As AI takes on more of the production of work, drafting, summarising, analysing, coding, the most valuable human contribution is quietly shifting from making things to checking them. This sounds straightforward. It is not. Verification is a genuine skill, it takes real effort, and the evidence suggests we are not naturally good at it.
The Harvard and BCG study of consultants using GPT-4 found that the people whose performance was harmed by AI shared a common habit: they accepted its output with too little scrutiny, what the researchers called unengaged interaction. The separate Microsoft and Carnegie Mellon research on critical thinking found the same dynamic from another angle: the more people trusted the AI, the less cognitive effort they applied to checking it. Fluent, confident output lulls capable people into waving work through.
Why verification is hard
Checking AI output is harder than checking a colleague's work, for a few reasons. The output is polished, so errors do not announce themselves through clumsy phrasing. The model is confident even when wrong, so there is no hesitation to alert you. And the errors are often plausible rather than absurd, a slightly wrong figure, a citation that does not quite exist, a conclusion that almost follows. Spotting these requires active, knowledgeable attention, which is precisely the effort AI tempts us to save.
This is why verification cannot be left to good intentions. People who are busy, trusting and under time pressure, which is to say most people, will default to accepting fluent output unless the system and the culture make checking the expected behaviour.
Build this capability in your organisation
Growth Performance designs bespoke programmes that turn insight into measurable behaviour change.
Making checking the default
The good news is that verification responds well to structure. Simple practices help: stating up front what would make an output wrong and checking specifically for that, requiring sources to be traced to their origin rather than trusted on sight, and adding a deliberate pause before high-stakes outputs are accepted. Forming a provisional view before seeing the AI's answer keeps your judgement independent rather than anchored to the machine's. None of this is exotic. It is the discipline of treating AI output as a draft to be tested rather than an answer to be trusted.
What this means for leaders
If verification is the new core skill, it deserves to be taught, resourced and rewarded like one. That means training people in how to check, not just how to prompt, building checking time into workflows rather than assuming it happens, and making it clear that catching an AI error is valued work, not pedantry. The organisations that build a strong verification habit will quietly avoid the costly, confident mistakes that catch out those who do not.
Try this
For your team's most important AI-assisted output this week, require a named person to write two sentences: what they checked, and what they would expect to see if the output were wrong. Make it routine. The habit of articulating the check is what keeps the check honest.
To assess how rigorously your team interrogates its reasoning and outputs, try the Decision Trace Readiness diagnostic.