AI hallucinates when it gets overwhelmed. Ask it one, focused question at a time.
Don’t ask broad questions, hoping it’ll magically hone in on the most important issue. Sometimes it will, but you have to check, and it may also miss things.
For systematic investigations, have a framework for sequencing your questions.
Avoid sourcing your inputs from the AI itself. For example, when I investigate stocks, I don’t ask the AI for which stocks to investigate. I use a list from finviz. The more external data you inject into the AI prompt, the less surface area for hallucination there is.
The impact of hallucinations accumulates if your sequence of questions depends on the accuracy of previous answers, so double check.
AI answers are always in authoritative language (using my system prompt), but treat them as as speculative hypotheses and argue with them. You can change the AI’s mind within that chat context through a logical rebuttal.
AI doesn’t just hallucinate empirical facts. It also conjectures broad, mechanistic frameworks for understanding topics under discussion that aren’t always appropriate.
What AI’s great at is surfacing hypotheses and frameworks over the empirical data you provide to it for you to consider. This is one of its best use cases for research. It’s often wrong and you need to use your judgment, but it lets you move much more quickly and with more confidence into new domains by supplying you with the professional common sense specialists in that area would apply to your questions.
Gemini 3.1 Pro technique diary (5/1/2026):
AI hallucinates when it gets overwhelmed. Ask it one, focused question at a time.
Don’t ask broad questions, hoping it’ll magically hone in on the most important issue. Sometimes it will, but you have to check, and it may also miss things.
For systematic investigations, have a framework for sequencing your questions.
Avoid sourcing your inputs from the AI itself. For example, when I investigate stocks, I don’t ask the AI for which stocks to investigate. I use a list from finviz. The more external data you inject into the AI prompt, the less surface area for hallucination there is.
The impact of hallucinations accumulates if your sequence of questions depends on the accuracy of previous answers, so double check.
AI answers are always in authoritative language (using my system prompt), but treat them as as speculative hypotheses and argue with them. You can change the AI’s mind within that chat context through a logical rebuttal.
AI doesn’t just hallucinate empirical facts. It also conjectures broad, mechanistic frameworks for understanding topics under discussion that aren’t always appropriate.
What AI’s great at is surfacing hypotheses and frameworks over the empirical data you provide to it for you to consider. This is one of its best use cases for research. It’s often wrong and you need to use your judgment, but it lets you move much more quickly and with more confidence into new domains by supplying you with the professional common sense specialists in that area would apply to your questions.