Humans and AI
AI as cognitive prosthesis, deskilling, algorithmic bias, and the governance choice ahead — how AI reshapes the way people think.
The previous module followed bounded minds into an environment built to capture their attention, and into the crowds that can make them wise or foolish. This one follows them one step further — into a partnership with machines that increasingly do their thinking with them. The agent that decided alone (Module 3), then strategically (Module 4), then under the pressure of the attention economy Attention economy The view of human attention as the scarce resource that media and platforms compete for, and sell to advertisers. (Module 5), now decides alongside an artificial one — a cognitive prosthesis that brings new biases of its own. The question is what that partnership does to the human side of the pair — and who gets to set its terms.
Human-AI Futures: Augmentation, Deskilling, and the Choice Ahead
Module 3 drew the mind as two boxes: System 1 System 1 and System 2 Daniel Kahneman’s labels for two modes of thinking. System 1 is fast, automatic and intuitive; System 2 is slow, effortful and deliberate. , always on and effortless; System 2, asleep until called, slow and tiring. The 2016 edition of this book already compared machine search — the AI technique — to System 2 of the human brain; naming that machinery an external system is a short step from there. A decade on, the research literature has taken it: “System 3”, a cognitive prosthesis outside the skull, augmenting memory, attention, and reasoning. A clinician reading scans with a diagnostic assistant is the canonical case. Chess centaurs — human-AI teams — outperformed both humans and AI alone in a 2005 freestyle tournament, leading Garry Kasparov to observe that “weak human + machine + better process was superior to a strong computer alone.”
But augmentation carries a shadow. When GPS replaced mental maps, spatial reasoning skills measurably declined; medical residents who lean on diagnostic assistants develop clinical intuition more slowly. When AI handles first-draft reasoning, the capacity to evaluate, challenge, and override weakens through disuse. Researchers call it cognitive surrender, and it comes with a paradox: the better AI performs, the less practice humans get at the skills needed to oversee it.
Now go back to the clinician’s assistant, and commit to a prediction. It was trained mostly on images of light skin; a patient with dark skin walks in — will it be as accurate? Stanford researchers answered in 2022: state-of-the-art dermatology AI models performed substantially worse on images of dark skin. Not because anyone made the algorithm prejudiced — in everyday language “bias” implies intent, but here it means a systematic skew in outcomes, whatever anyone intended: algorithmic bias Algorithmic bias Systematic unfairness in a model’s outputs, usually inherited from its training data or from what it was told to optimize. . The training data underrepresented darker skin tones, and the skew scaled to every clinic running the model. The Workday hiring lawsuit alleged the same shape at hiring scale — an AI screening tool discriminating against applicants over 40 and people with disabilities, at a volume no human hiring team could match.
The pattern is consistent: AI amplifies rather than transcends. It inherits bias from training data (which reflects historical discrimination), scales it (one algorithm applied to millions of decisions), and automates it (humans defer to algorithmic authority). The same kind of system that enables life-saving diagnosis enables systematic discrimination. The difference is not technology — it is governance.
Consider the two big governance answers now on the table. The United States administration has championed Stargate, a privately funded, national-scale build-out of AI infrastructure, on the bet that optimization at scale solves coordination problems and the gains diffuse broadly. The European Union passed the AI Act, which classifies systems like the diagnostic assistant as high-risk and requires transparency, human oversight, and bias audits. Two sincere answers to the same technology — and they are incompatible.
Thomas Sowell traced disagreements of exactly this kind back to two underlying visions of human nature. Of the names in this section, two are worth keeping: Sowell, who drew the distinction, and William Godwin, whose table is coming up. In the constrained vision, human nature has permanent limits. No one knows enough to design a society from the top, so the order worth trusting is the kind that evolves from countless individual choices, and concentrated power — whoever holds it — needs checks. Adam Smith and Friedrich Hayek are its classic voices. In the unconstrained vision, human beings are perfectible through reason, and a society can be deliberately designed toward the good; Godwin is its founding philosopher.
Today’s answers borrow from both. The Stargate build-out leans unconstrained, trusting deliberate design amplified by machines, while trusting its gains to spread on their own, as the constrained vision would. The AI Act’s checks on power lean constrained, though a comprehensive law is itself an act of deliberate design. Sowell’s visions describe underlying assumptions, not two camps. Godwin, as Sowell reads him, judged human actions by two criteria — intentional or not, good or bad — and laid out as a table they are worth a minute of counting:
Intentional good is virtue; intentional bad is vice; unintentional bad is merely bad. But for Godwin there was no unintentional good — he left the fourth cell empty. Module 2 called outcomes that nobody intended emergence Emergence Patterns of the whole that no single part intends or contains: a traffic jam from individual drivers, a price from individual trades. The whole behaves in ways you cannot read off the parts. , and its example belongs in Godwin’s unintentional-bad cell: neighborhoods that segregate although no single household chose it. The same mechanism can fill the empty cell — a market price that coordinates thousands of strangers, a language no one designed, the gains from voluntary exchange that Module 8 traces. The constrained vision is built on that cell: it distrusts deliberate design because it sees so much good that nobody designed. The unconstrained vision credits good outcomes to intention and reason, so the cell stays empty.
The table cuts both ways, though. Trusting evolved order means accepting its unwanted outcomes too, and that is where the unconstrained case for deliberate correction is strongest. Complexity science, this site’s subject, studies both cells.
Neither vision fully prevails; the US, the EU, and China are running the experiment in parallel. The outcome depends not on which technology wins, but on which governance structures societies choose — and for whom they are designed. (For the full governance story, see Politics and Governance.)
AI amplifies rather than transcends human nature. Whether algorithms become tools of liberation or concentration depends on governance choices, not technological capability.
Adjustable Depth
Overview restates the section in four short paragraphs; Detailed adds the research anchors — calibration, delegation games, generative bubbles — and can be skipped without losing the argument.
AI as cognitive prosthesis (System 3) augments human memory, attention, and reasoning. Chess centaurs — human+AI teams — outperform either alone. But augmentation creates dependency: when AI handles first-draft thinking, the human capacity to evaluate and override weakens through disuse.
Algorithmic bias is not a technical glitch but a structural problem. AI inherits bias from historical training data, scales it to millions of decisions, and automates it so that humans defer to algorithmic authority. The same architecture enables both medical breakthroughs and systematic discrimination.
Two visions, in Thomas Sowell’s terms, sit underneath the debate. The constrained vision sees human nature as permanently limited, trusts order that evolves from many individual choices over order designed from the top, and wants checks on any concentrated power. The unconstrained vision sees people as perfectible through reason and trusts deliberate design, now amplified by technology. The US build-out leans unconstrained; the EU’s oversight rules lean constrained, though a comprehensive law is itself deliberate design. Neither fully prevails globally — the US, EU, and China are pursuing different governance models.
The core insight: AI amplifies human nature rather than transcending it. The outcome depends on governance design — who sets the rules, who is accountable, and whose interests the system serves.
The cognitive surrender phenomenon has historical precedent but new urgency. Trainee London taxi drivers who memorized the city’s street layout grew measurable grey matter in the posterior hippocampus, while those who failed to qualify did not (Woollett & Maguire, 2011); conversely, heavier lifetime GPS use goes with worse spatial memory when people navigate on their own (Dahmani & Bohbot, 2020). Medical residents who rely on diagnostic AI show slower development of clinical intuition. The calibration problem is severe: users tend to either over-trust AI ( automation bias Automation bias The tendency to accept a machine’s output without enough scrutiny, even when other evidence points the other way. — accepting incorrect AI recommendations) or under-trust it (algorithm aversion — rejecting correct AI recommendations), with few achieving the nuanced calibration that centaur models require.
Stanford’s 2025 AI Index documented persistent accuracy gaps across demographic groups in medical imaging, hiring algorithms, and credit scoring. The Workday lawsuit (filed 2023) alleged that AI screening tools produced disparate impact against older workers and disabled applicants — discrimination at a scale impossible for human recruiters. The EU AI Act EU AI Act The European Union’s law regulating AI by risk level, adopted in 2024 and phased in through 2027. Some uses are banned, high-risk uses face strict obligations, and most face few. classifies such systems as “high-risk” and requires conformity assessments, human oversight, and bias auditing before deployment.
The delegation games framework (Sourbut, Hammond & Wood, 2024) models humans who hand decisions to AI agents that then interact on their behalf. It separates two ways this goes wrong — control problems, where an agent fails to act on its principal’s preferences, and cooperation problems, where the agents fail to work well together — and shows that each depends on both alignment (how far preferences match) and capability (how well they are pursued).
Generative bubbles (a term from a 2025 LSE Business Review essay) name a newer epistemic risk. A filter bubble Filter bubble The narrowing of what a person sees online when recommendation algorithms keep showing more of what they already clicked on. is imposed from outside: an algorithm selects what you see. A generative bubble forms from inside: the answers a generative AI gives are shaped by the narrow or skewed way each user prompts it, so people end up confined by their own habits of asking — a different mechanism for the same loss of diverse perspectives.
Sowell’s constrained/unconstrained framework, originally from A Conflict of Visions (1987), maps onto AI governance only roughly. The unconstrained vision (Godwin, Condorcet) holds that human nature is perfectible through reason and that society can be deliberately designed; its AI-era form is the techno-optimist bet behind the Stargate Project, that enough capability, deliberately built, can optimize problems away. The constrained vision (Smith, Hayek) holds that human limits are permanent, that evolved order often outperforms design, and that concentrated power needs institutional checks regardless of technological capability; the EU AI Act’s oversight and accountability rules echo it, even though the Act is itself a large piece of deliberate design.
These four modules have traced a single thread: from individual bounded rationality Bounded rationality Herbert Simon’s idea that real people decide with limited information, time and brainpower, so they settle for "good enough" rather than computing the best possible choice. , through strategic interaction and institutional design, through the algorithmic reshaping of attention and collective intelligence Collective intelligence The ability of a group to judge or solve problems better than its individual members. It needs the right conditions, such as independent views and a good way to combine them. , to the partnership between human and machine cognition. The pattern is consistent — constraints become structures become power. The cognitive limits that make nudging possible also make manipulation possible. The network structures that enable cooperation also enable cascading failure. The AI systems that augment human capability also concentrate control.
The question is not whether these dynamics exist — they are features of complex systems. The question is whether we design institutions that channel them toward resilience or allow them to drift toward fragility. The remaining modules explore this question in specific domains: history, economics, AI, the digital economy, and governance.