The Verification Tax

Why AI makes your team faster and worse at the same time, and what to do about it

A while back, one of our engineers shipped a piece of work with a confident, specific, completely invented fact in it, generated by a AI, that sailed through because it looked right.

This was not a careless person. This was someone who had, in writing, warned the rest of us about exactly this failure situation would happen in the past. He knew the AI will confidently make things up. He'd said so. And he reproduced the error anyway.

That's not a story about one engineer having a bad day. It's a story about what happens to good judgment when it meets a tool that's right often enough to be trusted and wrong often enough to hurt you. I've started calling the cost of that gap the verification tax. The work you have to do to catch what the machines gets wrong. And I've become convinced that most teams are paying it in the most expensive way possible, with their people's attention, one busy afternoon at a time.

This has happened before

Every wave of automation makes the same promise and springs the same trap. The tool takes over the hard part. Output goes up. And the skill that used to live in doing the hard part starts to quietly drain away, because nobody's doing the hard part anymore.

We have the receipts on this now. There's a study of endoscopists, doctors who detect polyps during colonoscopies, whose unaided detection rates dropped after they got used to working alongside an AI assistant. They got worse at the thing they were expert at. The AI wasn't bad. The doctors just stopped practicing the skill that made them experts, and those skills declined. There's research on students using generative AI that found their output improved while their actual knowledge didn't budge. The authors call it "metacognitive laziness." There's work on computer programmers who reach a working answer without ever understanding why it works, an illusion of competence that outlives the task. Left alone, AI assistance erodes the expertise it appears to extend. You feel more capable while becoming less so. The loss of capability is what happens by default.

The trap is that verification is invisible

Here's what makes this so hard to manage. When verification works, it leaves no trace.

A well-checked piece of work and a lucky, unchecked one look identical in the final output. You can't audit your way to confidence by reading the finished product, because the finished product doesn't record whether anyone actually thought about it. We learned this the expensive way. We built a benchmark to measure the quality of an AI assistant's finished answers, and it told us almost nothing about whether the checking that mattered had happened. The indication is in the process behind the output, and most teams aren't conditioned to see the process at all.

So the fabrication incident wasn't a surprise, in hindsight. Nothing in the system was watching the part of the work where the mistake was made. We were trusting an individual's attention to catch it.

It doesn't have to work this way

If I stopped there, this would just be another doom post. But there's a genuinely hopeful finding underneath it, and it's the whole reason I care about the design question.

The erosion of skill is the default, but it's a default, not an absolute. The same research that shows AI deskilling people also shows the opposite happening when AI-enabled tooling is built differently. When the design forces you to engage, when there's friction that makes you think before you accept an answer instead of letting you click through, the same tools start building skill instead of spending it. The evidence for this is thinner and earlier than the evidence for skill erosion, and it comes mostly from classrooms rather than offices. But the indication in its evidence lends itself to establishing the OG cognitive tool we are all born with.

Verification becomes apprenticeship only when the checks built into the work force the person to think. Supervising AI output doesn't teach you anything by itself. It teaches you something when, and only when, the work is built so you can't Claude and chill.

The two-step job

So what do you actually do? The answer has two halves, and you need both.

Train the judgment. Someone on the team (why not you?) has to be able to recognize what good skepticism looks like, to feel the wrongness of a too-clean answer. You build this the way you build any skill - with friction on purpose (make people engage before accepting), with structured, repeated feedback on the AI's output, and by protecting the entry-level tasks through which the next expert actually gets made. An AI-enabled tool is very good at doing the exact work a junior person needed to do in order to become a senior level expert. If you let it, this is how you stop producing experts.

Next, build the judgment into the work. This is the half people skip, and it's the half that survives turnover and bad afternoons. Concretely - ground AI answers in verified sources before its own training data, give it a real refusal threshold so it can say "I can't confirm this" instead of inventing some confident nonsense, and put a verification step between the output and anyone who relies on it, a review before merge, an evaluation before release. The check has to read the process, not the product.

The clearest public example of this done right is Stripe. Every week, more than 1,300 pull requests written entirely by AI agents merge into systems moving over a trillion dollars a year in payments, each one reviewed by a person first. Their own standard says: "a mostly correct integration is a failure; payments require 100% accuracy." That's what a company looks like when an unverified error is expensive enough that they had to build around it. Most teams carry that same risk. They just haven't measured what an unverified error would cost them in a similar capacity.

Both halves matter. Our engineer was trained and aware, but nothing in the process was built to catch him, so the error he'd personally warned about went through anyway. Training without checks built into the work is a person promising to be careful forever. Nobody is careful forever.

The part I'm least sure about

One honest mention here, because I'd rather you trust the argument than have me oversell it. If calibrated skepticism is the competency that matters, the obvious next move is to hire for it. I don't think that works as a primary strategy. Good judgment still matters, but individual judgment is exactly the thing that doesn't survive at scale. You can't hire your way out of needing checks and stage gates built into the work. The process has to do the work either way. Finding a validated way to select for this trait before someone's on the team is a real problem, and a billion dollar solution if it gets figured out.

The choice you're already making

Using AI at work is already settled. What's not is whether you build the checks in and that decides whether your team's expertise compounds or leaves with the people who had it.

Leaders are making this call right now, most of them without even realizing it.

This is a plain-language adaptation of to-be published, "The Verification Tax: Designing AI-Augmented Work So Expertise Compounds Instead of Collapsing." The company and its data are anonymized. Thanks to Casey Smith, whose verification-first assistant is one of the cases behind this, and who let me write about it.

A few of the studies referenced: Budzyń et al. (2025) on endoscopist deskilling. Fan et al. (2025) on metacognitive laziness. Dawson et al. (2025) on the illusion of competence in programming education. Ganuthula (2026) on AI-induced skill atrophy. Kazemitabaar et al. (2023, 2024) on cognitive-engagement design. Parasuraman & Manzey (2010) on automation complacency. Beane (2019) on how automation removes the apprenticeship path.