n o ren
AI & Technology

Does the “Adjacent Possible” Limit AI Adoption?

A consulting team spent a full quarter refining AI outlines before anyone asked whether the underlying question was wrong.

2026-08-271 min read

Stuart Kauffman, the theoretical biologist, named the adjacent possible: at any moment a system can only reach the states that sit one step from where it already is. Steven Johnson later borrowed the idea to explain innovation, arguing that good ideas arrive by opening doors the current room makes available rather than by teleporting across the building. A generative model is a very fast door-opener. It surfaces the step-away options in seconds, which is genuinely useful, and which is also where the trouble starts.

A fluent list of adjacent options reads like a complete map. When a model returns nine framings of a problem in twenty seconds, the tenth framing — the one nobody prompted for — stops feeling like a gap and starts feeling like it does not exist. Teams stop searching not because they concluded the search was over, but because the output looked finished. The frontier gets confused with the inventory.

Picture a boutique consultancy that adopts a drafting assistant for client memos. Throughput rises immediately, and partners review several times as many outlines per week as before. A quarter later, every memo answers some version of the question the client already asked, and none proposes the service line that would have required asking a different question. Nothing failed. The team simply never left the room it started in.

The loop tightens because each cycle trains the reflex. The easier the adjacent option becomes to generate, the more expensive the non-adjacent one feels by comparison, and that reflex outlives any single tool.

Key insights
The adjacent possible is a real constraint on innovation, and a generative model both widens it and makes it look like the entire space.
Fluency is the failure mode: a complete-looking list ends the search earlier than an incomplete one would.
The fix is procedural rather than technical — reserve an explicit slot for the option the tool did not produce.
Why it matters

A team that mistakes the model’s output for the option space will spend its quarter optimizing a question it never chose.

The habit compounds: every fast adjacent answer makes the slow non-adjacent one look more expensive to attempt.

Use this tomorrow
1
Take the last AI-drafted document you sent and count how many of its section headings restate a question the requester had already posed; if all of them do, the whole document came from inside one room.
2
Before your next planning meeting, write down three options the model did not offer, then put a name and a date next to whichever one is cheapest to test.
Go deeper

Kauffman introduced the adjacent possible to describe biological evolution: a genome can only reach forms a small number of mutations away, so the set of reachable forms expands one step at a time. Steven Johnson carried the idea into the history of invention in Where Good Ideas Come From, where he argues that most breakthroughs are recombinations of parts that had only just become available. Both versions share a quiet assumption — that the explorer can still see past the frontier. A tool that renders part of that frontier in high resolution changes the assumption, and not obviously for the better.

The collapse is sharpest in organizations that already measure output rather than originality. When the visible metric is drafts shipped per week, the adjacent option wins every internal argument on cost, and the non-adjacent one has to justify itself against a number it cannot produce yet. Nobody in that system decides to stop exploring; exploration just loses on the scoreboard, quietly and repeatedly. Teams that keep the habit tend to protect it structurally — a standing agenda item, a budgeted fraction of time — rather than relying on anyone’s appetite for risk in the moment.