Which tool to choose
The tool is once again the instructions: the personal ones if the preference always applies ("always favor authoritative sources"), or the start of the single chat if it applies only to that job. For sources, indicate the type you trust (official sites, documents you provide) and the type to avoid. For the AI's tools — web search, file reading, calculations — say when to use them ("search online before answering about recent facts"). Automatic memory isn't enough for these rules of method: they should be written, because they guide how the AI arrives at the answer.
How to do it
- Decide your method preferences: which sources to trust, what to avoid, which tools to use when.
- Write them as rules in the instructions: "always cite the sources", "for current events search online", "base yourself only on the documents I give you".
- Make a request and check that the AI follows the rules: does it cite the sources? does it use the right tool?
- Adjust the rules that aren't respected, making them more precise.
A concrete example
Luca uses the AI to inform himself on health and regulatory topics, where sources matter enormously. He put in the instructions: when you search, favor official and institutional sources, always cite the link, and if you only find weak sources tell me instead of answering anyway. Since then the answers arrive with the references, and when reliable material is scarce the AI flags it instead of filling the gaps. Luca no longer has to ask each time "but where did you get that?": he established it once as a working rule.
When it DOESN'T work (and how to fix it)
If the AI doesn't respect the preference on sources
The rule is probably too vague. "Use good sources" leaves too much room; "cite only from official sites and give me the link" is verifiable. Make the rule checkable, so you immediately notice whether it's being followed.
If it asks for a tool it doesn't have
Not all assistants search online or read files. If you ask it to use a tool it doesn't possess, it'll tell you or fall back on its knowledge. Check which tools your assistant has and adapt the rules to what it can really do.
If the cited sources are weak anyway
Even with the right rule, the search can pick up poor pages. Ask it to look for better ones, or indicate the sources to use yourself by pasting or naming them. For delicate topics, you choose the source and give it to the AI.
A tip from someone who really uses it
Treat sources as the most important part of the answer, not as a final detail. A confident answer with no verifiable sources is an answer to take with a grain of salt: you don't know if it's solid or invented. Impose on the AI, as a fixed rule, that it always tell you where it takes the information from, and admit when it has no reliable basis. This changes your relationship with the tool: you stop receiving assertions to believe on faith and start receiving answers you can check. On what matters, the difference between trusting and verifying runs precisely through the sources.
Frequently asked questions
Can I force the AI to use only certain sources?
You can steer it strongly ("use only official sources", "base yourself only on the documents I give you"), and to be sure you provide the sources yourself. A written rule strongly shapes behavior, but total control you have when you're the one giving it the material.
How do I push it to always cite sources?
By making it a fixed rule in the instructions and checking that it respects it. If it cites, great; if it doesn't, call it back. On facts and data, demanding the source should become your habit.
Does this apply to tools like web search too?
Yes: you can tell it when to use online search, when to rely only on your files, when to show you the steps of a calculation. As long as they're tools your assistant actually has.
If I say nothing, does the AI choose good sources anyway?
No, and it's the reason why so many answers seem reliable and aren't. Without a direction from you, the AI draws on what it finds or what it "knows", without distinguishing as you would between an authoritative source and any old one, and often without telling you where it comes from. The result is plausible but not checkable. Establishing yourself, in plain terms, which sources to favor and the obligation to cite them is what turns an answer to believe into an answer to verify. Letting the AI choose without rules is the most common way to trust a source you'd never have accepted.