Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs™. Each week, I dig into research on AI, knowledge management, and the future of work so you don't have to.
EDITOR'S NOTE
Last week I ended on a question: if judgment is what makes any of this legitimate, somebody has to be doing the checking. So I went looking for the people actually doing it.
There aren't many. In that same global survey of 48,340 employees, two in three said they rely on AI output without evaluating it, and more than half said they've made mistakes at work because of it.
I don't think AI is making people careless. People are adapting to a new tool the way they always do. The problem is that every adaptation shifts a cost somewhere else, and in this case that cost is increasingly the time, attention, and judgment required to verify what the system produces.
So this week, I went looking for where the checking went.
———————————————————————————
THIS WEEK'S READING:
The Work AI Index 2026 — Glean Work AI Institute
Six thousand digital workers across the US, UK, and Australia report that AI saves them eleven hours a week. They also report spending 6.4 hours a week on what the study calls "botsitting," roughly 37 percent of all the time they spend working with AI.
Why It Matters: This is vendor research, so weigh it accordingly. But the interesting part is where those hours go. The biggest share isn't spent catching errors. It's spent giving the system enough context to do useful work in the first place, ahead of both supervising it and correcting its output.
That's my week too. I can ask ChatGPT to map a database schema or find a dataset, and I almost always end up correcting it. Not because it's necessarily wrong, but because it doesn't know my specific context. Getting it there is the job.
That's why "just check the output" undersells what verification actually involves. The hours disappear into the back-and-forth required to produce something worth auditing in the first place, and most of that effort never shows up anywhere. The eleven hours saved gets reported upward. The 6.4 hours spent making the tool useful usually doesn't.
So checking is expensive. How much of it is actually happening?
The Impact of Generative AI on Critical Thinking — Microsoft Research and Carnegie Mellon
Researchers walked 319 knowledge workers through 936 real AI-assisted tasks. One hundred fourteen checked the output against an outside source. Twenty-three checked the sources the AI itself had cited.
Why It Matters: Twenty-three is the number that sticks. At first glance, it looks like laziness. Then you get to the more interesting finding: confidence in the tool predicted less critical thinking, while confidence in yourself predicted more.
Fluent output creates exactly the wrong kind of confidence. The better it reads, the less likely anyone is to look twice. And even checking the sources doesn't always save you.
Search for something like ear wax removal, and an AI summary can quietly turn an ordinary problem into something that sounds like it requires a specialist. Nothing has to be fabricated. The model can simply summarize doctor's-office pages written, quite reasonably, to get people through the door. You check the answer against those sources, arrive at the same conclusion, and walk away feeling careful.
The problem isn't only whether the AI invented something. It's whether the reference material itself gives you an independent reason to trust the conclusion.
And even when someone does look closely, the label can change what they see.
Why Do Experts Miss AI's Errors? — PNAS Nexus
Researchers showed 1,339 teachers the same incorrect grade. The only thing that changed was whether the grade was labeled as coming from a human or an algorithm. When it was labeled as AI-generated, the harm caused by the error increased by 22 percent.
Why It Matters: The teachers who deferred most strongly were younger, better educated, and more technologically confident, almost the opposite of what people tend to assume. They gave the algorithm more credit for competence and, with it, more of the responsibility. The decision felt less like theirs to make.
That suggests the missing skill isn't simply "be more suspicious." That's too vague to be useful. The real skill is knowing how much judgment to hand to the model and how much to keep for yourself.
That line is difficult to hold because nothing tells you exactly where it belongs. And the more comfortable you become with the tool, the easier it is for that line to drift. What begins as assistance can gradually become deference without anyone consciously deciding that the model should now own more of the call.
Let it drift far enough, and eventually the mistake stops being your problem alone.
AI Hallucination Cases — Damien Charlotin
A researcher has been logging court decisions worldwide in which a judge found that someone had filed AI-fabricated material. The tally is now above nineteen hundred cases across more than fifty jurisdictions, and it continues to grow.
Why It Matters: The sanctions aren't the part that interests me. Case law runs on precedent. If precedent gets built on citations, quotations, or authorities that were never real, the problem compounds.
It's the same failure as the medical summaries above, only one layer deeper. A reference layer is supposed to be the thing everybody else checks against. Corrupt that layer and careful people can reach the wrong answer while doing everything correctly.
It's also worth looking at who filed these cases. More involved people representing themselves than lawyers. That matters because this isn't simply a professional discipline that lawyers forgot. It's a verification skill most people were never taught in the first place.
We tend to talk about hallucinations as a problem between one user and one model. But once fabricated material enters a system other people rely on, the failure becomes cumulative. The next person can be cautious, diligent, and completely unaware that the thing they're checking against has already been compromised.
Now count how many people are left to catch any of it.
Hiring Benchmarks 2026 — Greenhouse
Across 640 million applications, the average recruiter went from handling 146 applications a year in 2022 to 746 in 2025. At the same time, the average recruiting team shrank from ten people to fewer than five. Time to fill grew 37 percent anyway.
Why It Matters: Greenhouse sells screening software, and to its credit, it doesn't claim AI caused all of this. It calls AI "a key lever" and stops there. I'll draw the connection myself rather than borrow one the source never made.
What the numbers show is a trade nobody consciously agreed to: good, easy, trusted. Pick two.
Applications became easier to create. Once they became easier to create, they became harder to trust. And the cost of figuring out which ones were actually good landed on roughly half as many people.
That's the part I keep coming back to. You can control whether your work is good. You can control whether it's easy to produce. But trust belongs to the person receiving it. Once volume rises faster than verification capacity, trust is usually the first thing to disappear.
———————————————————————————
Put all five studies together and I don't see a story about people getting sloppy. I see people adapting sensibly to a new tool while the cost of that adaptation lands somewhere nobody was measuring.
Some of it shows up as hours spent supplying context that never make it into the productivity calculation. Some of it appears when fluent answers suppress the second look. Some of it comes from an AI label quietly shifting ownership of a decision. Some of it accumulates inside a reference layer that degrades one fabricated citation at a time. And some of it lands on a shrinking group of people asked to verify a growing pile of work.
Good, easy, trusted. Pick two.
That helps explain why this doesn't get fixed by telling everyone to try harder. But critical thinking isn't one of the three. You don't trade critical thinking for quality, speed, or trust, and giving it up doesn't buy you anything.
The problem is that we've been treating judgment as though it were free. It isn't. Every act of verification takes time. Every second look has a cost. Every source you open, assumption you test, and answer you challenge consumes attention that has to come from somewhere.
When nobody explicitly decides who is supposed to pay that cost, it gets spent quietly through adaptation. The employee spends more time supplying context. The expert gives the system a little more deference. The recruiter absorbs another hundred applications. The lawyer, researcher, or analyst assumes the reference layer is still intact. Individually, each decision can make sense. Collectively, the system becomes harder to trust.
Which leaves us with one of the oldest rules there is: trust, but verify.
It's worth repeating now because everything above tells us what the second half actually costs. Most of us have been acting as though it were free.
One thing I'm still turning over is what happens when the reference layer itself starts to degrade. What happens to the people whose entire job is to be the thing everyone else checks against?
That's next week.
Know someone who's quietly eating the cost of all this checking? Forward this to them.
Have a take on any of this? Reply directly. I read every response.