How do you separate emotion from data when making a big decision?

Separating Feelings From Facts: A Founder Decision Method

You can't remove emotion from a decision — Kahneman's Thinking, Fast and Slow shows System 1 delivers its verdict first and recruits data afterward. So audit it instead: write the felt verdict down, run a premortem against a current model like GPT-6 Astra (rolled out September 3, 2026) or Claude Opus 5, check the outside-view base rate, and only then compare the two ledgers.

This question comes up in my coaching work more than almost any other, usually from a founder holding a term sheet or a shutdown decision, and it smuggles in a wrong assumption I want to name first: that emotion is contamination you can filter out, leaving clean data behind. Daniel Kahneman spent a career demonstrating the opposite. In Thinking, Fast and Slow he shows that System 1 — fast, associative, emotional — delivers its verdict before the analysis begins, and System 2 mostly arrives afterward to write the justification. The line of his I quote most often is that subjective confidence is a feeling, not a judgment: founders routinely mistake the intensity of their conviction for the quality of their evidence.

I learned this on my own decisions before I found it in the literature. When I was deciding whether to keep or kill a product line, the spreadsheet said the same thing for three consecutive months while I kept finding fresh reasons to postpone. That is loss aversion doing exactly what Kahneman documented — losses loom roughly twice as large as equivalent gains — and no amount of staring at the data fixed it, because the data was never the problem. The verdict had arrived first; the analysis was decoration.

So the honest goal is not separating emotion from data. It is making each visible on its own ledger so neither secretly authors the other. This is where current AI models have earned a place in my process: they have no skin in your game. GPT-6 Astra, which OpenAI began rolling out on September 3, 2026 with state-of-the-art results on professional and knowledge work, and Claude Opus 5, the everyday frontier model Anthropic released in July 2026 at half the price of its flagship, are both competent premortem partners. One warning from someone who builds with these systems daily: models mirror your framing. Hand one your preferred option first and it will anchor on it as reliably as a junior analyst who knows what the boss wants to hear. Prompt discipline matters more than model choice.

Kahneman's favorite institutional debiasing tool was the premortem — assume the decision has already failed, then write the history of the failure — because it legitimizes doubt in rooms where doubt reads as disloyalty. He also wrote that whenever we can replace human judgment with a formula, we should at least consider it. The protocol below operationalizes both, and I run it on any decision where being wrong costs more than a quarter of runway.

I call it the Split-Ledger protocol, and it has five steps. One: the felt-verdict entry — before opening any data, write the decision in one sentence plus the answer your gut has already given; it counts only when it is on paper, because writing the emotion down stops it leaking into the other ledger. Two: the data ledger — list the five to seven facts a disinterested stranger would need, each with a source and a number, no adjectives allowed. Three: the AI premortem — in a fresh chat, tell GPT-6 Astra or Claude Opus 5 that it is twelve months later and the choice failed, and ask it to write the post-mortem; run it once per option, in separate chats, so the model cannot anchor on your favorite. Four: the outside view — ask the model for base rates on similar decisions, then verify the two numbers that matter most yourself, because models still invent statistics with a straight face. Five: compare ledgers — if the felt verdict and the data verdict disagree, name the specific bias driving the gap: loss aversion, sunk cost, anchoring on what you paid. You are still allowed to choose the gut answer. But now it is a labeled input, not the hidden author.

Where this fails, so you do not overuse it: values decisions. Choosing a cofounder, or deciding whether you still want the company you are building — there the emotion is the data, and forcing it through a premortem produces confident nonsense. The protocol also feeds chronic ruminators, who will happily run step three forever as a sophisticated form of postponing; if you have run it twice on the same decision, you are no longer analyzing, you are hiding. Reserve it for consequential, hard-to-reverse calls, decide within a week of running it, and let the small stuff stay fast. Kahneman never promised debiased humans — only that we can build procedures that catch the worst of it before we sign.

Sources: OpenAI — GPT-6 Astra announcement, Anthropic — Introducing Claude Opus 5, 9to5Mac — GPT-6 Astra rollout details.