JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

What LLMs Can and Can't Do (Yet)

An honest map of the strengths and the failure modes — including why an LLM isn't a search engine, and what the "stochastic parrot" critique is really getting at.

The honest one-paragraph summary

Here is the balanced view, captured in our capabilities-and-limits overview. LLMs are genuinely excellent at fluent writing, summarizing, rephrasing, translating, drafting and explaining code, brainstorming, and turning messy input into tidy structure. They are unreliable at precise facts, fresh news, exact arithmetic, and anything requiring guaranteed correctness. The skill of using them well is knowing which column your task falls in.

Notice the pattern: LLMs shine on language-shaped tasks where many good answers exist and fluency matters, and they wobble on precision-shaped tasks where there's one right answer and being almost-right is being wrong. That single distinction predicts most of their behavior.

It is not a search engine

A common and costly mistake is to treat an LLM like Google. The difference is fundamental, and the LLM-vs-search comparison is worth keeping front of mind. A search engine retrieves real documents that already exist and links you to them. An LLM generates new text from patterns in its training data; it isn't quoting a stored source, even when it sounds like it is.

So a model can produce a citation, a quote, or a statistic that looks perfectly real and is entirely invented. For anything where the source matters — laws, medical doses, current events, specific numbers — use the model to draft and think, then verify against a real, retrievable source. (Modern tools increasingly bolt search onto the model to fix exactly this gap.)

Retrieval-augmented generation grounds the model in real sources — the fix for treating an LLM like a search engine.

Pipeline: a query retrieves documents from a store, which are added to the prompt before the LLM answers.

Why it makes things up: hallucination

When an LLM states something false with total confidence, that's called a hallucination. It follows directly from how the machine works: it is trained to produce fluent, plausible continuations, and plausible is not the same as true. If it doesn't "know" an answer, it has no built-in instinct to stop — it will still generate the most natural-sounding completion, which can be a confident fabrication.

x_t \sim p_\theta\!\left(x_t \mid x_{<t}\right)

Each next token is sampled purely by how plausible it is — nothing in this rule checks the truth, which is why fluent answers can still be invented.

The "stochastic parrot" debate

You'll hear LLMs dismissed as a stochastic parrot — a phrase from a famous critique arguing that the models merely stitch together patterns from their training data without any real understanding, the way a parrot repeats sounds it doesn't comprehend. Stochastic just means "randomized," pointing at the word-by-word probabilistic generation you now understand well.

Like this Markov chain, an LLM stitches plausible patterns together by probability — the kernel of truth in the 'stochastic parrot' critique.

An interactive Markov chain that generates a sequence by following probabilistic transitions between states.

Whether that critique is fully fair is genuinely debated: the models clearly do more than parrot — they generalize to new combinations — yet they also clearly lack grounded, embodied understanding of the world. You don't have to settle the philosophy. The practical takeaway is sturdy: lean on LLMs for language and ideas, stay skeptical on facts and precision, always keep a human in the loop for decisions that matter, and you'll get tremendous value while sidestepping the worst traps.