What LLMs Can't Do (Yet)
A practical inventory of the places where language models still fail in ways that matter — and what that tells us about intelligence.
I use language models every day. I find them genuinely useful — not in the hype-inflected way where everything is revolutionary, but in the mundane way where they save me real time on real tasks. I also spend a lot of time noticing where they fall apart.
The failure modes are more interesting than the successes. They tell you something about what these systems actually are.
The Calibration Problem
The most practically important failure is bad calibration. LLMs don’t know what they don’t know. They produce confident, fluent text whether they’re describing well-established facts or hallucinating something plausible-sounding.
This isn’t a retrieval problem — it’s a structural one. The model has no mechanism for distinguishing “I learned this from ten thousand sources” from “I inferred this from statistical patterns in adjacent text.” Both come out sounding the same.
The practical consequence: you can’t trust an LLM on anything where the stakes of being wrong are high, unless you can verify the output independently. Which often means you’d have been better off doing the research yourself.
The Context Collapse Problem
LLMs are very good at local coherence — the next sentence makes sense given the previous one — and surprisingly bad at global coherence — the whole document hangs together as a unified argument.
Ask a model to write a long essay and you’ll notice it tends to circle back to the same points, lose track of its own thread, and occasionally contradict something it said six paragraphs ago. It’s optimising for sentence-by-sentence plausibility, not document-level integrity.
Humans with good writing skills do something different: we hold the whole structure in working memory, feel the tension between sections, notice when an argument has a hole in it. Models aren’t doing that.
The “Why” Problem
Perhaps most fundamentally: language models can describe processes but don’t understand mechanisms. They can tell you that certain antibiotics cause C. diff infections. They cannot reason from first principles about why — what the causal chain is, what would change if you altered a variable.
This is the difference between pattern-matching on training data and having a model of the world. LLMs have the former. They don’t have the latter in any robust sense.
What This Means
None of this is a criticism. A tool that’s useful 80% of the time is valuable. The skill is knowing which 20% to be careful about.
The mistake is treating LLMs as oracles — as if their confident output is a signal about truth. It isn’t. It’s a signal about what text patterns appear plausible given the training distribution.
Used well, they’re a first draft machine, a brainstorming partner, a search engine that can synthesise. Used carelessly, they’re a hallucination engine with very good PR. The difference is entirely in how you use them.