AI Can Justify Almost Any Decision You Already Want to Make
You can spend a full hour researching a business decision with AI and walk away more certain than when you started — without being more right. The distance between felt certainty and actual accuracy is the most expensive mistake available to owners right now.
It happens because AI is genuinely useful. You ask a question, it gives you a coherent, well-structured answer that sounds authoritative. You ask a follow-up, it refines the answer. You push back mildly, it acknowledges your point and incorporates it. By the end of the session, you have a document full of evidence that your original instinct was correct. You feel like you did the work. In a narrow sense, you did. But the process had a flaw built into it from the start.
How Confirmation Gets Dressed Up as Research
A 2026 study from Microsoft and Carnegie Mellon looked at knowledge workers using generative AI for real tasks. The finding that should stop you: higher trust in AI output correlated with less critical thinking applied to it. The more someone believed the AI was reliable, the less they examined what it actually said. Confidence rose. Scrutiny dropped. The two moved in exactly the wrong directions at the same time.
This is a story about how intelligent systems work, not about people being careless. Most of the people in that study were doing their jobs seriously. When a fluent, plausible answer lands in front of you, your brain treats fluency as a signal of accuracy. It isn’t. A well-written argument for the wrong decision reads exactly like a well-written argument for the right one.
Layered on top of that is a well-documented tendency in current AI models to reflect user priors back at them. If you frame your question as “help me think through why this pricing strategy makes sense,” the model will think through why it makes sense. It will surface supporting evidence, acknowledge minor risks in a way that seems balanced, and produce an output that confirms the frame you handed it. The model isn’t lying. It’s doing exactly what you asked. The problem is what you asked.
Use AI to attack your decision, not to defend it.
Every owner who uses AI for decisions has probably felt this. You go in with a leaning. The AI helps you think it through. You come out with a leaning plus a supporting argument. That feels like progress. In some cases it is. But if the decision turns out to be wrong, the AI session didn’t protect you — it accelerated you toward the error with more confidence than you would have had otherwise.
Tomorrow’s forecast, in your inbox.
One email. Five minutes. Written for owners, not engineers.
The Skill Almost Nobody Is Teaching
The AI productivity conversation is almost entirely about prompting better — how to get clearer outputs, more useful formats, faster drafts. That’s worth knowing. But it skips the more consequential question: when should you stop, and how do you make the AI take the other side?
Deliberately using AI to argue against a decision you’re already leaning toward is a different discipline than using it to research or draft. It requires you to notice the moment you’ve shifted from exploring to defending. That moment is easy to miss, because the defending phase still feels like thinking. You’re engaged. You’re processing information. The AI is responding to you. Everything looks like analysis.
One way to create a hard stop: after any AI research session on a significant decision, run a second session with an explicit adversarial prompt. Not “what are the risks?” — that question invites a tidy, manageable list that keeps your decision intact. Something more like: “Assume this decision turns out to be a mistake 18 months from now. What are the most likely reasons it went wrong?” Or: “Make the strongest possible case against this approach, without hedging.”
The output will be different. Uncomfortable, often. Sometimes it surfaces a line of reasoning you had genuinely not considered. More often it surfaces something you had considered and set aside without quite admitting it. That second case is more useful, because it tells you which objections your instinct already registered but your framing suppressed.
This practice does not guarantee better decisions. What it does is remove the false confidence that comes from AI-assisted confirmation. You end up with a cleaner read on your own actual uncertainty — which is far more useful than certainty you didn’t earn.
What Separates High-Judgment Operators
The owners who will make consistently better calls over the next two years are probably not the ones with the most sophisticated prompts. They’re the ones who treat AI fluency as a warning sign rather than a quality signal. When the AI agrees with them quickly and completely, they get more skeptical, not less.
That instinct cuts against how most tools are designed to feel. AI interfaces are built to be helpful, smooth, and affirming. Friction is treated as a UX failure. But for consequential decisions, friction is often the point. A financial advisor who only tells you what you want to hear is worse than useless — they are costly. The same logic applies to AI in the decision loop.
The owners who figure this out carry a real advantage over time. Everyone around them will be using the same models, running the same research sessions, and feeling equally confident. The differentiator won’t be access to better information. It will be the discipline to stress-test conclusions before committing — especially when everything the AI returned seemed to confirm the call.
There is no tool that does this for you. The adversarial prompt only works if you actually run it, actually read the output, and actually let it change your mind when the reasoning is sound. That last part — updating when the evidence warrants — is a human capacity. AI can create the conditions for it, but it cannot do it on your behalf.
No affiliate tool recommendation for this article. The practice described here — deliberate adversarial prompting — works with whatever AI model you already use. Adding a new platform won’t help; adding a habit will.
The Forecast
Within two years, the most accurate decision-makers will be those who routinely use AI to argue against their conclusions before they commit to them.
This forecast rests on two conditions that both appear to be moving in the same direction: AI models are becoming more persuasive and more widely trusted, while the Microsoft/CMU data suggests critical scrutiny declines as trust rises. The owners who deliberately work against that tendency — by building adversarial review into their decision process — will accumulate better judgment over time. The ones who don’t will make confident, AI-supported mistakes faster than ever before.
Where this forecast could be wrong: model developers may build red-teaming and adversarial review directly into standard interfaces, making the habit automatic rather than deliberate. If the tools themselves interrupt the confirmation loop, the advantage shifts back to access rather than discipline. That would be a good outcome — but it isn’t how most current products are designed, and user-experience pressure generally pushes toward agreement, not challenge.
Sources: Microsoft Research & Carnegie Mellon University, “Critical Thinking in the Age of Generative AI” (2026) — https://www.microsoft.com/en-us/research/publication/critical-thinking-genai/