Seven strange AI analogies that change how you work.

Bad AI analogies make people ask for the wrong product. If AI is an oracle, the product becomes answer extraction. If AI is an intern, the product becomes delegation theater. The analogies worth keeping are stranger because they change the work surface: what gets written down, what gets checked, what gets handed to a tool, and what the human still has to practice.

Why analogies are not decoration

Cognitive science gives a useful warning here. Dedre Gentner's structure-mapping work argues that strong analogies preserve connected systems of relations, not surface resemblance. That is why "AI is a brain" is usually weak for operators. It matches the mystique, but it does not tell you what to log, what to bound, or what to verify.

The better move is to pick analogies that force operational decisions. If you say the context window is a cockpit, you have to decide which instruments belong on the panel. If you say an agent is a surgical count, you have to count tools before and after the risky step. The analogy becomes a build rule.

1. The prompt is an agar plate

The prompt is not a spell. It is a medium that makes invisible assumptions grow where you can inspect them. A weak prompt is useful if it reveals the contaminants: missing actors, unearned certainty, vague criteria, and an answer shape you were secretly steering toward.

The operating rule: before asking for an answer, write what you expect to grow. Then ask the model to name the assumptions your prompt cultured. The output is not the product; the visible assumption colony is.

2. The context window is a cockpit

A cockpit is not an attic. You do not throw every relevant object into it and hope the pilot remembers where the altimeter landed. You design a panel where the state that changes the next action is visible at the moment of action.

Edwin Hutchins' distributed-cognition work is useful here because it treats cognition as happening across people, artifacts, procedures, and environments. In AI work, that means the context window is part of the thinking system. Bad context makes the system think badly, even when the model is strong.

The operating rule: separate context into four cockpit instruments: facts, open questions, constraints, and authority. If a fact does not change the next action, do not pin it to the panel.

3. The agent is a surgical count

An agent is not one clever person. It is a team process that can lose track of what entered the work site. The surgical-safety checklist study is not interesting here because medicine is dramatic. It is interesting because a small communication artifact changed outcomes across complex teams.

The operating rule: every meaningful tool call should name the object, action, reversibility, proof, and owner before the action. Afterward, count what changed. That is the agent version of "nothing left inside the site."

4. The model is a prosthetic sense

A prosthetic sense does not merely add ability. It changes what the body attends to. This is the uncomfortable side of the extended-mind idea: tools can become part of the cognitive loop, which means they also reshape what the human stops practicing.

Microsoft Research's 2025 critical-thinking study is a useful warning signal: higher confidence in GenAI was associated with less critical-thinking effort, while higher self-confidence was associated with more. The question is not "did AI help?" It is "which human faculty did the workflow exercise, and which one did it let atrophy?"

The operating rule: any AI workflow that saves effort should declare the muscle it still trains. If the answer is "none," you built a crutch, not a prosthetic sense.

5. Benchmarks are weather reports

A weather report matters, but it is not climate and it is not your exact street corner. Benchmarks tell you something real about conditions. They do not tell you whether your task, repo, review bar, latency budget, or operator expertise will create a storm.

METR's early-2025 developer-productivity study is the useful shock: in that setting, experienced open-source developers took longer with AI, even while believing AI had sped them up. METR also warned readers not to overgeneralize the finding. That is exactly the weather-report analogy. The report matters because it forces local measurement, not because it ends the argument.

The operating rule: pair every benchmark claim with a local forecast card: task type, human expertise, review cost, accept/reject rate, and the failure you will count as negative even if the model sounded fluent.

6. A hallucination is a compass spinning near metal

Treating hallucination as "the model lied" often hides the engineering move. A spinning compass is not morally defective. It is an instrument in a bad field. The question is what nearby metal, missing anchor, or bad map made the reading unreliable.

Anthropic's interpretability work gives a better mental model than folk psychology: models contain features and circuits we are only beginning to inspect. A wrong answer is not a confession from a little person inside the machine. It is a reading that needs external anchors.

The operating rule: when the model makes a claim that matters, add one external anchor before continuing: source link, code path, small execution, customer artifact, or falsifying counterexample.

7. A good AI community is sourdough starter

Prompt libraries go stale because they preserve strings, not living judgment. A starter culture survives because it is fed, tested, discarded, and carried forward. The strange community lesson is that the valuable artifact is not the prompt. It is the before and after delta that shows what changed in the operator.

The operating rule: share deltas, not magic words. For every serious AI workflow, publish the unaided answer, the prompt, the model output, the external proof, the final decision, and what confidence changed.

The analogy-to-receipt protocol

Use this when a team is stuck in generic AI language. It is short enough to run inside a single work session and concrete enough to expose whether an analogy changed the behavior or only made the deck sound smarter.

• Name the analogy: Write the exact sentence, then delete any metaphor that does not imply a concrete action.
• Extract the control surface: Ask what must be visible, bounded, counted, practiced, or falsified if the analogy is true.
• Create the receipt: Add one artifact the next operator can inspect without trusting the author.
• Run the reversal: Ask when this analogy would mislead you and write the counter-condition into the workflow.
• Ship one small behavior change: A new context panel, a tool-call count, a delta log, or a local forecast card is enough.

The claim worth stealing

The hidden analogies that matter are the ones that change the work. Cockpit changes how you prepare context. Surgical count changes how you grant tools. Apprenticeship changes how you teach. Prosthetic sense changes what you keep practicing. Weather report changes how you read benchmarks. Starter culture changes what your community preserves.

That is the test. If the analogy does not leave behind a receipt, it was only a line. If it changes the surface, it can change the way people work.

Sources

• Dedre Gentner: Structure-Mapping in Analogy and Similarity
• Edwin Hutchins: Distributed Cognition
• Collins, Brown, and Newman: Cognitive Apprenticeship
• Haynes et al.: A Surgical Safety Checklist to Reduce Morbidity and Mortality
• NIST: AI Risk Management Framework
• Microsoft Research: The Impact of Generative AI on Critical Thinking
• METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
• Anthropic: Mapping the Mind of a Large Language Model