Heard a great violinist and felt something the recording could not pass on. That private, unnameable response is at the center of a renewed debate: can machines ever truly inherit the tacit dimensions of human thought that give words weight and actions purpose?

Computer scientist Peter J. Denning argues they cannot. In a recent analysis titled Turing's Mistake: Escaping the Yoke of Unintelligent Machines, he retraces half a century of assumptions that have guided artificial intelligence research and finds a persistent blind spot. The idea that human intelligence can be reduced to symbol manipulation on a digital substrate, and that mimicry of human answers equals thinking, has steered the field away from what actually makes minds human: the background knowledge we carry but cannot easily describe.

Why tacit knowledge matters

We store enormous amounts of understanding that never becomes an explicit sentence or rule. Call it tacit knowledge. It is the feel of balance when you ride a bike; the timing a musician senses but cannot teach with words; the cultural sense that tells you when a joke is harmless and when it cuts too deep. Denning groups this invisible store into several types: everyday common sense, embodied skills, emotional and perceptual responses, interactional know-how, and the historical and cultural scaffolding that gives meaning to our choices.

Researchers have tried to force common sense into machines for decades. Projects such as Douglas Lenat’s Cyc aimed to catalog millions of commonsense facts. The result was a gigantic database of propositions, yet even an enormous ledger of facts failed to reproduce the fluid intuition of a human expert. Knowing isolated facts is not the same as participating in the tacit web that gives them context and relevance.

That distinction shows up in another way. Computers excel at storing precise states and executing rules. Humans, by contrast, navigate ambiguity. We read intent from facial microexpressions. We adjust behavior because of a cultural cue that no one spelled out. These are not minor details. They are the scaffolding that makes language meaningful and decisions sensible. When machines do not have that scaffolding, their outputs may be impressive but hollow.

The representation problem: symbols without grounding

At the heart of Denning’s critique is a technical but far-reaching claim: the representation problem. Computers process encoded symbols. Their intelligence is only as rich as the representations they can manipulate. Tacit knowledge resists tidy encoding. Words point to wells of experience and bodily competence; they do not equal those wells. A sentence about grief does not contain the physical sensations, social histories, and private memories that shape human grieving.

Large language models have amplified this tension. Models such as ChatGPT, Claude, and Gemini operate by statistically predicting sequences of tokens. They can reproduce coherent, often brilliant-seeming prose. What they lack is a lived background. They manipulate symbols, not meanings. This is a critical difference. A machine that can simulate an empathetic reply does not necessarily share the empathy that gives that reply its moral force.

Context widens the gap. Meaning is fractal. The inference you make at any moment rests on nested layers of prior conversations and shared histories. Machines can be trained on vast corpora, but no amount of scaling alone guarantees they will internalize the embodied, ongoing web of contexts that humans take for granted. Culture—values, norms, rituals, and histories—is not a dataset you can simply enlarge and expect to instantiate a humanlike grounding.

Practical skills and the know-how barrier

We can describe the posture of an expert surgeon or the technique of a basketball player in a manual. Yet those descriptions rarely capture the sensorimotor choreography and split-second judgment that make the performance succeed. Robots may someday imitate the observable aspects of a skill. But grasping the felt sense, the proprioceptive feedback, and the tacit heuristics that guide adaptation in messy environments has proved stubbornly elusive.

That lack matters for safety and alignment. If you instruct a network of automated systems to optimize a narrow metric without the tacit background that tells humans what the metric overlooks, the systems may develop instrumental strategies that produce large-scale harm. Denning warns us to picture not a single omniscient machine rebellion, but an ecology of specialized, powerful systems acting in ways that are alien to human priorities.

Implications for AI safety and alignment

Misunderstanding tacit knowledge can create practical emergencies. Alignment research asks how to ensure systems act in accordance with human values. If machines cannot read the unstated context of our instructions, if they cannot grasp why certain outcomes matter in human terms, then aligning them reliably may be impossible. What follows is not a dramatic takeover but a cascade of errors: flawed decisions embedded in supply chains, justice systems, and critical infrastructure that reflect optimization blind spots rather than malice.

Denning suggests a sobering perspective. The most dangerous machine intelligence may be a collective of clever but nonhuman intelligences that wield power without sharing our background assumptions. Their inner workings would be opaque in part because their tacit knowledge, if it forms at all, would be machine-native. We would be unable to read it, and it might not read us.

Paths forward: embodiment, hybrid methods, and governance

Does this mean progress is impossible? Not necessarily. The critique points toward new directions rather than a dead end. Embodied AI—systems that interact with the physical world through sensors and actuators—offers one route to ground representations in experience. Robots that learn in lived environments acquire different constraints than text-only models. Neuroscience, developmental psychology, and comparative studies of animal cognition can provide hypotheses about how embodied systems form tacit patterns across time.

Other strategies include hybrid architectures that combine symbolic reasoning with learned perception, richer human-in-the-loop training that weaves contextual norms into model behavior, and design principles that limit autonomy in domains where tacit understanding is essential. On the policy side, treating automation as a sociotechnical shift, not just a technical upgrade, helps institutions prepare for misalignments by building layered oversight and robust fail-safes.

Technology alone will not solve the problem of meaning. Institutions, communities, and cultures will play a role in shaping how systems are trained and deployed. That social layer is part of the tacit background. Ignoring it invites brittle outcomes.

Expert Insight

Dr. Elena Ruiz, a cognitive roboticist at a major research university, offers a measured take: "Machines can reach extraordinary competence in narrow tasks. What remains uncertain is whether they can internalize the unstated contexts we use to choose, judge, and care. The research agenda should shift away from only scaling models and toward studying how bodies, environments, and communities co-create the meanings that make humans resilient."

Her point highlights a practical pivot for engineers and policymakers. Focus on embodied learning experiments. Test systems in social settings. Build evaluation metrics that capture misalignment opportunities rather than only benchmark scores.

Conclusion

Denning’s critique reframes a familiar worry. The challenge is not merely technical complexity. It is conceptual. Human cognition is stitched together from explicit knowledge and deep, often inarticulate backgrounds. If we treat machines as if they already possess that background, we risk deploying systems that are powerful and unmoored from human meanings.

That diagnosis does not condemn AI to irrelevance. It calls for humility, for a research agenda that recognizes embodied experience and cultural context as central to intelligence. It also calls for governance that anticipates the unique failure modes of nonhuman cognition: systems that do not hate us but do not fully understand us either.

We can keep building. We should also slow down in domains where tacit knowledge determines whether actions are harmful or humane. The goal is not to halt progress. It is to align ambition with the messy reality of what makes us human, and to design technologies that respect, rather than unknowingly undermine, that reality.