← Research Research · Positions

What I argue, and what the evidence shows

This site makes three different kinds of statement, and they are not interchangeable. This page is the third kind, gathered in one place so it can be quoted accurately.

Empirical#

What a study found. Graded, sourced and listed in the evidence base, with a note on what it does not support.

Definitional#

What a term means here. Set out in the glossary. A definition is a convention, not a finding.

Interpretive#

What Rahim Hirji argues follows from the evidence. This page. These are readings, and reasonable people reject some of them.

Every argument below carries the graded evidence that bears on it and a statement of what it does not claim. Where a source complicates the argument, it is listed under the argument it complicates rather than left out. Each has a permanent address, so a single position can be cited without citing the whole page.

Interpretive

AI reaches human judgement before it reaches jobs.##

The public argument about AI is almost entirely about employment. The effect that arrives first, and that is already measurable, is on the quality of the thinking people do while still holding the same job. Judgement degrades before headcount does, and it degrades quietly, because nothing about the output announces it.

What this does not claimIt does not claim that jobs are safe, that employment effects will not arrive, or that the judgement effect has been demonstrated causally at population scale. The strongest single study behind it is a field experiment, not a longitudinal measurement, and the self-report work is self-report.

Evidence that bears on it: Dell'Acqua (2023) · Lee (2025) · Gerlich (2025) · Parasuraman (2010). Argued at /research/ai-and-human-judgement.

Interpretive

Most organisations arrive at their AI position without deciding it.##

Adoption happens through hundreds of small local choices, none of which was the decision. The alternative is to decide in advance which judgements stay human and to design the work around that answer. The difference between the two is not sophistication or spend. It is whether anybody chose.

What this does not claimIt does not claim that designed adoption has been shown to outperform drift. No controlled comparison of the two exists, and this framework has not been tested against a control. It is a way of seeing the problem, offered as such.

Evidence that bears on it: BAuA (2025) · Japan Institute for Labour Policy and Training (JILPT) (2025) · Parasuraman (1997). Argued at /research/design-versus-drift.

Interpretive

Measuring usage tells you almost nothing about whether capability improved.##

Nearly every organisation measures logins, licences and prompts, and reports them as adoption. Those numbers describe activity. They are silent on whether anybody got better at anything, and they can rise while the underlying capability of the organisation falls.

What this does not claimIt does not claim usage data is worthless. It is useful for licensing, cost and support. The objection is to reporting it as a capability measure, which is a different quantity that nobody is measuring.

Evidence that bears on it: BAuA (2025) · Japan Institute for Labour Policy and Training (JILPT) (2025). Argued at /research/how-do-you-measure-ai-adoption-properly.

Interpretive

The damage falls on how skill is formed, not on skill already held.##

AI is efficient at removing exactly the work through which people used to become competent: the first draft, the routine analysis, the small case nobody senior wanted. Those who already have expertise keep it for a while. Those who were going to acquire it have lost the route.

What this does not claimIt does not claim that graduate hiring has collapsed because of AI. The French statistics office, whose data is the sharpest signal available, cautions explicitly against attributing the fall to AI alone, and that caution is carried here rather than dropped.

Evidence that bears on it: Dauth (2021) · Kissin (2026) · Ericsson (1993) · Arthur (1998). Argued at /research/missing-rungs.

Interpretive

A human in the loop is not, by itself, oversight.##

Placing a person at the end of an automated process satisfies most policies and very little else. Where the machine is usually right, attention decays, and the reviewer stops reviewing while continuing to approve. Oversight is a property of how the work is designed, not of who is nominally responsible for it.

What this does not claimIt does not claim human oversight is impossible or that the requirement should be dropped. It claims the requirement as usually written does not produce the thing it names.

Evidence that bears on it: Parasuraman (2010) · Dzindolet (2003) · Saudi Data and AI Authority (SDAIA) (2023). Argued at /research/human-in-the-loop-is-not-a-safeguard.

Interpretive

Checking AI output is work, and almost nobody counts it.##

Time saved in production is reported. Time spent establishing whether the output is true is absorbed silently by whoever is accountable. Where verification is genuinely done, the saving is smaller than claimed. Where the saving is as large as claimed, verification is usually not being done.

What this does not claimIt does not put a number on the cost. No study here measures how much time proper verification takes across a real workload, and any figure offered would be invented.

Evidence that bears on it: Huemmer (2026) · Divisional Court of England and Wales (Dame Victoria Sharp P and Johnson J) (2025) · Magesh (2025). Argued at /research/the-verifiers-discount.

Interpretive

How AI arrives in a team predicts the effect better than which tool arrived.##

The variable that moves outcomes is not the model. It is whether staff were consulted, whether training was funded, and whether anyone said out loud what the tool is for. This is inconvenient, because the tool is the part that gets procured and the introduction is the part that gets skipped.

What this does not claimIt does not claim the choice of tool is irrelevant, and it does not claim good introduction guarantees benefit. One of the studies behind it found effects on experts to be individual and currently unpredictable, which cuts against any confident promise in either direction.

Evidence that bears on it: Japan Institute for Labour Policy and Training (JILPT) (2025) · BAuA (2025) · Yu (2024). Argued at /research/ai-workforce-strategy.

Interpretive

Deskilling is a design problem, not a discipline problem.##

The common response to capability loss is to tell people to be more careful, or to run awareness training. Both put the burden on the individual at the moment of use, which is the moment they are least able to carry it. The decisions that matter were made earlier, by whoever designed the workflow.

What this does not claimIt does not claim individuals have no agency, and it does not claim training never helps. It claims awareness alone is a weak control, which is a measured finding rather than an opinion.

Evidence that bears on it: Dzindolet (2003) · Parasuraman (1997) · Infocomm Media Development Authority (IMDA) (2026). Argued at /research/capability-debt.

Interpretive

Expertise is built through effortful work that AI is very good at removing.##

Competence comes from doing difficult things repeatedly, with feedback, including the attempts that go badly. Every one of those is a candidate for automation, and each is individually easy to justify removing. The cost appears years later in people who were never required to struggle.

What this does not claimIt does not claim that practice volume alone produces expertise. One of the sources above is a direct challenge to the strong form of the deliberate-practice claim, and it is listed here because it complicates the argument rather than despite that.

Evidence that bears on it: Ericsson (1993) · Macnamara (2019) · Arthur (1998). Argued at /research/how-humans-learn-with-ai.

Interpretive

This is an argument for using AI well, not for using it less.##

Nothing here recommends refusing the technology or slowing adoption. The position is that the gains are real and that they are being taken in a way that quietly spends something not on the balance sheet. Deciding where judgement stays is what makes the gains keepable.

What this does not claimIt does not claim that careful adoption is costless, that every organisation should adopt, or that the optimistic case is wrong. One of the sources above argues seriously that AI is expertise-widening rather than expertise-replacing, which is close to the opposite of the thesis here.

Evidence that bears on it: Brynjolfsson (2023) · Autor (2024) · Storm (2015). Argued at /research/the-superskills-thesis.

How to quote this#

If you are writing about this research, the three sentences are different and the difference matters. "Research reviewed by SuperSkills finds" belongs to the evidence base. "Rahim Hirji argues" belongs to this page. "SuperSkills uses the term" belongs to the glossary. Each argument here has its own address, in the form of this page followed by a hash and the identifier shown beside it.

The same holds in the structured data. These are schema.org Claim nodes authored by a named person, not Article findings, and they are deliberately not typed as anything a machine would read as established.

Box of Amazing

Rahim’s free weekly letter on AI and human capability

If this was useful, the weekly letter is where the thinking happens first. Most of what ends up on this site starts there. Weekly essays on AI, capability and the future of work. Read by 25,000 people, every week since 2017. Free, and one click to stop.

Opens Substack to confirm. No pitch in it, unsubscribe in one click, and nobody follows up because you read something.

Running an event, or responsible for how AI arrives in your organisation? Keynotes  ·  Advisory and coaching  ·  Enquire