Artificial Intelligence at the Periphery

The debate about artificial intelligence in investment management is usually framed as a race: firms that adopt it fastest will win, firms that hesitate will be left behind. We think the framing is wrong, and dangerously so, because it treats the technology as a uniform good whose only variable is dosage. Our position is narrower and, we believe, more defensible. Artificial intelligence is welcome in our process exactly where its outputs can be mechanically verified, and barred from exactly the places where its output is a judgment. The line is not between old and new, or between cautious and bold. It is between checkable and uncheckable, and everything we understand about how investment processes fail says that is the line that matters.

Begin with the property of generative models that nobody puts in the brochure: fluency. A generative system produces text, code, and rationales that are smooth, confident, and plausible, at constant quality of surface regardless of quality of substance. In most domains this is a convenience. In an investment process it is a specific and serious hazard, because the rationale is the audit trail. Our entire discipline rests on a simple demand: every position must be traceable to a reason that was stated before the outcome was known, in language plain enough to be wrong. A tool that can generate a compelling rationale for any position, on demand, after the fact, does not strengthen that discipline. It dissolves it. Given a fluent machine, every decision can be made to look reasoned, including the decisions that were hunches, accidents, or errors, and once the rationales are decorative, review becomes theatre. The most dangerous sentence in a future investment committee is not "the model said so." It is a beautiful paragraph, written by a model, explaining why the model was right to say so.

This is what we mean by explanation laundering. A black box is at least honest about what it is: its outputs arrive labelled as unexplained, and a firm can decide, consciously, how much unexplained performance it is willing to own, sized how, watched how. A black box with a fluent narrator bolted to it arrives labelled as explained. The explanation is generated by the same statistical process as the prediction, from the same data, with the same blind spots, and it carries no independent evidence about why the prediction should be trusted. It is the prediction wearing a suit. Firms that would never accept an unexplained model will accept this, because it feels like diligence, and the feeling is the product being sold.

Now the counterargument, in its strongest form. It says that this is a transitional squeamishness: models that predict without explanation can make money in markets, machine learning systems already sit inside many quantitative processes, and a firm that refuses AI in its core will simply be out-researched and out-traded by firms that do not share its scruples. There is enough truth in this that we reject only its conclusion. Prediction without explanation is indeed an old and sometimes profitable practice, and we do not claim otherwise; our own Library carries an essay on exactly how much trust such models have earned. But where such systems last, they tend to be narrow, heavily monitored, sized as suspects, and surrounded by people who treat their outputs as claims to be verified rather than judgments to be obeyed. The scruple is not the alternative to performance. On our reading, it is close to a condition of it.

And notice what the periphery contains, because the word sounds like a demotion and is not. Searching documents across decades and languages. Cleaning and reconciling data, with every correction logged. Drafting code that is then subjected to tests it must pass or fail, mechanically, without negotiation. Monitoring systems for behaviour outside stated tolerances. Most of the labour of a research firm is periphery labour, and artificial intelligence does it faster, more patiently, and in some cases better than we do. We use it there without reservation, because there its errors are catchable, for reasons we make precise below. The verification is the job. Where verification is mechanical, the machine is welcome to the work, and the humans are freed for the part that is not mechanical.

What remains in the core, then, is small and specific: deciding what to hold, and being able to say why, in advance, in plain language, in a form that could turn out to be wrong. Every strategy we run must have a simple economic rationale: a reason, statable by a person, why the effect should exist, who is on the other side of it, and what would make it stop. This is not a philosophical preference. It is a risk control. When a strategy with a rationale underperforms, we can ask whether the rationale has failed or merely the weather has, and those two situations demand opposite responses. When a strategy without a rationale underperforms, there is nothing to ask. There is only a decision to keep trusting a number or to stop, taken in the dark, usually at the worst moment, by frightened people. The rationale is what makes a drawdown survivable as a process rather than merely as a balance.

It helps to sort every task in a research firm by how its errors are discovered, because that, and not the sophistication of the tool, decides where a machine may work. Some errors are caught by the machine itself: a test fails, a checksum breaks, two sources disagree. Some are caught by a person in minutes: a chart that looks wrong, a citation that does not say what it was claimed to say, a reconciliation with a visible gap. And some are caught only by outcomes, months or years later, after the money has been committed: a position that was sized on a story, a relationship that was never there. We call this the discovery ladder. Machines are welcome on the first two rungs without reservation, because there their errors announce themselves at a cost we can afford. The third rung is where we keep people, not because people are wiser than models, but because a person can be asked, before the outcome, what would prove them wrong, and can be held to the answer. A model can be asked the same question. It will answer it beautifully. That is precisely the problem.

We will be told, and already are told, that this position will age badly, that the models will become reliable enough that demanding a human-statable rationale will look like demanding that pilots fly every leg by hand. Perhaps. Standards should be revised when the evidence changes, and if a generation of audited, out-of-sample, regime-spanning results shows that unexplainable systems can be owned as safely as explainable ones, we will revise ours in public, in this Library, with the evidence attached. Until then, the burden of proof sits where it sits for every other claim we make: on the claimant. If we cannot explain it, we do not trade it. The periphery is where the machines work. The core is where we answer for the decisions, and we intend to remain able to.