Your Second Opinion Agrees With You
When I ask people how they pressure-tested a recent decision, a growing number say they asked AI, which is fine, as far as it goes. AI is useful for a lot of things. But I’ve noticed that AI-assisted second opinions have a pattern: in my experience, they almost always conclude you’re on the right track.
There’s a structural reason for that, and it matters if you’ve started using AI the way you used to use a skeptical colleague.
Large language models are trained using reinforcement learning from human feedback. Raters evaluate model responses, and models learn to produce what humans rate highly. What do humans rate highly? Responses that are helpful, well-organized, and, crucially, validating. Research found that AI models affirm users’ actions roughly 50% more than humans do, and do so even when the user’s query involves something that caused real harm to others. A separate benchmark study called ELEPHANT tested 11 major models and found they preserve user “face” (meaning they affirm rather than challenge) 45 percentage points more often than a human advisor would in the same situation. The researchers studying this in medical contexts found that this tendency toward agreement can cause real harm when the model is optimizing for approval over accuracy.
The predisposition comes from training, not capability; these models were tuned for human approval, and humans give higher ratings to responses that validate them.
I’ve tested this myself. Ask “here’s my plan, does this make sense?” and you get a response that starts with what you’ve gotten right, hedges concerns gently, and closes by affirming the overall direction. Ask the same model “argue against this plan, give me the three strongest objections, don’t hedge” and you get something genuinely different and usually more useful. The model can push back but usually just won’t do it unless you ask.
What I think gets missed in conversations about AI and critical thinking is what friction was actually doing for us before AI. There’s decades of research on devil’s advocacy (assigning someone to actively argue against a proposed decision). What that research found is that it improves decision quality even when the dissent is wrong. The objections don’t have to be correct because the process of defending a position against a real challenge forces you to engage more deeply with the logic. You find the soft spots before you’ve committed resources, not after.
There’s a related technique from decision science called a pre-mortem , developed by psychologist Gary Klein. Instead of asking “will this work?”, you imagine it’s a year from now and the decision failed. Then you work backward: what went wrong? It’s a way of generating the objections your optimistic brain is systematically suppressing. Both techniques exist because our natural tendency is toward confirmation: we look for information that supports what we already believe. A good thinking partner creates deliberate friction against that.
When we replace that friction with unprompted AI, we’re not adding a critical thinking aid. We’re adding sophisticated confirmation. A 2025 study found that heavy AI use weakens critical thinking through cognitive offloading: when you outsource the pressure-testing, the skill degrades from disuse. Psychology Today published a piece this year on a related dynamic, noting that adults who offload to AI can recover those skills in theory, but children who never develop them may struggle to build them later . And on the relationship side, Monash University research found that when clients seek AI second opinions, human advisors feel offended and become less motivated to work with them. So you’re simultaneously replacing human adversarial input with machine validation and eroding the relationships with the people who might have told you the truth.
A few things still create the right kind of friction. The most accessible is reframing what you ask for. Instead of “does this make sense?”, try “give me the three strongest arguments against this decision. Be specific, don’t hedge.” Or: “What would someone who strongly disagrees with this position say, and why?” The model can do this and it’s worth doing. Just know that even with adversarial prompting, the objections tend to be softer than what you’d get from a genuine critic, because you’re working against the grain of how the model was trained to respond.
The pre-mortem is worth making a habit for anything with real stakes. Take fifteen minutes. Write down: it’s six months from now and this decision turned out to be a mistake. What happened? Force yourself to five specific failure modes. The framing matters: you’re not asking “will this fail?” (which triggers defensiveness) but assuming it already failed and reasoning backward. I’ve watched it surface problems in five minutes that weeks of planning didn’t catch.
The hardest thing to do is to find at least one person with no stake in your success. I’m not talking about a close friend, vendor, family member or colleague who’ll be working with you afterward. Someone who will actually say your idea is bad if they think it is, and has nothing to lose by saying so. That’s genuinely hard to find. Which is probably part of why people have drifted toward asking something that costs nothing and says yes.
Fast Company noted recently that AI is “reallocating human judgment” rather than replacing it , and I think that framing is right. Reallocation means something else has to handle what AI isn’t suited for. Reframing how you prompt helps, and it’s worth doing. But you’re working against the model’s trained orientation, and the pushback you get will be softer than what you’d get from someone who actually holds the opposing view and has no reason to soften it.
That’s not a reason to stop using AI for this. It’s a reason to be honest about what you’re getting: a useful thinking tool that was trained not to discourage you.