Imagine being asked to approve a report assembled by an AI assistant. The numbers look reasonable. The explanation is clear. When you ask whether the conclusion holds up, the assistant supplies a careful-sounding review of its own work.
Your name goes on the report.
What would let you disagree with it?
The original records would help. So would a calculation you can reproduce, or a colleague with the relevant expertise. A test could help, if you can see what it tested and what happened, rather than relying on the assistant's description of it.
When the explanation and the reassurance are all you have, checking stops being a check. It becomes another act of trust.
Two ways to approve the same report
Self-review deserves to be taken seriously. In the 2023 Self-Refine study, having a model critique and revise its own answers improved average performance across the seven tasks the researchers evaluated. Self-Refine, May 2023 version That is a real possibility, not a guarantee that any particular critique is right. Whether a model agrees with itself or a second model agrees with it, the question is the same: what evidence changed, and what check was actually performed?
Take a hypothetical retention report. Every calculation is correct. But the definition of an active customer changed between the two periods, and the report presents the two rates as directly comparable.
In the first arrangement, the reviewer receives the report and the assistant's assurance that the figures were double-checked. They were. The arithmetic is fine, so the approval goes through, and the misleading comparison goes with it.
In the second, the reviewer also has the definition behind each figure, time to look, and the standing to hold approval until the comparison is explained. The same person, with the same skill, now has a real chance of catching the problem.
The difference is not diligence. It is access, time and authority. Assigning responsibility takes one sentence. Giving someone the practical ability to exercise it takes more work.
None of this is new. Separating the person who prepares work from the person who approves it is an old idea. What changes with AI is that the work and the reassurance about the work can now come from the same place, instantly, fluently and in whatever volume you ask for.
Less effort is not necessarily a loss
A CHI 2025 survey by Microsoft and Carnegie Mellon researchers collected 936 examples from 319 knowledge workers. The more confident people were in AI's ability to do a task, the less effort they reported on five of six thinking activities. These are people's perceptions, not measured long-term skill loss. Study, sections 5.1 and 6.3
Lower effort can mean the support is working. It can also mean questions went unasked. The survey alone cannot tell you which happened in a particular piece of work.
There is a serious counterweight to any simple story of decline. In a study of 5,172 customer-support agents, AI assistance increased issues resolved per hour by 15% on average, and the authors found evidence consistent with workers learning from it. Yet the most experienced and highest-skilled workers saw small declines in quality. These are findings from one workplace, reported in the paper's November 2024 revision, not a forecast for every job or for today's models. Generative AI at Work
Neither study settles how a specific team should delegate its work. Together, they suggest a more useful question than whether AI makes people sharper or duller: what does this arrangement help the person do, and what can they still check?
A review with something behind it
Try this on one recurring, low-risk task where an AI tool is already approved for use. Before reading its answer, write down what would make the result unacceptable.
For a summary, a missing exception might change the reader's decision. For a spreadsheet, the totals might need to reconcile with the original records. For a customer response, the proposed remedy might need to match what the company can actually provide.
Then locate the evidence for that check.
Do not stop at a generated citation or a statement that the test passed. Open the source or inspect the result. Ask whether it addresses the requirement, and whether it belongs to this version of the work. A genuine source can still be irrelevant; a passing test can cover the wrong thing.
If the evidence is unavailable, make that limitation visible before approval. That may mean asking for a source, narrowing the conclusion or getting help from someone qualified to assess it. It need not mean abandoning the task.
This is an exercise for ordinary work. High-stakes decisions still need their established professional and organizational safeguards.
For a manager, there is a further question. Does the person doing the checking have time to investigate and permission to raise an unresolved concern? If they cannot delay approval or get help, what do you expect their sign-off to accomplish?
What is worth keeping within reach
Nobody needs to reconstruct every delegated answer from first principles. That would defeat much of the purpose of working with other people or tools.
A more practical standard is to identify the consequential parts, find evidence that bears on them, and recognize when you need help. The more consequential the work, the more convincing the check should be.
For the next piece of AI-assisted work you approve: what evidence would let you say no?
Disclosure: George Pu is the founder of SimpleDirect, an independent AI lab whose research ships under Vinci and includes work on verification systems. This essay does not discuss or recommend any Vinci product. Founder Reality is his publication. About George and Founder Reality