The memorandum took a few seconds to produce. It compares three scenarios for a private credit allocation, summarises a dozen research notes, weighs the arguments on each side and closes with a measured conclusion, written in the tone of a senior analyst who has seen several cycles. Not long ago, a document of that shape would have occupied a small team for the better part of a week. Its owner reads it on his phone between meetings, finds it persuasive, and notices, with some surprise, that he has no real idea whether it is right.
That small moment of doubt matters more than the speed. For most of the history of investing, an advantage could be built on knowing more, sooner, than others did. Research was expensive and good analysis was scarce. Language models are steadily eroding that scarcity. When competent analysis can be summoned on demand, the edge does not disappear, but it moves. Producing an answer and judging whether it deserves to be believed have always been different skills. They are now being valued very differently.
Key Takeaways
- AI makes competent financial analysis fast and cheap, so the edge moves from producing answers to judging which ones deserve to be believed.
- A language model will strengthen or attack the same thesis with equal fluency, removing the effort that once kept confirmation bias in check.
- Polished machine output can look more reliable than the reasoning beneath it, and people tend to over-rely on automated recommendations.
- Bini, Cong, Huang and Jin found that LLM biases vary by task and model, and shrink when the models are told to decide rationally.
- Choukhmane, de Silva, Lin and Akuzawa found that AI advice moved simulated lifetime outcomes towards life cycle principles, but that the advice depended heavily on how people phrased their questions.
- The scarce asset becomes judgment: asking the right questions, spotting the hidden assumption and treating a persuasive output as a hypothesis.
- Who: Private investors, family offices and advisers who now use AI tools for research and analysis.
- What: Why better judgment, rather than more analysis, becomes the core financial skill once sophisticated answers are abundant.
- When: Now and through the coming decade, as language models take on more of the work once done by analysts.
- Where: Wherever investment decisions are prepared with AI, from a memorandum read on a phone between meetings to a family investment committee.
- Why: Because an answer can be fluent and persuasive without being right for the person asking, and what comes back depends on how the question is framed.
More Answers, Not Fewer Questions
It is tempting to assume that more analysis must produce better decisions. The relationship has never been that simple, and AI pulls it in two directions at once. It lowers the cost of gathering and summarising evidence, which is a genuine gain, and it multiplies the coherent interpretations available at any moment, which is a more ambiguous one.
The technology is also unusually obliging. Ask it to test an investment thesis and it will look for weaknesses. Ask it to strengthen the same thesis and it will return a more elegant version of the case in its favour. It does not distinguish between the investor who wants to be challenged and the investor who wants to be reassured, and serves both with equal fluency. Confirmation bias, the old habit of seeking evidence for what one already believes, used to be limited by the effort required to find that evidence. That constraint has largely gone. A conviction can now be furnished with professional-looking support within minutes, which turns intellectual independence from a default into a deliberate choice.
The Authority of the Well-Written Answer
Machine output carries a particular kind of authority. It is fluent, orderly and free of the hesitations that make human experts sound unsure. Researchers studying automation have long observed that people tend to over-rely on automated recommendations, accepting them with less scrutiny than they would apply to a colleague. A neatly formatted analysis, with scenarios and probabilities attached, can create an impression of reliability that runs well ahead of the evidence beneath it. The confidence of the presentation and the quality of the reasoning are separate variables, and only one of them is visible on the screen.
The research that is emerging suggests a more nuanced picture than either enthusiasts or sceptics tend to assume. Pietro Bini, Lin William Cong, Xing Huang and Lawrence Jin put prominent families of language models through a broad set of experiments originally designed to expose human biases. The results depended on the task. In tasks eliciting preferences, responses became more human-like as models grew larger and more advanced. In tasks concerning beliefs and probabilities, the most advanced models frequently answered as a rational decision-maker would. Simply instructing the models to decide rationally reduced the biases they did display. It is a working paper, and its findings vary by model and task, but the lesson is useful in both directions. These systems are neither neutral oracles nor straightforward mirrors of human error, and the way a question is framed can change what comes back.
Taha Choukhmane, Tim de Silva, Weidong Lin and Matthew Akuzawa approached the matter from the side of the user. A representative sample of people wrote their own prompts asking a leading model for spending and investing advice, and the researchers then simulated what following that advice would mean over a lifetime. The guidance broadly moved people towards the patterns economists tend to favour: wider participation in diversified equity funds, equity exposure that declines with age, and larger savings buffers. Yet the advice also differed systematically by gender, prior experience with AI and financial literacy, and in the case of gender roughly two thirds of the gap in recommended equity shares came from how men and women phrased their questions rather than from the model treating identical questions differently. The outcomes were simulated, not observed, and the study concerns ordinary households rather than private clients. Its central point travels well nonetheless. The answer depends heavily on the question, and the question depends on the person asking it.
Producing a recommendation is also not the same as showing that it suits a particular person. An answer can be well reasoned in general and still wrong for an owner whose liquidity needs, obligations, tax position or temperament never made it into the prompt.
Judgment as the Scarce Asset
If analysis becomes abundant, the advantage shifts to whatever remains scarce, and what remains scarce is judgment. Not the bold call, but the quieter work of knowing which questions to ask, noticing the assumption on which an elegant argument silently depends, and recognising when a probability has been dressed up as a certainty. It includes the humility to treat a persuasive output as a hypothesis rather than a conclusion, the independence to request the strongest case against one’s own view before acting, and an honest awareness of one’s own biases, including those a question can be quietly shaped to satisfy. It also includes the recognition that, beyond a certain point, further information stops improving a decision and merely postpones or decorates it.
This is why behavioural finance is becoming more than a catalogue of investor mistakes. Its deeper subject has always been how people reason under uncertainty, how they weigh evidence, and how readily they mistake fluency for truth. The widening distance between access to knowledge and the quality of judgment is the territory Behavioral Arbitrage sets out to explore, and the arrival of capable machines makes that territory larger rather than smaller.
None of this is an argument against the tools. Used critically, they can widen the range of scenarios an investor considers, expose blind spots and make good reasoning cheaper to practise. What they change is the definition of financial intelligence. For a long time the clever investor was the one with the better answer. In the decade ahead it may be the one who can tell when an answer, however well written, has not yet earned the right to be acted upon.
Related thinking on judgment as the skill that preserves wealth can be found in Behavioral Arbitrage and the Hidden Skill Behind Modern Wealth Preservation.
Sources: Pietro Bini, Lin William Cong, Xing Huang and Lawrence J. Jin, “Behavioral Economics of AI: LLM Biases and Corrections”, NBER Working Paper No. 34745, 2026, which applies classic behavioural experiments to prominent large language models and finds that biases vary by task and model scale, and that prompting for rational decisions reduces them. Taha Choukhmane, Tim de Silva, Weidong Lin and Matthew Akuzawa, “AI Financial Advice: Supply, Demand, and Life Cycle Implications”, NBER Working Paper No. 35574, 2026, which uses prompts written by a representative sample to show that AI financial guidance can move simulated lifetime outcomes towards life cycle principles, while varying with how users frame their questions.
We last reviewed this analysis on 16 November 2026.




