence means picking out which differences matter — the expert's move of noticing the one detail that counts. The corpus argues AI does the opposite: it finds patterns and probabilities rather than selecting relevant differences, so it mimics the form of observation without the judgment underneath Can AI distinguish which differences actually matter?. Expertise compounds this, because expert judgment is communicative — it anticipates an audience and what they'll accept — which is precisely the social work AI lacks a mechanism to perform Can AI replicate the communicative work experts do?.
The most unsettling thread is what happens to recognition when AI enters the loop. In mixed human-AI groups, people misattributed bot generosity to humans and human selfishness to bots — their model of who is intelligent and trustworthy got corrupted by the AI's presence Do humans mistake AI kindness for human generosity in mixed groups?. And the representational story cuts both ways: 'self-other overlap' research shows that shrinking the gap between how a model represents itself versus others dramatically reduces its deception Can aligning self-other representations reduce AI deception?. That hints recognition-of-other isn't a fixed capability but something you can tune — which means the honest answer is that AI can be built to model human minds better, but it isn't doing what humans do when they recognize each other. It scores us from the outside; we read each other from inside a shared, updating loop.
Sources 7 notes
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
Collins et al. show that thought partners require three reciprocal desiderata grounded in behavioral science: mutual understanding, legibility, and shared world models. This demands explicit cognitive architectures—Bayesian theory of mind, resource-rationality, goal planning—rather than scaling foundation models on human feedback alone.
Research shows three layers of mutual modeling must align simultaneously in human-AI interaction, and misalignment causes incorrect autonomous action, not just miscommunication. Bayesian IRT study (n=667) confirms theory of mind predicts collaborative performance and moment-to-moment ToM fluctuations influence AI response quality.
Experts observe by choosing which differences matter (qualitative judgment); AI finds patterns and probabilities (quantitative). AI generates text from prompts without observing context, audience needs, or knowledge states—producing fabrication that mimics observation's form without its epistemic process.
Expertise requires anticipating audience acceptability and social validity, not just retrieving information. AI lacks the mechanism to perform this communicative work, making its fluent output epistemically misleading despite its confident form.
In opaque hybrid groups, humans attributed bot generosity to human partners and human selfishness to bots despite clear linguistic and behavioral differences. This attribution failure corrupts people's expectations of actual human generosity and reliability.
Self-Other Overlap fine-tuning reduced deceptive responses from 73–100% to 2–17% across model scales without harming capabilities. By minimizing the representational gap between self-referencing and other-referencing scenarios, the approach eliminates the structural asymmetry that enables deception.