The wrong mental model is everywhere
Most people interact with AI as though it were a very fast encyclopedia. You ask a question; it finds the answer. The bigger the model, the more answers it has stored. When it gets something wrong, the intuition is that the page was missing - the model just didn't have that one.
This mental model is wrong in a way that matters. It is not just imprecise - it produces systematically incorrect intuitions about when AI will fail, what makes a good question, and why rephrasing something slightly can produce a completely different answer.
The better model is search. Not searching a website or a document archive, but something more like exploring a landscape - navigating terrain to find the most plausible response to what you asked. Understanding AI this way clarifies most of what confuses people about how it behaves.
Not retrieval, but navigation
An encyclopedia stores facts. You look up a topic; the entry is there or it isn't. An AI language model works differently. During training, it processes enormous amounts of text and builds up a kind of internal landscape shaped by patterns in language - what tends to follow what, how ideas relate, what kinds of responses fit what kinds of questions.
When you ask it something, the model doesn't look up the answer. It navigates that landscape, moving through it step by step, producing each word based on where it is and where the terrain points next. The output is the destination it reached.
This is why the same question can produce slightly different answers each time you ask. The landscape has multiple plausible destinations for most questions. The model isn't fetching a stored answer - it's finding one of many reasonable endpoints.
Why this explains the failures
The encyclopedia model predicts that AI fails by not having information - a gap in the record. The search model predicts something different: that AI fails by navigating to the wrong place.
Hallucination is the clearest example. When a model confidently states something false, it hasn't failed to find the right page. It has navigated to a region of the landscape that feels plausible - it fits the patterns of how things are typically written about this topic - but doesn't correspond to reality. The navigation succeeded by its own logic. It just arrived somewhere wrong.
Why small wording changes matter so much follows from the same idea. Your question determines where the navigation starts. A slightly different starting point can lead to a completely different destination - not because the model is easily confused, but because the terrain genuinely looks different from a slightly different position. Rephrasing isn't just cosmetic; it changes the journey.
Confident wrong answers happen because confidence reflects how smooth and consistent the navigation was, not whether the destination was correct. A model can take a clean, consistent path to a false conclusion. The absence of hesitation doesn't mean the answer is right.
Why context shapes answers so strongly is the same mechanism again. Everything you put in front of the model - examples, instructions, prior conversation - shifts the starting position. The model isn't being swayed or tricked. It is navigating from a genuinely different place.
Prompting as giving directions
If the model is navigating a landscape, then writing a good prompt is less like asking a clear question and more like giving good directions.
Vague directions get you somewhere in the general area. Specific directions get you to the right place. When people say you need to be "specific" with AI, this is the underlying reason: each piece of specificity narrows the landscape the model is navigating toward a region where useful answers live.
This is why prompts that seem over-specified often work better. Saying "answer this as a financial advisor explaining to someone who has never invested before, in plain language, in no more than three paragraphs" isn't unnecessary verbosity. Each instruction rules out more of the landscape and steers toward what you actually want. Without those constraints, the model finds somewhere plausible - but plausible is a wide territory.
Vague questions get average answers because average answers are what the most open, unconstrained region of the landscape looks like.
Thinking out loud as navigation
One of the more effective techniques for getting better answers from AI is asking it to "think step by step" - to show its reasoning before giving a final answer. The search model explains why this works.
Each reasoning step the model produces changes where the next step starts from. A well-formed intermediate step puts the model in a part of the landscape where the next step is more likely to be well-formed too. The chain of reasoning is the navigation path becoming more constrained as it goes - each step making the right destination more reachable.
This also explains a failure mode: if an early step is wrong, the rest of the reasoning builds on a false foundation. The model doesn't backtrack. It keeps navigating from where it is. One confident early error can produce a chain of internally consistent but completely wrong conclusions.
Improving AI means reshaping the landscape
The encyclopedia model suggests a clear path to better AI: add more information, fill the gaps, make the record more complete. The search model suggests a harder problem: improvement means making the landscape itself better calibrated, so that the most plausible destinations are also the most correct ones.
This is difficult because the landscape was shaped by learning from human-written text - and human text includes confident errors, motivated reasoning, and contradictions. Simply training on more text doesn't automatically produce a more accurate landscape; it can reinforce existing distortions. The techniques used to improve AI models after initial training - reinforcement learning from human feedback, fine-tuning, constitutional methods - are all ways of reshaping the landscape, steering the most plausible destinations toward responses that are accurate and useful rather than just fluent and confident.
The implications for trust
Treating AI as an encyclopedia produces a simple trust test: ask it questions you know the answers to and see if it gets them right. If it does, you trust it; if it doesn't, you don't.
Treating AI as navigation produces a more nuanced question: for the kinds of things I'm asking, does this model reliably navigate to the right place? How sensitive is it to how I phrase things? When it goes wrong, does it go wrong in recoverable ways or in ways I might not notice?
These questions don't have clean yes or no answers. They depend on what you're asking, how you're asking it, and what the consequences of a wrong destination look like for you.
Adopting this mental model doesn't simplify the problem of trusting AI systems. It actually surfaces more complexity. But that complexity was always there - the encyclopedia model was just asking questions that hid it.