On the morning of 7 August 2007, dozens of the world's most sophisticated quantitative hedge funds began losing money at a speed their own risk models said was almost impossible. Renaissance Technologies told investors its Institutional Equities Fund had lost 8.7% in the first eight days of the month. Highbridge Capital's Statistical Opportunities Fund fell 18% over the same stretch. Goldman Sachs's own Global Equity Opportunities Fund lost 28% of its value in two weeks, forcing the firm to arrange a $3 billion capital injection just to keep it running. These weren't reckless traders. They were some of the most rigorous, data-driven investors alive, running strategies built on decades of academic research into value, momentum and mean reversion.

What went wrong wasn't the data. It was the assumption sitting quietly underneath it: that the future would keep behaving like the past. Researchers Andrew Lo and Amir Khandani later traced the "quant quake," as it became known, to hundreds of funds running similar factor-based strategies without realising how much their portfolios overlapped. When one large fund began unwinding its positions, likely for reasons unrelated to those specific stocks, it forced prices to move in ways that triggered similar models at other funds to sell too. Each fund's model was working exactly as designed. Collectively, they had built a single, fragile structure without anyone seeing the whole of it.

The instinct to lean on data over instinct is, most of the time, the right one. Human judgement is reliably unreliable: it overweights recent events, it's swayed by how a question is framed, and it's prone to seeing patterns in noise. A well-built model doesn't have those flaws, and there's real evidence that data-driven decisions outperform gut instinct in stable, repeatable conditions.

The trouble is that markets aren't always stable and repeatable. AI-driven investment tools inherit the same blind spot the quant funds had in 2007, potentially in amplified form, because they're trained on more historical data and can obscure their own assumptions more completely. A model doesn't know when it's operating outside the conditions it learned from. It just keeps producing an answer, with the same apparent confidence, whether the underlying regime has shifted or not. The CFA Institute has drawn the comparison directly: AI risk increasingly resembles model risk, and just as backtests routinely overstate how a strategy would have performed in real time, AI outputs can overstate how reliable a decision actually is.

This isn't a historical problem. In July 2026, a heavily leveraged fund built around a concentrated AI-infrastructure thesis lost the majority of its value within weeks when the underlying stocks fell 30–47%, and margin calls forced a rapid unwind. The mechanism was different from 2007 (leverage rather than crowding) but the underlying lesson was the same: a correct or reasonable view, pursued without enough room to be wrong for a while, can still produce a disastrous outcome even if the underlying thesis ultimately proves right.

There's a second, quieter risk in how people actually use these tools. Because AI-generated answers arrive instantly and read as complete, they invite quick acceptance instead of scrutiny, especially when the output already matches what someone expected. The result can be a process that looks more thorough on the surface while running on less genuine analysis underneath.

None of this is an argument for ignoring data or distrusting models. Arbra's own research process depends on both. The distinction worth drawing is between two different jobs, both of which matter, and neither of which can substitute for the other. Data is very good at answering questions within a known structure: Given these conditions, what has historically happened next? Judgement is what's needed to ask a different, harder question: Has the structure itself changed, and would the historical pattern even apply here?

That second question is exactly where the quant funds of 2007 fell short, and it's the same question that determines whether an AI-assisted investment process holds up under stress or fails quietly, all at once, the way theirs did. A model can tell an investor what a thousand similar situations looked like in the past. It can't tell them whether the situation in front of them is actually one of those thousand, or something the data has never seen. That judgement call is the part of the job that doesn't automate.

In practice, this means treating a model's output as a strong opinion rather than an answer; one that still needs to be checked against the specific, current conditions the model wasn't built to recognise. It means asking, deliberately, what would have to be true for this pattern not to hold this time, rather than accepting a confident output at face value. And it means keeping enough experienced human judgement in the process to notice when a market is behaving in a way the data hasn't seen before, since a model trained entirely on the past cannot reliably recognise when the assumptions underpinning it have shifted.

The investors who came through August 2007 in reasonable shape weren't the ones who ignored their models. They were the ones who kept asking whether the conditions the models assumed still held, and who treated a moment of confusion as a signal worth investigating rather than noise to filter out. That instinct – knowing when to stop trusting the pattern and start asking why it might be different this time – is judgement doing the job data was never built to do.