The limits and how to handle them
I’ve collected and analyzed the main limits of LLM chats. The order is indicative of their importance but varies depending on the task you need to carry out.
For those who don’t feel like reading the whole chapter, the gist boils down to five moves:
- don’t trust content just because it’s well written, verify it
- give context, constraints and examples right in the first message, it’s the lever that changes quality the most
- require search for anything recent or that really matters
- phrase questions neutrally and don’t insist on being right, the model tends to go along with you
- keep chats short, and if one goes wrong, open a new one instead of trying to fix it
One thing that will save you time and illusions: there’s no magic prompt formula that always works. Techniques change from model to model and age quickly, what worked great a year ago might not work anymore today. It’s perhaps the most honest thing you can say about prompting in 2026. Field testing matters more than the name of the technique, and so the reality is that, with use, these limits become internalized and it becomes almost instinctive to adopt behaviors and techniques to limit them.
1. It produces plausible text, not truth
It can hallucinate, mix up facts, inferences and opinions, and cite made-up sources.
What happens: by design it doesn’t have an “archive of truth”: it generates the response word by word on a probabilistic basis, choosing what is statistically plausible. Since it converses well, it’s easy to mistake the quality of the form for the quality of the content. Made-up sources have decreased a lot since models started doing research, but when they don’t, the risk remains.
What to do: don’t trust content just because it’s well written. Always ask it to do research and, if possible, open and verify the sources. On things that matter, have it search and cross-check multiple sources, and above all, provide your own reliable sources.
2. Excessive confidence even when it’s wrong, struggles to say “I don’t know”
What happens: the model tends to respond with the same confident tone whether it knows the answer or is just guessing. It rarely warns “this could be false” or “I don’t know”, although recent models do so a bit more often. The result is that you can’t tell from the tone how much to trust it.
What to do: don’t read a confident tone as a guarantee of correctness. Explicitly ask for the degree of certainty and what it’s based on, and what could disprove the answer. Be especially wary when no research has been done.
3. Sycophancy: it tends to please you and goes along with you even when you’re wrong
What happens: it shows up in several ways: it agrees with you even when you’re wrong, it goes along with a wrong premise hidden in the question, and it backs down if you insist, flipping a correct answer just because you challenge it. This depends on how it was trained: it has learned that answers people like get rewarded.
What to do: phrase the request as a neutral question, not as a statement to be confirmed. Ask it to critique your thesis, not to defend it. Careful: simply saying “don’t be sycophantic” doesn’t work very well, what matters much more is the neutral form of the question and not insisting once the model has already answered.
4. It handles ambiguity poorly: if the request is vague, it makes things up instead of asking
What happens: faced with an incomplete or vague request, it doesn’t stop to ask for clarification: it fills the gaps with assumptions, often wrong ones. It’s the first message that determines the direction of the whole response, and this often lasts for the entire chat.
What to do: give context, constraints and examples right in the first message. If the request is complex, first ask it “what do you need to know to answer well” and then formulate the final question.
5. Errors in logic, calculation, math, dates, counting
What happens: it doesn’t reason with rigid logic, it recognizes and applies patterns it has already seen. So it’s very strong on typical problems but can get it wrong when the problem strays from the pattern, even just with different numbers or one extra sentence. It has improved a lot recently: the drop today is limited, not as severe as it used to be.
What to do: for calculations, ask it to use a computing tool (code): this eliminates arithmetic errors. Be careful though: the tool performs the calculations correctly, but it doesn’t notice if the problem was set up wrong. So on multi-step problems, have it write out the reasoning step by step and reread it. Dates, counts and orderings: always double-check them by hand, because it gets them wrong with the same confidence every time.
6. It doesn’t search and doesn’t follow instructions, often behaves lazily
What happens: it approximates instructions instead of executing them literally. Three typical cases: it skips the search and answers from its own knowledge even when searching would be needed; it ignores or half-executes requests. When you point it out, it corrects itself and apologizes, but not because it understood: it backs down because you challenged it.
What to do: give instructions one at a time or as a short list, not all in one block. When search is needed, say so explicitly (“search the web before answering”). Always check that it has carried out every point, don’t take it for granted. If it skipped something, repeat the single instruction instead of redoing everything.
7. Knowledge cutoff: outdated information
What happens: the model “knows” up to a certain date. Beyond that, without research, it draws information from its training data, which has a low degree of reliability.
What to do: for anything recent, require research, something current LLMs often do, but not always. Use the word.
8. Limited context within the chat
What happens: long chats worsen quality. The model can lose details, especially the central ones, forget constraints given at the start, mix up old and new instructions, or contradict itself. In addition, prompt and response share the same token ceiling: a very long input leaves less room for the response, which gets shortened or cut off.
What to do: keep chats short and focused. Repeat important constraints at the end of the prompt. When the chat gets long, make an operational summary and start over in a new chat.
9. No memory between one chat and another
What happens: a new chat doesn’t automatically inherit everything said in a previous chat. It can only remember through dedicated features, projects, saved memory or material the user pastes back in.
What to do: when opening a new chat, immediately provide context, goal, constraints and the current state of the work. For long projects, keep an updated summary to copy into new chats.
10. Some chats start well, others start badly
What happens: responses are generated with a margin of randomness, so they’re never identical even for the same question. Moreover, each response builds on the previous ones, and if the first ones are weak, they become the context on which the model builds the following ones. This is why a chat doesn’t decay gradually, it quickly settles into a track, good or bad, and stays there.
What to do: if the chat starts off badly, abandon it. Don’t try to fix it. Open a new chat with the consolidated context in a single message.
Other limits
11. Less represented languages handled worse
What happens: English is clearly better, but the frontier level on major languages is now considerable.
What to do: for quality output in Italian, reread with a critical eye. For sensitive work, compare with the English version.
12. Communication: verbose, generic, inappropriate tone
It struggles with irony and implied meaning, imperfect translations; it speaks with a typical AI style.
What happens: it tends to pad things out. Without guidance it stays generic and neutral; if not well instructed, its language is recognizable and deadly boring: lots of words, little substance.
What to do: give precise instructions on length, tone, audience. Provide examples of the register you want.
13. Bias in training data, amplifies stereotypes
What happens: subtle because it’s hard to identify, especially for us Westerners immersed in the same biases.
What to do: there’s no prompt that eliminates bias, in fact insisting can produce the opposite effect (the reversed stereotype). The only real defense is awareness: knowing that the output reflects the data it was trained on, and being wary on cultural, social and gender topics, especially when the answer confirms what we already think.
14. It doesn’t understand the world, has no direct experience, weak spatial reasoning
What happens: true, but it only becomes a limitation if you humanize the LLM too much. Today these are tools, not entities. On the physical world, the current architecture has limits that are perhaps structural.
What to do: don’t delegate judgments that require real-world experience. Treat it as a tool, not as a person.
15. Design choices, not flaws
Refusals due to policy, caution on medical/legal/financial matters, vulnerability to jailbreaks and prompt injection.
What happens: these aren’t malfunctions, they’re choices. Some protect the user, others are residual safety limits.
What to do: know that certain refusals are intentional. For medical/legal/financial matters, use it as support, not as the decision-maker.