The Verification Tax

Nikita Melashchenko · (19 April 2026) · opinion

Two publications landed on my desk that appear to contradict each other, and I think both are right.

The first is the randomised controlled trial by Daniel Schwarcz and colleagues on AI-powered lawyering. Law students completed legal tasks with a retrieval-augmented legal AI tool, a reasoning model, or no AI, and the AI groups produced significantly better work, with productivity gains of 50 to 130 per cent in five of six tasks.^[1] Quality went up, not just speed, which is a break from the earlier GPT-4-era studies. The second is Joshua Yuvaraj’s “verification-value paradox”. Because generative AI is disconnected from ground truth and lawyers carry paramount duties of honesty and not misleading the court, the author argues that every efficiency gain is met by a correspondingly greater duty to verify the output manually, so the net value of AI in legal practice is often negligible.^[2]

The way to hold both findings at once is to notice that the verification tax is not flat. It varies with how checkable the output is. A persuasive letter can be verified by reading it, where the claims are the drafter’s own, and the quality is on the page. A research memorandum citing forty authorities is different. Each citation is a promissory note that must be individually redeemed, and the court filings that have earned lawyers disciplinary referrals in Australia and elsewhere are exactly this failure to redeem. Daniel Schwarcz and colleagues’ own data gestures at the same point. The retrieval-grounded tool avoided adding hallucinations while the bare reasoning model did not. Retrieval grounding is, in effect, a technology for lowering the verification tax, because it turns “trust me” into “here is the source, check it”.

So the paradox is real, but it is a schedule of rates, not a flat prohibition. Tasks where verification is cheap relative to generation, drafting, restructuring, first-pass analysis, are where the tools deliver value. Tasks where verification costs as much as doing the work, authority-heavy research for filing, are where Yuvaraj’s warning bites hardest, and where the profession’s disciplinary cases will keep coming from. However, the paradox is arguably the consequence of using the tools in the wrong way, or largely the wrong too for the wrong tasks.

There is a third paper I want to mention, from a different literature entirely. Giray catalogues “AI shaming” among academic writers, the practice of dismissing AI-assisted work as lazy or inauthentic, and documents its effects: inhibited adoption, stifled innovation and self-censorship born of anxiety about how disclosed use will be judged. His own prescription is transparency, declare the use, take responsibility for the output.^[3] I would push the logic one step further than he does. A profession that shames disclosed AI use does not thereby get less AI use; it risks getting undisclosed use instead, and undisclosed use is precisely the use that never goes through a verification step, because admitting to verifying would mean admitting to using. It is similar to trade secret and patents debate, where without the right incentives the socially optimal level of disclosure is not achieved. Therefore, if the verification tax is the price of these tools, the worst possible policy is one that incentivises people to evade the tax collector.

I argued in an earlier note that the pragmatic response to AI is neither prohibition nor surrender but adjustment of practice. The adjustment now has a name and a price. Choose the right tools, pay the verification tax, structure work so the tax is low and stop pretending the alternative is a world where nobody uses AI tools.