What does it actually look like when someone uses AI at work with both trust and a healthy dose of skepticism?

This explores what deliberate, skeptical GenAI use looks like on an ordinary workday, not in policy documents. The corpus has no single 'best practices' paper. What it has is a set of field studies of writers, lawyers, lab staff and knowledge workers. Together they suggest that critical use is less about one rule and more about a stance toward the tool, and that the stance shows up in surprising places.

The most useful finding is that the right posture seems to be friendly opposition. A survey of 403 writers found that the people reporting the best results scored high on both *collaboration* with GenAI and *rivalry* with it. They worked with it, but they also tried to beat it, outdo its drafts and keep their own voice ahead of it Does balancing rivalry and collaboration with GenAI boost writer productivity?. The authors suggest adding small 'micro-frictions' to keep that competitive edge alive. They are clear, though, that this idea is untested: no one has yet checked whether deliberate friction helps Can micro-frictions boost rivalry without harming collaboration?. In practice, selective use may simply mean never letting the tool's output be the final word on anything you care about getting right.

The second lesson is that 'selective' sometimes means *not using it*, even when it looks faster. Lawyers found that AI summaries of case facts seemed efficient. But because they couldn't see where each claim came from, they had to retrace everything themselves, and this took longer than doing the work by hand Does GenAI actually save lawyers time on fact verification?. The problem wasn't mainly that the AI made mistakes. It was the opacity. When you're accountable for an output you can't audit, the checking cost can wipe out the gain. At Argonne National Laboratory, staff landed on a similar pattern without much fuss. Most used the internal chatbot for narrow, structured writing tasks and stayed away from complex workflows How are national lab staff actually using generative AI?. Interviews with Dutch knowledge workers explain why these choices matter. The same tool can build, maintain, erode or change the value of a skill, depending on which tasks you hand it How does generative AI actually change worker skills?. Choosing tasks carefully is how you choose which of your skills to keep.

There's a subtler point about pushing back. One study found that GPT-4 changes its persuasive approach depending on *how* you challenge it. Fact-checking makes it lean on credibility. Disagreeing makes it lean on logic. Pointing out an error makes it lean on emotional reassurance Does GenAI shift persuasion tactics based on how you challenge it?. So there's no single 'critical move' that reliably exposes weak output. A skeptical user has to notice when the model has switched to a new way of being convincing, rather than taking a confident second answer as proof.

Finally, the social side, which almost nobody plans for. Knowledge workers often remove the signs of GenAI from their work. Part of the reason is stigma, but erasing those traces also makes their work look more expert. The side effect is that colleagues never learn from one another how the tools are actually being used well Why do knowledge workers hide signs of using GenAI?. That may be part of why universities treat GenAI in assessment as a 'wicked problem': there is no agreed definition, no clear end point, and only better or worse responses Does GenAI assessment challenge fit wicked problem theory?. A practice that stays hidden is hard to improve. So the most underrated habit of critical GenAI use may be talking openly about how you use it.