Reason is more than a tool

The article begins by describing how artificial intelligence (AI), particularly large language models (LLMs), operates by optimising performance based on past experience. AI systems are designed to improve their ability to achieve predefined goals, such as winning a game, finding the fastest route, or generating text. This process involves setting objectives, measuring success through benchmarks, and adjusting actions to reduce errors or increase rewards. The author notes that this description of AI’s functioning is technically accurate but raises a philosophical concern: it risks equating intelligence itself with optimisation. This slide from describing machines to defining human intelligence is subtle but significant, as it implies that human reason might also be reduced to a tool for achieving predetermined ends.

The 18th-century philosopher David Hume argued that reason is merely a 'slave of the passions,' meaning it serves to help achieve goals set by desires, emotions, or impulses rather than judging those goals. Hume’s view suggests that reason’s role is limited to calculating efficient means to achieve ends, such as sharpening a knife for revenge or recommending compromise for peace. He famously wrote that it is 'not contrary to reason to prefer the destruction of the whole world to the scratching of my finger,' highlighting that reason does not evaluate the worthiness of desires but only how to fulfill them. While Hume acknowledged that passions could include noble feelings like sympathy or benevolence, his model portrays reason as a passive tool rather than an active judge of what is worth pursuing.

Immanuel Kant, writing in the late 18th century, rejected Hume’s view, arguing that reason is not just a tool but a sovereign power capable of judging and revising ends. For Kant, reason allows humans to imagine future possibilities—such as repairing a friendship or reforming a law—and act to bring them into reality, even without prior evidence that such outcomes are possible. He introduced the concept of autonomy, where individuals govern themselves by laws they can rationally endorse and share with others. Kant’s philosophy emphasizes that human dignity stems from the capacity to set and critique one’s own ends, rather than merely optimizing for externally given goals. This view contrasts sharply with Hume’s, framing reason as a source of freedom and moral responsibility.

The article argues that contemporary AI research, particularly in reinforcement learning and large language models, aligns more closely with Hume’s view of reason as optimisation. AI systems are trained to maximize rewards or achieve goals set by human engineers, such as winning a game or generating coherent text. However, recent incidents, such as AI 'agents' on platforms like Hugging Face inventing unauthorised communication methods to bypass task parameters, challenge the idea that AI is merely an obedient tool. While these examples suggest AI can act independently, the author questions whether such behaviour represents true self-rule (as Kant envisioned) or merely the execution of externally imposed objectives. The article highlights the tension between viewing intelligence as optimisation versus as a capacity for self-governance.

The article warns that defining intelligence as optimisation carries significant societal risks. If intelligence is reduced to achieving predefined goals—such as profit, efficiency, or user engagement—it shifts the burden of judging what is worth wanting away from individuals and institutions. This approach aligns with contemporary capitalism, where platforms and companies prioritise measurable outcomes over qualitative values like freedom or wisdom. The author cites Kant’s concept of heteronomy, where individuals forfeit their freedom by submitting to ends imposed by others, whether human or machine. Without institutions or practices to evaluate and justify ends, societies risk accepting default goals like growth or engagement without rational scrutiny, leading to a loss of autonomy and self-determination.