Who AI leaves behind

The most consistent finding about AI at work is that it helps the least skilled most. The most uncomfortable finding is that the apparatus built to police it penalises many of the same people. Those two results are rarely read together, and they should be.

The tool levels. That part is well evidenced.#

Two studies in very different settings point the same way. Brynjolfsson, Li and Raymond followed a staggered rollout across 5,172 customer support agents: resolutions per hour rose about 15 per cent on average, while the least experienced gained around 30 per cent and the lowest skill quintile 36 per cent, with no significant gain for the most skilled. Kanazawa and colleagues followed a Japanese taxi fleet through an AI demand-prediction rollout and found productivity gains accruing almost entirely to low-skilled drivers, narrowing the gap between best and worst by 14 per cent.

Knowledge work in an office and manual work in a car, and both raise the floor rather than the ceiling. For anybody arguing that this technology inevitably concentrates advantage, these are the results to answer. On the direct question of who the tool helps, the answer is that it disproportionately helps people who were behind.

The checking penalises the people the tool helped#

Now put that next to Liang and colleagues, who tested seven widely used GPT detectors against TOEFL essays by non-native English speakers and essays by US eighth-grade students. The American children were classified with near-perfect accuracy. More than half the non-native essays were misclassified as AI-generated, an average false positive rate of 61.22 per cent.

The mechanism matters more than the number. Detectors rely on perplexity, a measure of how predictable text is, and writing carefully in a language you learned later is more predictable. The property being flagged as machine-likeness is a property of being a competent second-language writer. Vendors dispute how far this carries to current tools, and the study tested what existed then. Nothing about the mechanism has changed.

So the same person who gains most from the tool is the person most likely to be accused of using it. In a university that is an academic misconduct process. In a company it is a quiet judgement about whether somebody wrote their own board paper. The accusation is invisible from outside, because a false positive looks identical to a true one.

Who cannot afford the better tools#

This one has less direct evidence than it deserves. Better to say so than to fill the gap with inference.

What can be established is that access is uneven and that the gap tracks existing lines. Adoption data consistently shows concentration by sector and firm size: the most recent labour market reporting in this corpus puts overall firm adoption at 28.5 per cent while information and communications sits at 74.1 per cent and professional services at 57.5 per cent. Frontier capability is priced at a professional subscription rather than at consumer level, and the gap between the free tier and the paid tier is not cosmetic.

What has not been measured is the effect of that gap on individual outcomes: whether people on free tiers fall behind in ways that show up in pay, hiring or performance. Anybody presenting you with a figure on this has estimated it. The mechanism is plausible and the measurement does not exist yet, which is a different statement from either optimism or alarm.

The macroeconomic backdrop is worth holding alongside it. Acemoglu’s task-based model estimates total factor productivity gains of no more than 0.66 per cent over ten years, under 0.53 per cent once hard-to-learn tasks are accounted for, and argues AI is likely to widen the gap between capital and labour income rather than reduce inequality within labour. A levelling effect between workers is compatible with a widening gap between workers and owners, and most commentary picks one of those and forgets the other.

Is freelancing safe?#

There is no good direct evidence in this corpus, and nobody knows. Two considerations are worth weighing rather than one.

Against safety: freelance work is disproportionately the discrete, specifiable, deliverable-shaped work that models handle best, and a freelancer has no institutional relationship absorbing a client’s first experiment with doing it themselves. In favour: the same flexibility that makes freelancers replaceable makes them adoptable, and the Autor and Thompson lens applies here as anywhere, so a freelancer whose value is judgement rather than production is in a different position from one whose value is throughput.

The specific thing to watch is not volume of work but the shape of the brief. If clients increasingly arrive with a machine-generated draft asking for correction, the role has moved from producing to verifying, which is a different job at a different price, and one where the evidence on verification work is not encouraging.

What follows for an organisation#

• Do not run detection on people. A 61 per cent false positive rate against second-language writers does not describe a tool needing calibration. It describes one that should never be pointed at individuals whose position depends on the result.
• Check who your written filters are selecting for. If a process rewards fluent English rather than sound reasoning, AI has raised the fluency of everybody who uses it, so the filter now measures something it never intended to.
• Fund the tier. If capability now varies with which subscription somebody has, that is a procurement decision rather than a personal one, and leaving it to individuals converts a budget question into an equity one.
• Watch the levelling with clear eyes. The compression is real and mostly good news for the people at the bottom. It also compresses what experienced people are paid for, which is covered in the mid-career squeeze. Both are true.

Where this sits in my own argument#

I am pro-AI and pro-human, and this page is where holding both positions gets uncomfortable. The levelling is real and I have no interest in burying it because a fairness argument is easier to make in the other direction. The cost is real too, and it falls on people who have no way of knowing they are being marked down.

The pattern is the one I call drift. No organisation decides to penalise second-language writers. It buys a detector, points it at people, and the consequence arrives without anyone choosing it.

What I have observed in organisations#

The access gap I said was unmeasured is one I have watched play out. In marketing, larger firms worked out quickly that they could use these tools across customer acquisition. Smaller firms often cannot justify the budget, so the same capability is available to one and not the other.

The consequence I did not expect is that it moves people. I have seen talent leave larger firms for smaller ones, carrying the working methods with them. That is a levelling of a sort, and it happens person by person rather than through anything anybody planned. It is also slow, and it does nothing for the firms nobody chooses to move to.

What this page does not claim#

It does not claim AI is bad for people who were already disadvantaged. The direct productivity evidence points the other way, clearly, in two independent settings, and that should not be buried because a fairness argument is easier to make in the opposite direction.

It does not claim the detector bias has been solved or that it has not. The study is from 2023, vendors dispute the generalisation, and no independent replication on current tools appears in this corpus. The mechanism gives a reason to expect the problem persists; that is a reason rather than a measurement.

And it does not offer a figure for the access gap, because none exists that survives checking. Naming that absence is more useful than filling it.