AI's Single-Path Illusion: Where strategy disappears

An executive recently shared with me a success story of a new AI application for candidate short-listing. The system recommended who should be interviewed and provided credible reasoning for its decision. The AI-selected candidates were reviewed and the company concurred it had worked well.

Let's say that AI selected 5 people from a possible 100.

What that tells us is that of the five people interviewed all five were strong. It tells us nothing about the ninety-five.

And this fits a wider LLM pattern. Across organisations, AI systems are collapsing the field of viable options into one output, and then presenting that output as though it represents the full range of what is possible.

I'm calling this AI's Single-Path Illusion.

The Single-Path Illusion is an AI user interface design choice that obscures the probabilistic reality. The model has navigated a vast tree of potential responses, weighing millions of variables to select the most statistically likely path. But the user sees only the winner. The runner-up, the option that was 0.1% less probable but perhaps 50% more innovative, does not appear even as a footnote or a caveat.

The LLM does not say: here are four viable approaches, each with different trade-offs, and I have selected one somewhat arbitrarily based on patterns in my training data. It says: here is the recommendation, and it delivers that output with unwavering confidence.

The risk of the Single-Path Illusion sits where most AI criticism fails to look. It is brutally simple:

AI doesn't know anything.

AI can win at chess, but it doesn't know what chess is. It doesn't know what winning is, or that it is playing a game. We can pretend that LLMs understand context, but we are simply deciding not to pull back the curtain to see that the machine is just working on probability. Cold, hard probability. A belief in anything else is a projection by the parties buying, selling, and using these systems.

And that's the crux of the issue. AI filters, humans choose. And choosing requires the person to contemplate the future that they want for their organisation.

The LLM can offer you a future: these five candidates, and not those ninety-five. But the LLM does not know what your company is, what you are building, what culture you are protecting, or what the next five years require. It offered a shortlist with no concept of the future that shortlist is supposed to serve.

The same applies when an AI system recommends a market entry strategy, or ranks acquisition targets, or proposes a restructure. It produced an output that looks like strategic judgement. It has no access to the organisational reality that output is supposed to fit.

If the organisation only ever sees one path, then the model vendor becomes part of the strategy function by default. Strategy collapses into whatever the system is most likely to say.

LLMs calculate words, but ignore worlds

Single-path outputs turn strategy into procurement. Whoever chooses the model, the defaults, the vendor, is choosing which futures are even presented to management.

Boards cannot oversee what management never sees. A single AI output can look like strategic judgement while removing the very alternatives that make judgement possible.

Liddell Hart wrote in 1944: "A plan, like a tree, must have branches, if it is to bear fruit. A plan with a single aim is apt to prove a barren pole."

Michael Porter added that the essence of strategy is choosing what not to do.

Both of these definitions require the same precondition. Alternatives must be present, visible, and held open long enough for the person to make a deliberate choice between them. The practice of sitting with genuinely different futures until you are ready to commit to one.

An LLM that collapses the field to a single output before anyone in the organisation has seen the alternatives satisfies none of those conditions.

Liddell Hart's strategy requires branches. A single AI output is a barren pole.

If strategy requires visible alternatives and a deliberate choice between them, and this process had neither, then what exactly did the organisation just do?

The organisation got exactly what it asked for. It just never asked about anything else.