30 September 2026
Artificial intelligence is becoming extraordinarily good at producing answers. Investment professionals face a different challenge: deciding what to do when the right question is not yet clear.
That distinction matters because financial markets rarely operate in environments where every possible outcome and its probability can be reliably specified. Unexpected geopolitical events, technological disruption, changes in investor behaviour and structural economic shifts confront investors with something fundamentally different from measurable risk: uncertainty.
In Extinction Is the Easy Story for AI, published by Enterprising Investor, Markus Schuller challenges the assumption that increasingly powerful AI systems will inevitably become superior decision makers across every domain. His argument is particularly relevant to investment management: machines can perform exceptionally well when problems are clearly defined and supported by sufficient data, but the frontier of investment judgement frequently begins where those conditions end.
Risk is not the same as uncertainty
The distinction starts with a concept familiar to economics since Frank Knight: risk and uncertainty are not interchangeable. Under risk, potential outcomes are known and probabilities can be estimated. Such environments are well suited to computational tools capable of processing enormous datasets, detecting patterns and optimising decisions. Under genuine uncertainty, however, neither all the possible outcomes nor their probabilities are necessarily known in advance. Historical data may offer limited guidance precisely because the event or structural change investors need to understand has not happened before.
This creates a natural boundary for machine intelligence. AI systems learn from historical or simulated distributions. According to the research discussed by Schuller, they can perform extremely well in stable, data-rich environments and, on well-defined forecasting questions, recent models have achieved results comparable to expert forecasters.
But identifying the problem itself is different from solving an already specified one. When markets face an emerging phenomenon, investors may need to determine which evidence matters, formulate new hypotheses, interpret incomplete information and decide which trade-offs are relevant before a model has a well-defined problem to optimise.
Fluency should not be confused with understanding
Generative AI makes this distinction particularly difficult because its outputs can be extraordinarily persuasive. A coherent investment thesis, polished scenario analysis or sophisticated portfolio recommendation can create an impression of understanding even when the system generating it is primarily identifying and recombining patterns.
Schuller argues that this risks creating a category error: equating optimisation with understanding. Research cited in the article shows that machine-learning systems can learn shortcuts that perform successfully against benchmarks without necessarily capturing the underlying concept assumed by their users. Larger models, meanwhile, do not automatically eliminate such limitations.
For investors, the problem is obvious. Markets contain many situations in which errors cannot immediately be identified against an objective benchmark. A coding error may produce a failed programme. An investment thesis based on the wrong causal interpretation may remain apparently plausible for months or years. AI fluency can therefore become most dangerous precisely where the quality of its reasoning is hardest to verify.
Human judgement has weaknesses too
This is not an argument that human investors are naturally superior. Uncertainty is difficult for people as well. Investors are vulnerable to cognitive biases, emotional responses and the desire to impose simple narratives on complex environments. Schuller describes uncertainty as cognitively hostile territory for both humans and machines.
The more interesting question is therefore not human or machine, but under what conditions their combination produces better decisions. The article points to evidence that human ingenuity, when supported by an appropriately designed decision environment and selectively augmented by AI, can outperform machine intelligence alone under uncertainty. Human strengths include framing emerging problems, generating hypotheses, interpreting incomplete evidence and reasoning towards observations that do not yet exist.
AI brings complementary capabilities: searching, calculating, synthesising information and challenging existing conclusions at extraordinary speed. Used together, those capabilities can be powerful. But only if augmentation does not gradually become substitution.
The hidden cost of cognitive convenience
This is where the argument intersects directly with the investment profession. The risk is not simply that AI occasionally produces an incorrect answer. It is that repeated delegation of analytical work may gradually weaken the skills professionals need to recognise when an answer is wrong.
Schuller cites research indicating that reliance on generative AI can reduce critical-thinking effort and, in some settings, weaken learning or unaided professional skills. If that dynamic extends into investment organisations, the long-term consequence could be a paradox: firms gain increasingly sophisticated analytical tools while becoming less capable of independently interrogating their outputs. For investment teams, this raises questions about how junior professionals develop expertise. Building an investment thesis manually, challenging assumptions, examining contradictory evidence and learning from unsuccessful decisions are not simply inefficient steps on the way to an answer. They are part of the process through which judgement develops.
Removing too much of that friction could accelerate today’s analysis while weakening tomorrow’s analyst.
Design the decision process, not just the AI system
Schuller’s proposed response is what he calls “cognitive immunization”: designing decision environments that preserve independent reasoning, critical scrutiny and accountability while still exploiting the capabilities of AI. For investment organisations, that principle could have practical implications.
AI-generated conclusions need to remain open to challenge. Professionals should be able to reconstruct the evidence behind important decisions rather than merely approve an output. Teams need room to develop competing hypotheses rather than converge immediately around the machine-generated consensus. And responsibility for investment decisions must remain clearly human.
This is especially important because accountability cannot simply be delegated alongside analysis. If professionals cannot explain the evidence, assumptions and reasoning behind a decision, retaining a human approval step does little to provide genuine oversight. As Schuller puts it, a human who cannot reconstruct the evidence or interrogate the representation may technically remain “in the loop”, while playing an increasingly ceremonial role.
The scarce resource may be judgement
Investment firms are understandably investing heavily in AI infrastructure and capabilities. The potential benefits—in research productivity, data processing, modelling and information retrieval - are substantial. But competitive advantage may ultimately depend on something less scalable.
If AI makes information processing and sophisticated analytical outputs increasingly abundant, independent judgement becomes more, rather than less, valuable. The ability to frame a problem that has never previously existed, distinguish evidence from apparent certainty, challenge a plausible consensus and remain accountable for a decision cannot be treated simply as another task to automate.
For investment professionals, the objective should therefore not be to keep humans involved for its own sake. Nor should it be to preserve inefficient processes that technology can genuinely improve. The more demanding task is to identify where machines have the comparative advantage, where humans still do, and how investment organisations can combine the two without allowing one to erode the capabilities of the other.
AI can make answers cheaper, faster and more plentiful. In markets characterised by genuine uncertainty, knowing which questions deserve to be asked - and which answers should not yet be trusted - may become the scarcer investment skill.