Can AI Help Me Make Better Decisions? How AI Improves Human Judgment

You're about to accept a job offer. The salary is good, the role is exciting, and the deadline is tomorrow. Your gut says yes. Your spreadsheet says maybe. And somewhere between the two, you know you're probably missing something you can't quite name.

Yes - AI can help you make better decisions. The mechanism isn't magic and the result isn't certainty, but the effect is real and measurable: AI acts as a structured thinking partner that surfaces your blind spots, stress-tests your assumptions, and forces your reasoning into the open where it can actually be examined. The improvement comes less from AI knowing the right answer and more from AI preventing you from stopping too early at the wrong one.

That's the core of it. AI doesn't decide for you. It makes the process of deciding harder to shortcut.

When I work through consequential decisions now - career pivots, investment logic, arguments I'm about to make in writing - I treat AI as an adversary by default. I state my conclusion, then ask it to dismantle it. What comes back is rarely devastating, but it's almost always useful. Something I glossed over. A framing I borrowed without examining.

The job offer I mentioned at the top? I walked through it with Claude in about twenty minutes. The outcome wasn't a decision. It was clarity about what was actually bothering me.

---

The Cognitive Problem AI Actually Solves

Human decision-making fails in patterned, predictable ways. Daniel Kahneman's decades of research, summarized in Thinking, Fast and Slow (2011), established that we rely on cognitive shortcuts - heuristics - that are efficient but systematically biased. Availability bias makes recent events feel more probable than they are. Confirmation bias makes us collect evidence for what we already believe. The planning fallacy makes us perpetually underestimate time and cost.

These aren't character flaws. They're features of a brain optimized for speed, not accuracy.

AI doesn't share most of these biases by design - though it carries others, which matters and I'll come back to. More importantly, AI has no social stakes in your decision. It doesn't get tired, doesn't want to please you enough to stop pushing back, and doesn't experience the emotional constriction that narrows human thinking under pressure.

A 2023 paper published in PNAS by Noy and Zhang found that professionals using large language models for complex cognitive work showed not just speed improvements but measurable quality gains - particularly in tasks requiring synthesis and structured argumentation. Decision-framing falls squarely in that category.

What this means practically: AI is most valuable at the front end of decision-making, when you're still building the frame, not at the back end when you're choosing between two pre-defined options.

---

Structured Externalizing: Making Your Reasoning Visible

There's a concept in cognitive psychology called distributed cognition - the idea that thinking doesn't happen purely inside the skull but across tools, environments, and other minds. Coined by Edwin Hutchins in his 1995 book Cognition in the Wild, it was developed to explain how ship navigators and cockpit crews make decisions that no individual member could make alone.

AI extends this framework to individual decision-making. When you write out your reasoning to an AI - not just your conclusion, but the chain of logic that got you there - something happens that doesn't happen when you think silently. Your assumptions become visible. You can see where the argument is held together with vague language. You can notice where you've assumed causation from correlation.

The specific prompt structure matters here. Asking "What do you think I should do?" extracts an AI opinion. Asking "Here's my reasoning. Where is it weakest?" extracts something more useful - structured pressure on your own logic.

I've found the most useful framing is what I call the Steel-Man Audit in my book The Last Skill: before you commit to a decision, ask AI to generate the strongest possible case for the option you're currently not choosing. If you can't refute that case, you're not ready to decide.

---

Where the Evidence Gets Complicated

A 2022 study by Dell'Acqua et al. at Harvard Business School - involving 758 consultants using GPT-4 - found that AI assistance raised performance significantly on tasks within the model's capability. But on tasks outside that range, AI users performed worse than non-AI users. The researchers called this the "jagged frontier" of AI capability.

The implication for decision-making is uncomfortable. If you're making a decision in a domain where AI is well-trained (financial modeling, logical argumentation, medical literature synthesis), the augmentation is real. If you're making a decision that requires deep local context, embodied experience, or the kind of tacit knowledge that can't be textualized - and this is a genuinely hard category to define - AI assistance can actively degrade your judgment by making you more confident in an incomplete picture.

Which means the first decision you need to make is whether AI can actually help with this decision. That's not a trivial question.

There's also the sycophancy problem. Large language models are trained on human feedback and tend toward outputs humans rate positively - which often means outputs that agree with the user. Some models are better than others at sustained disagreement, but as a rule, if you present your reasoning confidently, AI will often shore it up rather than challenge it. You have to explicitly invite pushback. Repeatedly. With some friction built in.

---

The Pre-Mortem Method, Scaled

The pre-mortem is one of the most empirically validated decision tools in organizational psychology. Developed by Gary Klein and formalized in his 2007 research, the technique involves assuming a decision has already failed - catastrophically - and working backward to identify what went wrong.

Klein's original research found that prospective hindsight (imagining a future failure as already occurred) increased the ability to identify reasons for failure by approximately 30% compared to standard risk assessment.

AI runs this process faster and with fewer social dynamics. In a team setting, the pre-mortem requires someone willing to voice the bad scenarios out loud, which is socially costly. AI has no such reluctance. You describe the decision, tell it to assume it backfired completely eighteen months from now, and ask it to generate failure narratives.

What you get back isn't necessarily correct. But it's generative. It populates your decision with scenarios your optimism-biased brain was quietly refusing to render.

The technique works best when you resist the urge to immediately rebut each failure scenario. Sit with them. The discomfort is the point.

---

Honest Constraints

The evidence that AI improves decision quality is real but narrow. Most of it comes from professional, high-structure tasks - consulting, writing, coding - not from the messier decisions that actually define lives. Relationship choices, values conflicts, decisions made under grief or fear or obligation. There's very little rigorous research on whether AI-assisted deliberation produces better outcomes in these domains, and some theoretical reason to think the structured rationalism AI encourages could be actively misaligned with decisions that are fundamentally emotional or ethical in nature.

AI also doesn't know what you haven't told it. Context you consider obvious but haven't stated, history you've forgotten to mention, the felt sense of a situation - none of this enters the process. A model working from incomplete framing can produce confident analysis of the wrong problem.

And finally: better process doesn't guarantee better outcomes. You can reason more clearly and still be wrong. The improvement AI offers is probabilistic, not protective.

---

FAQ

Will AI make the decision for me if I ask it to?

It will generate a recommendation if you push for one, but that output reflects pattern-matching on training data, not knowledge of your life. The useful function is pressure-testing your reasoning, not outsourcing the conclusion. Treat any AI recommendation as a starting point for interrogation, not an endpoint.

What's the best way to prompt AI for decision help?

Lead with your current conclusion and the reasoning behind it, then explicitly ask for weaknesses, counterarguments, and failure scenarios. Avoid asking "what should I do?" before you've stated your own position - otherwise you're inviting the model to decide rather than to audit.

Does this work for small decisions or only big ones?

The overhead makes it impractical for trivial choices. Where it earns its time is in decisions with significant irreversibility, high stakes, or situations where you notice you've been avoiding thinking something through. The avoidance itself is usually a signal.

---

The question of whether AI improves decisions connects naturally to a deeper one - whether AI changes how we think over time, not just in individual moments. That's worth examining alongside research on cognitive offloading and the long-term effects of tool-mediated reasoning. The decision you make today about how to use AI may matter more than any single decision you make with it.