The Promise of Artificial Intelligence: Reckoning and Judgment
Source: https://mitpress.mit.edu/9780262043045/the-promise-of-artificial-intelligence/ ↗
A philosopher-computer scientist with fifty years in the field argues that the difference between calculation and judgment is categorical, not a matter of degree.
Smith distinguishes between reckoning — the formal manipulation of representations that computers excel at — and genuine judgment, which requires registration with the world, commitment, and the kind of embodied understanding that current AI systems lack.
The book is subtle and important: it does not deny AI's power but insists on precision about what that power is and is not.
For product leaders deploying AI, Smith provides the conceptual vocabulary to think clearly about where machine capability ends and human judgment begins.
Central argument
Brian Cantwell Smith argues that current AI systems are powerful reckoners — they manipulate formal representations with extraordinary speed and accuracy — but they do not exercise judgment, which requires genuine registration with the world, embodied commitment, and the capacity to take responsibility for a stance. The distinction is categorical, not a spectrum: no increase in computational power or data scale closes the gap between reckoning and judgment. Consequently, Smith contends that framing AI as 'intelligent' in any full sense is not merely imprecise but actively misleading, because it obscures what kind of thing these systems actually are.
Critique
Smith's framework, developed primarily through philosophical analysis, risks becoming difficult to falsify: if judgment is defined by properties that computational systems are structurally excluded from possessing, the argument is secured by definition rather than by empirical engagement with what frontier systems actually do. A thoughtful critic might press him on whether 'registration with the world' and 'embodied commitment' are genuinely necessary conditions for judgment, or whether they reflect a particular philosophical tradition — broadly phenomenological — that not all serious accounts of cognition would accept. The book was also completed before the generative AI wave, which does not invalidate its framework but does mean it never confronts the cases that most stress its distinctions.
Why it matters for product
For a CPO deploying AI in product workflows, Smith's reckoning/judgment distinction provides a precise diagnostic tool: tasks that can be fully specified as formal operations on representations are good candidates for AI automation, while decisions that require contextual commitment — prioritisation calls, strategic framing, stakeholder alignment — cannot be delegated without losing something categorical, not just something marginal. This has direct implications for team design: augmenting researchers or PMs with AI reckoning tools is not the same as replacing their judgment, and conflating the two leads to accountability gaps that surface painfully during product failures. Smith also implicitly warns against metric-driven product cultures that mistake high-confidence AI outputs for grounded understanding, which is a live risk when teams interpret model confidence scores as proxies for truth.