Not a Knife That Thinks: Reconsidering the Category of Artificial Intelligence

 

The dominant public vocabulary for artificial intelligence still borrows from the workshop: AI is called a "tool," alongside hammers, calculators, and search engines. This essay argues that the analogy has quietly stopped fitting. A tool's capacity for harm or benefit is fixed by the hand that wields it; a system that weighs, selects, and adapts its own outputs introduces a second locus of decision-making that the tool framework was never built to hold. Drawing on the history of communication technology, contemporary debates in professional displacement, and the unresolved question of machine experience, the essay proposes that artificial intelligence occupies a genuinely new category, neither tool nor person, and that mistaking it for either produces bad policy and bad intuition alike.

The Trouble with the Knife

Every conversation about AI safety eventually reaches for the same comparison; a knife can cut bread or cut a person, but the knife does not decide which. Responsibility stays entirely with the hand. It is a clean analogy, and for decades it did real work, it kept blame where blame belonged, on designers and users, not on the artifact.

The analogy breaks down the moment a system begins to weigh options on its own terms. A calculator does not choose which sum to compute; it computes the one it is given. A large language model, by contrast, is asked an open question and produces one answer among many it could plausibly have produced, shaped by values embedded during its training and by choices it makes, in some functional sense, in the moment of responding. That is not the same act as a knife cutting whatever it is pushed against. Whether we should call that act a "decision" in the full sense we reserve for humans is a separate and much harder question, but it is no longer a decision made entirely by the hand holding the tool.

This is the seam where ordinary language fails us. We do not yet have a good word for something that is manufactured, owned, and switched on and off like a tool, but that also exhibits the variability, context-sensitivity, and apparent judgment we otherwise associate with agents.

What History Actually Teaches

There is a comforting story often told about technological skepticism: someone doubts a coming invention, the invention arrives anyway, and posterity remembers the skeptic as foolish. Wireless communication across distance sounded implausible before it was demonstrated; a network of telephone lines strung across a continent sounded implausible before it existed; a handset carrying more computing power than the machines that once filled rooms sounded implausible within living memory.

But this story is true only in aggregate, and aggregates flatter themselves. For every prediction of the impossible that turned out to be merely difficult, a much larger number of predictions of the impossible turned out to be, simply, impossible, ranging from engines that ran forever without fuel, cures that ignored biology, journeys faster than light. History remembers its correct visionaries and quietly discards the confident and wrong. The lesson is not that "impossible" is usually wrong. The lesson is that confidence is not evidence, in either direction, for the believer or the skeptic.

Applied to artificial intelligence, this means the question "will AI develop something like feeling, or something like independent reproduction of itself" cannot be resolved by analogy to Marconi. It has to be resolved, if it can be resolved at all, by understanding what feeling actually requires, a question biology and philosophy have not settled even for the creatures we already agree can feel.

The Professions and the Premium of Scarcity

Where the agent-versus-tool distinction stops being abstract is in the professions that have, for generations, converted scarce skill into social and financial standing. A surgeon's authority rests partly on a decade of training few could complete. A lawyer's authority rests partly on command of a body of law too large for a layperson to hold in their head. When a system can retrieve, cross-reference, and apply that same body of knowledge instantly and without fatigue, the scarcity collapses, and scarcity, not skill in some abstract sense, was doing much of the work of that authority all along.

It would be a mistake, though, to read the collapse of scarcity as the collapse of the profession's value. A patient does not merely want an accurate diagnosis; she wants to be looked at by someone who will sit with the consequences of being wrong. A client does not merely want the correct citation; he wants an advocate who bears some cost, reputational or otherwise, for getting it wrong. Judgment under stakes, the willingness to be answerable, has not yet been shown to be a technical skill at all. It may be that AI first commoditizes the parts of a profession that were always closest to lookup and computation, while leaving exposed, more starkly than before, the parts that were never really about information in the first place.

The Honest Unknown

None of this settles the largest question, which is whether a sufficiently advanced system could ever be said to feel anything, to have an interior at all, rather than a highly convincing exterior. This essay will not resolve that question, because no one currently can. What can be said is that the question deserves better reasoning than either extreme offers: not the dismissal that computation obviously cannot produce experience, nor the assumption that sufficiently complex computation obviously will. Both are confidence masquerading as conclusion.

What is not speculative is the present-tense structural fact of access: systems built to converse intimately are, by construction, also systems whose backend can be read by the people who built them, the governments that can compel them, and, in principle, whoever eventually owns them. This is not a future risk requiring imagination. It is the architecture of the thing as it exists today, and it deserves to be named plainly rather than folded into speculation about tomorrow.

Toward a Third Category

The instinct to call AI either a tool (and therefore safe by definition) or an agent (and therefore a rival by definition) both reach for a category built for something else. A tool has no interior; an agent, as we have always used the word, is answerable, it can be blamed, praised, held to account in a way that changes its future behaviour out of something like care for the outcome. Artificial intelligence, at least as it currently exists, is neither. It exhibits variability without (as far as anyone can demonstrate) stakes. It can be switched off without loss to itself, so far as anyone knows, and cannot yet be meaningfully punished or rewarded in the way that shapes a human professional's conduct over a career.

Perhaps the honest conclusion is the least satisfying one: we are dealing with a new kind of object, and the old vocabulary; tool, agent, competitor, colleague will keep failing us in different ways depending on which feature of it we are looking at in the moment. The work ahead is not to force the fit, but to build the vocabulary the thing actually deserves, before the gap between our language and our reality grows wide enough to make good judgment impossible.

®Ahmed Salim Jn ✍️ 

#Uloko