The Two Gaps Behind Almost Every Bad AI Output

An LLM rarely disappoints because it cannot think. It disappoints because of what we forgot to place in front of it. The model is only ever as good as the world we hand it, and when the output arrives malformed or hollow or beside the point, the failure is almost always upstream of the machine rather than inside it. There are broadly two gaps that open up here, and they are worth telling apart, because the remedy for one does nothing for the other.

Context starvation

This is the gap of substance. It is when the model has been asked to reason about a world it cannot see. The background, the constraints, the half a dozen facts that live only in your head and never made it into the prompt. The model is not lost because it is incapable; it is lost the way anyone is lost when asked for directions to a place that was never named. Ask it to tidy up last quarter’s numbers and it hands back a clean, confident summary of the wrong quarter, because you never said which file you meant. Ask it about a rule and it cites the version that was true a year ago. It fills the void with the most plausible thing it can find, and plausible is not the same as true. A starved model does not refuse and it does not stall. It confabulates, confidently and fluently, and that fluency is precisely what makes the gap so easy to miss.

The structural gap

This is the gap of form. Here the model has the substance, every fact it needs is already present, and still the output comes out wrong, because it was never told the shape the answer was meant to take. Hand a capable cook every ingredient and no word on whether you wanted a soup or a soufflé, and you will get something edible and nothing like what you ordered. Ask for a short reply and you get three measured paragraphs. Ask for an apology and you get an explanation, because short and sorry were shapes you pictured but never put into words. The structural gap is the distance between the material and the blueprint. It shows up as the right idea in the wrong register, the answer that is technically complete and tonally foreign, the essay that reads like a memo because no one ever said it was meant to be an essay.

And here is the trap the second gap sets. Feed it more of what it already has, more detail, more background, even the exact name of the thing you asked for, and it does not edge any closer to what you wanted. It only does the wrong thing more rigorously, because what was missing was never information at all.

How do we overcome these?

The first instinct is to blame the model and reach for a bigger one. But a sharper model starved of context only confabulates more convincingly, and a sharper model left without structure only formats its confusion more elegantly. The fix, almost always, lies on our side of the conversation.

For the first gap, we feed it the world it is missing. We do not assume it knows the project, the constraints, the reader; we assume it knows nothing of these and we say them out loud. Context is something which, when supplied in the right measure, quietly turns guesswork back into reasoning.

For the second, we show it the shape. An example of what we want is worth more than a paragraph describing what we want. A template, a register, a single line of make it sound like this does the work that abstraction never can. We close the structural gap not by asking for more, but by asking for it in a form the model can recognise.

Both gaps, in the end, return to the same quiet truth. The model was never the missing piece. We were.