Availability Alone Should Not Justify Adopting an AI System

A presumption runs through much of the adoption of AI: what can be automated should be. It is seldom stated, because it seldom needs to be. The function is there, a competitor has switched it on, and whoever hesitates is the one asked to explain.

I draw a hypothesis from this: the availability of a tool can become a reason to use it without anyone having formulated that reason. The mechanism leaves an experimental trace, which I discussed in When an AI Answer Is at Hand, “I Don’t Know” Nearly Vanishes: once an answer is at hand, the threshold at which we decide to answer shifts, and people can feel more confident while becoming less accurate. If this recurs in other contexts, adoption stops being a neutral choice and the default deserves the scrutiny we usually reserve for the exception.

The Nobel laureate in physics Giorgio Parisi shows the other side. He and Francesco Zamponi published in July a proof for a statistical-physics problem open for over ten years, to which Claude contributed the missing idea. Parisi has said that without the “waste of time” of doing the calculations himself he would not have understood the problem deeply enough to steer Claude. The tool was useful, and its usefulness rested on a skill that the unassisted work had built. I add a distinction of my own, which Parisi does not make: solving a problem and recognising which problem is worth solving are different abilities.

In the Phaedrus (274e–275e), Thamus answers Theuth, who presents writing as a remedy for memory, that one person can generate an art but another must judge how useful or harmful it will be to those who use it; the father of letters, out of affection, credits them with the opposite of the power they have. Whoever proposes an invention is inclined to see what it makes possible and tends to overlook what it can take away from those who use it.

The structure repeats today. The promise of a more humane future is made by whoever sells the tool, and the direction depends on the incentives governing its adoption. Knowing that a technology can do something leaves open what happens when it is introduced, who benefits, and which capacity goes unused.

When two firms offer the same capability, the one that keeps a human step can find itself at a disadvantage against the one that removes it, if the market rewards mainly cost and speed. I treat this as a tendency: it depends on the value attached to quality, responsibility and control. It is, however, an incentive large enough to make any limit that rests on the will of a single firm fragile. I examined the same fragility in A Limit Indexed to the Competition.

In Oakland, in the states’ trial over teen safety on social media, a former data scientist from Instagram’s wellbeing team testified that leadership had rejected a proposal to turn “Take a Break” on by default, because the cost to the app’s core metrics was not desirable. Only a small percentage of teenagers used the feature. In the agreement reached with the states, some of the most incisive limits, a one-hour daily cap and a longer night block for under-18s, also depend on other social media adopting similar terms.

The example does not prove how AI companies behave. It illustrates a common problem: a protective measure can be technically available and still go unadopted, because its cost falls on whoever should introduce it. The tool exists, so the problem lies in the criterion by which its activation is decided.

Microsoft cut 4,800 roles in July 2026, stating that they were not being replaced by artificial intelligence, while the same memo adds that AI is changing how work gets done. Meanwhile the 2026 capital-expenditure estimates I cited in The Job No Machine Will Take put capex at $190 billion. The juxtaposition leaves causation open. It does show that a company can decline to replace a person with a machine and still direct a growing share of its resources to computing rather than to labour.

Replacement need not happen task by task. It can happen in budget priorities, in service design and in expectations about how much human work is still needed.

To defuse pervasiveness means to remove availability’s role as a sufficient reason. The first step is to shift the burden of proof onto whoever adopts: to show beforehand, with a benefit measured against a concrete alternative, that the tool is needed. The assessment should cover the immediate result, the costs of verification, responsibility for errors, reversibility, and the skills that risk no longer being exercised, including those needed to keep control.

Existing legal instruments answer narrower questions. Article 27 of the AI Act requires a fundamental-rights impact assessment before certain high-risk systems are deployed by specific actors, including public bodies and private providers of public services, as well as for certain uses in credit and insurance. The provisions for Annex III systems apply from 2 December 2027, under the calendar updated by Regulation (EU) 2026/1744.

Article 22 of the GDPR governs certain decisions based solely on automated processing that produce legal effects or similarly significantly affect the person. It does not, in general, require showing beforehand that a system yields a net benefit over a non-automated alternative.

The distinction matters. A risk assessment and a demonstration of usefulness answer different questions. A system can meet certain requirements and still be superfluous; it can be accurate and make answers harder to verify; it can shorten production time and lengthen the time needed to check the result.

The second step is to introduce common limits, because those that impose a competitive cost may be hard to sustain unilaterally. On 26 August Bill Gates proposed taxing AI tokens and robots, arguing that the tax system nudges companies to replace people with machines, and reserving some jobs for humans. In an interview with Axios he said he could imagine, in a very extreme form, 40% of jobs initially reserved. These are hypotheses, not adopted measures. They have a merit, though: they act on the economic conditions of the decision, before replacement happens on a large scale.

The letter Pacing the Frontier, signed by more than a thousand people from AI labs, illustrates the complementary problem I discussed in The Leash They Cut to Fit. The letter specifies no threshold, and I infer that a brake is credible only if whoever sets the threshold and whoever measures its crossing do not coincide entirely with whoever is supposed to be braked. Otherwise control risks turning into self-certification.

The third step is to separate uses. The Rust project adopted a policy that allows models to answer questions, analyse, distil, check, suggest and review, and strictly restricts their use to produce contributions: LLM-generated code is admitted only if non-critical, high quality, well tested and carefully reviewed, and must be disclosed. The policy distinguishes activities that can benefit from assistance from activities in which the community wants to preserve authorship, responsibility and understanding of the work.

For those who do not belong to an organisation able to set its own rules, an elementary measure remains, the one I proposed in When Nobody Has to Try: write your own answer before reading the machine’s. It does not guarantee avoiding errors, but it keeps visible the distance between your judgment and the tool’s.

None of this automatically stops a global tendency. It requires firms, institutions and governments to accept bearing the cost of alternatives that the market, by itself, may not reward. The AI Act sets some common limits for certain risks and uses; it does not require prior justification of every adoption.

The Monday-morning test is simpler. When an AI function is already switched on in your office, ask who decided to switch it on, what benefit was measured against which alternative, what skill it risks consuming, and who answers if it gets things wrong. Ask also whether it can be switched off without losing access to the service. If the answer is only that the function was available, you do not yet have a reason to use it.