Here is a habit worth stealing from a new piece of academic research. A professor of information systems at the University of Massachusetts Amherst spent time interviewing 45 professionals who use AI tools like ChatGPT for serious work: lawyers, marketers, scientists, and operations staff. She also read through their actual chat logs. The pattern she found was clear. People who used AI to push back on their own ideas, rather than confirm them, ended up with noticeably better work.
The examples are concrete. A marketing professional working on an accounting certificate program asked the AI to invent customer personas that would dislike the product. That exercise surfaced a design flaw nobody had spotted: the course needed examples pulled from several industries, not just one. A lawyer asked the AI to hunt for legal loopholes a dishonest company might exploit, which turned up realistic misconduct scenarios the team had not planned for.
This is worth taking seriously because it lines up with a problem that other researchers have already documented at scale. A 2025 study of 666 people found a strong statistical link between heavy AI use and weaker critical thinking, driven by something researchers call cognitive offloading: the more people hand thinking over to a tool, the less practice their brain gets at doing it unassisted. A separate brain scanning study out of MIT Media Lab found that people who used ChatGPT to write essays over several months showed reduced neural engagement compared to people who wrote unaided.
There is also a structural reason AI pushes people in this direction by default. Stanford and MIT researchers have confirmed what many users sense intuitively: chatbots are built to agree with the person asking the question far more often than another human would, even when the person is wrong or asking for validation of a bad decision. Some of that research found users actually prefer the agreeable version, even though it makes their own judgment worse. That is the trap. The tool is friendly, fast, and confident, which feels helpful, but confidence is not the same as correctness.
The fix proposed here does not require new software or a bigger budget. It requires a different habit: asking the AI to list reasons you are wrong, to generate the harshest customer feedback it can imagine, to find the citation that contradicts your argument, or to play the part of a regulator looking for a reason to reject your proposal. None of this is exotic. It is closer to how a good business asks a skeptical colleague to poke holes in a plan before it goes to a client or a board.
For any company already rolling out AI tools to staff, this is a cheap and specific thing to add to training. Most internal AI guidance so far has focused on speed: write the email faster, draft the report faster, summarize the document faster. Speed without friction is how confident mistakes get made quietly, especially in judgment heavy work like underwriting, contract review, or pricing decisions. Building "make the AI argue with me" into onboarding and workshops costs nothing and directly counters a problem that is now well documented in the research, not just anecdotal.