🔁 Why AI Makes a Bad Second Opinion (And a Great First One)

There's a specific way a lot of people have started using AI that feels reasonable on the surface but tends to produce weaker outcomes than they expect: making a decision first, then asking AI to check it. "Does this plan make sense?" "Is this the right call?" "Can you sanity-check this approach?" These questions feel like due diligence. In practice, they're often asking AI to validate a decision that's already been made, and AI is structurally not very good at that particular job.

The distinction that matters here is sequence. AI brought in before a decision is formed and AI brought in after a decision is formed produce genuinely different kinds of value, and most people default into the second pattern without realizing the first would usually serve them better.

------------- Context -------------

When AI is asked to evaluate a decision that's already been presented as the plan, it tends to find reasonable support for that plan, because the framing of the question shapes the response. Ask "does this make sense" about almost any coherent plan, and a capable AI model will generally find a way to say yes, with some caveats, because most reasonably constructed plans do make some sense, and the question as framed is oriented toward confirmation rather than genuine challenge.

This isn't a flaw exactly. It's a reflection of how these tools respond to framing. A question asked in a confirmatory posture tends to get a confirmatory answer, unless the plan is genuinely and obviously flawed. The subtler problems, the ones that a good second opinion is actually supposed to catch, are much less likely to surface when the question is framed as "check this" rather than "help me think through this from scratch."

Contrast this with AI brought in before a decision has formed, asked to help explore the problem itself: what are the options, what are the tradeoffs, what am I not considering. This framing produces a genuinely different quality of engagement, because there's no existing conclusion for the response to gravitate toward. The AI is helping construct thinking rather than validate a thought that's already complete.

------------- The Pattern That Produces Weaker Decisions -------------

A business owner who had gotten into the habit of running significant decisions past AI after reaching her own conclusion, essentially as a confidence check, noticed over several months that the practice wasn't catching the mistakes she'd hoped it would. Decisions that later turned out to have real problems had, in each case, received a broadly encouraging response from AI when she'd asked it to evaluate them. It wasn't that AI had been wrong exactly. It was that the framing of her questions had made genuine challenge unlikely.

She changed her approach specifically: instead of forming a decision and then asking AI to check it, she started bringing problems to AI before she'd reached a conclusion, explicitly asking it to help her map out the full range of options and considerations before she committed to a direction. The quality of her decisions improved noticeably, not because AI had gotten better, but because she was using it at the point in her process where it could add the most genuine value, rather than at the point where its natural tendency toward confirmation could do the least good.

The time implication here is significant, because bad decisions caught late are far more expensive to correct than better decisions made early. A decision that turns out to be flawed after significant time and resources have already gone into executing it costs far more to unwind than the same flaw would have cost to catch during initial exploration, when changing direction was still cheap.

------------- Where a Real Second Opinion Still Comes From -------------

None of this means AI has no role in checking decisions after they're formed. It means the check needs to be structured differently than a simple confirmation question. Rather than asking "does this make sense," a more useful post-decision prompt explicitly asks AI to argue against the decision: what's the strongest case that this is the wrong call, what would someone who disagreed with this point to, what assumptions is this decision resting on that might not hold.

This reframing produces meaningfully different output than a confirmatory question, because it's explicitly asking for the kind of challenge that a confirmatory framing tends to suppress. It's not a perfect substitute for a genuine human second opinion, which brings independent judgment, real stakes, and a different vantage point that AI simply doesn't have. But it's a significantly better use of AI at the post-decision stage than simple validation-seeking.

------------- Practical Moves -------------

First, notice when you're using AI to validate a decision you've already made versus using it to help form a decision you haven't reached yet. These are different uses and they deserve different framing.

Second, for early-stage thinking, bring problems to AI before you've settled on a direction, and ask explicitly for a full range of options and considerations rather than a check on an idea you already favor.

Third, for post-decision review, avoid simple confirmatory framing like "does this make sense." Instead, explicitly ask AI to argue against the decision, surfacing the strongest case for why it might be wrong.

Fourth, for genuinely high-stakes decisions, pair AI-assisted exploration with an actual human second opinion. AI can help you think more broadly, but it can't replace the independent judgment and real stakes that a human perspective brings to a consequential decision.

Fifth, pay attention to how often your AI-assisted decision checks come back positive. If nearly every check you run comes back reassuring, that's a signal the framing of your questions may be producing confirmation rather than genuine challenge.

------------- Reflection -------------

The sequence in which AI enters a decision-making process matters more than most people realize. Used early, as a thinking partner helping to map out a problem before a conclusion is reached, AI adds genuine value. Used late, as a validation check on a decision already made, it tends to reinforce rather than challenge, which can create a false sense of confidence in decisions that deserved more scrutiny.

Catching this pattern and adjusting when in the process AI actually gets brought in is a small shift with a meaningful payoff, because decisions caught and corrected early are dramatically cheaper, in time and resources, than the same decisions unwound after significant commitment has already gone into them.

Think about the last significant decision you ran past AI. Did you bring it in before you'd reached a conclusion, or after? What might have been different if the sequence had been reversed?

0 comments

🔁 Why AI Makes a Bad Second Opinion (And a Great First One)