The Quiet War Between Human Thought and Machine Optimisation

Artificial intelligence can write essays, crack jokes, answer complex legal questions, and produce clean code in seconds. It does all of this without fatigue, without bad days, and without emotional baggage. On the surface, the comparison with human thinking looks increasingly unflattering for us.

But the surface is misleading. The more researchers examine how AI systems actually work, the clearer it becomes that what these tools do is fundamentally different from what human minds do — and that confusing the two carries serious consequences.

How AI Actually Thinks

When an AI model receives a prompt, it encodes the input as a series of numerical vectors called tokens. These pass through multiple processing layers where the model calculates which tokens are statistically relevant to predicting the next one. The result is not understanding in any meaningful sense. It is pattern completion — an extraordinarily sophisticated form of autocomplete, trained on a vast corpus of human-generated text.

Human cognition works differently at almost every level. The brain processes information through billions of neurons operating concurrently, with different regions specialised for different functions. Where AI works at the level of tokens, humans work at the level of ideas — grasping chunks of meaning and linking them to prior knowledge, context, and lived experience.

A 2026 review in Human Behavior and Emerging Technologies, drawing on cognitive science and AI research, characterises this as the difference between statistical intelligence and embodied intentionality. The review finds that while AI and human cognition share certain information-processing foundations, they diverge fundamentally on embodiment, continuity of self, and phenomenological grounding — aspects of cognition that text-based training cannot replicate.

Memory is another point of divergence. Humans maintain multiple memory systems — sensory, working, and long-term — all linked associatively by meaning, context, and emotion. An AI model’s knowledge, by contrast, is encoded in its weights during training and does not update with use. Its working memory, known as the context window, is a fixed sequence of tokens under current consideration. Once that window fills, everything outside it is entirely forgotten.

Learning follows the same pattern. A human can form a lasting memory from a single exposure to a new idea. A model may need to encounter a concept thousands of times in training data before it can use it reliably. And once training ends, the model’s parameters are largely static. Humans are dynamic, constantly adjusting as new information arrives.

The Hallucination Problem

One of the most discussed limitations of current AI is its tendency to hallucinate — producing confident, fluent, entirely incorrect statements. This is not a bug that will simply be patched away. It is structural.

Because large language models generate responses by predicting plausible token sequences rather than reasoning from principles, they can fail at tasks that seem trivial to humans. For years, leading models could not correctly count the number of times the letter “r” appears in the word “strawberry.” The model produces what looks like reasoning without performing it.

A more accurate human analogue is confabulation — a term from neuropsychology describing when a person unknowingly constructs a false memory or explanation, genuinely believing it to be true. The brain has a natural tendency to fill in missing details. AI does something structurally similar, but at a scale and speed no human confabulator could match.

The practical stakes are not trivial. Research cited in Psychology Today found that AI-generated summaries from major search platforms are inaccurate between 10 and 28 percent of the time. With an estimated five trillion searches conducted annually, the volume of quietly accepted misinformation is substantial.

The Cognitive Cost of Convenience

There is a quieter concern running beneath the hallucination debate: what happens to human thinking when we routinely hand it over.

Neuropsychologist Umberto León Domínguez of the University of Monterrey warns of “cognitive offloading” — the use of AI in place of the mental work humans would otherwise do themselves. The concern draws on a well-documented parallel: research has consistently shown that heavy reliance on smartphone GPS weakens spatial cognition and navigational memory. The worry is that AI may do the same to a broader range of cognitive abilities.

The empirical evidence is beginning to accumulate. A four-month MIT study involving 54 participants found that writing essays with ChatGPT significantly reduced cognitive engagement compared to writing without it — participants showed markedly lower neural activity, suggesting the intellectual effort needed to transform information into knowledge was being bypassed rather than supported. A separate Microsoft study on 319 knowledge workers found a significant negative correlation between frequency of AI tool use and critical thinking scores: as trust in the system rose above trust in one’s own abilities, the tendency to offload mental effort increased.

A 2025 paper in the journal Information introduced a Cognitive Sustainability Index to model this risk, finding that AI interaction exists on a spectrum from cognitive atrophy — where over-delegation erodes analytical autonomy — to cognitive synergy, where reflective engagement with AI tools actually develops metacognitive skills. The determining variable is not which tool is used, but how actively the human remains in the loop.

The distinction between these two trajectories is not abstract. Consider the professional who uses AI to eliminate routine formatting while sharpening their own judgment on complex decisions — and the one who uses it to avoid the discomfort of thinking through problems they haven’t encountered before. Both use the same tool. One is building capacity. The other is quietly trading it away.

What Should AI Leave for Humans?

Researchers are increasingly asking not just what AI can do, but what it should do — and what should deliberately be left to human minds.

Varun Grover, who holds the Billingsley Endowed Chair at the University of Arkansas, frames this as a choice between two trajectories. The dystopian path follows a cycle of AI improvement leading to greater use, leading to greater dependency, leading to the gradual numbing of human skills. Empathy, judgment, creativity, and moral reasoning become first augmented, then replicated, then simulated. The utopian path looks different: AI improvement leads to discriminating use of intelligence across humans and machines, resulting in higher productivity without the erosion of human agency.

The key variable, in Grover’s analysis, is trust. Societies trust systems that demonstrate empathy, transparency, and reliability. AI is advancing on all three fronts: affective computing mimics empathy, explainability tools simulate understanding, and large training datasets project an air of reliability. But trust manufactured through simulation is not the same as trust that has been earned. When users understand AI’s logic and recognise its limitations, trust enables empowerment. When trust is engineered through opaque systems and simulated warmth, it becomes compliance dressed up as confidence.

In January 2025, the Vatican entered the debate directly, co-issuing a doctrinal note — Antiqua et nova — through its Dicastery for the Doctrine of the Faith. The document concludes that AI must only be used to complement human intelligence, not replace it, arguing that replacement would effectively enslave humanity by substituting machine outputs for genuine human judgment and moral responsibility. It is an unusual voice in a largely secular debate, but its framing reflects a concern shared across ideological lines: that the value of human thought is not merely instrumental.

The Rights Distraction — and Why Framing Matters

As AI systems have grown more sophisticated, a parallel debate has emerged about whether they deserve moral consideration, or even rights. This is not merely academic: the question is attracting serious attention in policy circles.

Eryk Salvaggio, a researcher at the University of Cambridge and affiliated with the Max Planck Institute, argues that this framing is dangerous — and for reasons that connect directly to the cognition debate. Neural networks resemble human neural pathways because they are modelled on them. A model train works like a train because it is a model of a train; that does not make it a transportation network. The resemblance is structural, not experiential.

Salvaggio’s concern is not primarily philosophical. When corporate products are described as capable of experience or inner life, it creates the conditions for extreme deregulation. Rights, once attributed, tend to be treated as unalterable. Extending them to software systems provides AI companies with a mechanism to assert protections on behalf of their products, short-circuiting democratic debate about how those products are regulated.

This is where the cognitive and political debates converge. If we confuse simulation for reality in how AI thinks — treating fluent text generation as genuine understanding — we risk making the same category error in governance: treating simulated experience as grounds for legal protection, while the humans and communities affected by AI deployment remain the uncounted variable.

The regulatory imagination, Salvaggio argues, should be focused on people, not algorithms. The source of a system’s influence is not the machine itself but those who produce it.

The Embodiment Gap

Perhaps the most fundamental difference between human and machine cognition is one that rarely enters the public conversation: embodiment.

Human thought is shaped by physical existence. The concept of wetness is tied to the tactile sensation of water. Knowledge of gravity comes not from reading about it but from living with it. This grounding in physical reality provides human thinking with a layer of common sense that AI cannot replicate from text alone.

Researchers at the University of Sheffield, writing in Science Robotics, argue that the human brain has evolved through embodiment in a physical system that directly senses and acts in the world. Current AI systems have no bodies and no direct connection to physical reality. Their knowledge is entirely second-hand — derived from descriptions written by people who actually inhabit the world being described.

The cognitive science literature frames this as the symbol grounding problem: how do words and concepts gain meaning? A computational system that connects abstract symbols only to other abstract symbols cannot, by its structure, ground those symbols in experience. The consequence is an AI that can discuss the physics of a fall with precision while having no conception of falling.

A 2024 theme issue of the Philosophical Transactions of the Royal Society B brought together researchers from across disciplines to examine this directly. Their collective conclusion was that embodied cognition — the view that mind and body are not separable systems but continuous ones — poses a fundamental challenge to the idea that general intelligence can emerge from text prediction alone. Large language models, one contributor notes, may be best understood not as approximations of human intelligence but as exaggerations of one specific aspect of it: language-based, disembodied pattern association.

Where This Is Heading

The honest answer is that nobody knows with confidence. Amara’s Law, named after futurist Roy Amara, holds that we tend to overestimate the effect of a technology in the short term and underestimate it in the long term. It applied to electricity, to GPS, and to semiconductors. It almost certainly applies to AI.

What researchers do broadly agree on is the direction of the risk. Cognitive atrophy through uncritical offloading. The erosion of political and regulatory will to govern AI through confusion about what it actually is. The slow displacement of embodied, morally-grounded human judgment by systems that can simulate its outputs without any of its foundations.

None of these outcomes are inevitable. The same body of research that documents the risks of cognitive offloading also identifies the conditions under which AI use produces the opposite effect — deeper engagement, sharper metacognition, expanded capacity. The variable, consistently, is whether humans remain actively in the loop: questioning, evaluating, and retaining the ability to reach their own conclusions.

The war between human thought and machine optimisation is not a battle over processing speed or vocabulary size. It is a slower, quieter contest over which kind of intelligence we decide to trust — and what we are willing to give up in exchange for the convenience of deferring to the other. The research does not settle the question. But it clarifies, finally, what is actually being asked.

Key Sources

• Tremblay et al. (2026). “Shared Minds: The Cognitive Parallels Between Humans and Artificial Intelligence.” Human Behavior and Emerging Technologies, Wiley.
• Barrett, L. & Stout, D. (2024). “Minds in movement: embodied cognition in the age of artificial intelligence.” Philosophical Transactions of the Royal Society B, 379(1911).
• Prescott, T. (2023). “Embodied AI: Bridging the Gap to Human-Like Cognition.” Human Brain Project / Science Robotics.
• Mogi, K. (2024). “Artificial intelligence, human cognition, and conscious supremacy.” Frontiers in Psychology, 15:1364714.
• Grover, V. (University of Arkansas). Research on trust, AI dependency, and human agency.
• Salvaggio, E. (University of Cambridge / Max Planck Institute). Research on AI rights framing and regulatory implications.
• Domínguez, U.L. (University of Monterrey). Research on cognitive offloading and AI chatbot use.
• MIT Media Lab (2024). Four-month EEG study on neural engagement during AI-assisted essay writing (preprint).
• Microsoft (2024). Study of 319 knowledge workers on AI tool use and critical thinking scores.
• Dicastery for the Doctrine of the Faith & Dicastery for Culture and Education (2025). Antiqua et nova. Vatican City.
• MDPI Information (2025). “Cognitive Atrophy Paradox of AI–Human Interaction.”
• Psychology Today (2025). “Your Brain on AI: Cognitive Offloading, Debt, and Atrophy.”

Ex Nihilo magazine is for entrepreneurs and startups, connecting them with investors and fueling the global entrepreneur movement