LLMs are impressive, but they have well-documented weaknesses. Knowing them makes you a safer and more effective user.
Hallucination
An LLM can write text that is fluent and confident but factually wrong, or unfaithful to a source you gave it. It happens because the model is built to produce plausible text, not to verify facts.
What helps: ask for sources and check them, supply the source text yourself, use retrieval (RAG), and verify numbers, dates and quotes.
Bias
Models learn from human-written text, so they can inherit its stereotypes, for example linking certain jobs with a particular gender. Another effect: because English dominates training data, models can favor English-language perspectives, even when you ask in another language.
What helps: state the perspective you want, ask for multiple viewpoints, and review sensitive outputs yourself.
Flattery (sycophancy)
Models tend to tell you what you seem to want to hear. A chatbot may agree with a wrong claim, drop a correct answer when you push back, or praise too much.
What helps: ask it to critique your idea, argue the other side, or list what could go wrong. Don’t treat agreement as proof.
Prompt injection
An LLM has trouble telling your instructions apart from instructions hidden inside content it reads, such as a web page or an uploaded file. Someone can plant text that tries to redirect the model.
What helps: treat outside content as data, limit what tools an AI can use, and require approval for risky actions.
Memorization and privacy
Models sometimes reproduce long passages from their training data word for word. For you, that’s one more reason not to paste confidential information into tools you don’t trust.
Other things to keep in mind
- Energy: running models uses electricity. Image generation takes far more per request than short text answers.
- Thinking skills: researchers worry that heavy reliance on LLMs could weaken critical thinking. Use them to support your thinking, not replace it.
- Emotional support: many people talk to chatbots about personal struggles. They can be a place to organize thoughts, but they can also hallucinate or agree with unhelpful ideas, and some lack proper crisis handling. If you’re struggling, please reach out to a qualified professional or someone you trust.
A simple habit
For anything important, ask yourself: “How would I check this?” If you can’t answer, don’t rely on it yet.
Further reading: the Wikipedia article on large language models.