Which tool to choose
The tool is the assistant's personal instructions section, the one that applies to all chats: that's where values make sense, because they have to steer every answer, not just one. Write them as operational criteria, not as abstract declarations. Automatic memory may catch a value in passing, but it's too vague for something this important: the principles you want the AI to reason with should be written by you, where you control them, not left for the machine to infer.
How to do it
- Open the personal instructions and add a section on your values.
- Write them so they guide choices: not "I care about the environment" but "all else being equal, always prefer the more sustainable option".
- Add the boundaries: what you never want suggested to you.
- Put it to the test with a request where values matter (a purchase, a decision) and check that the advice respects them.
A concrete example
Chiara cares a lot about conscious consumption and about not getting into debt. She wrote in the instructions: all else being equal, always prefer reuse and second-hand over new, and never propose installment or credit options. Since then, when she asks for advice on a purchase, the AI suggests the second-hand market first and never brings up financing. Without those lines she received "catalog" advice, the same for everyone; with those lines she receives advice that starts from how she wants to live.
When it DOESN'T work (and how to fix it)
If the AI ignores a value in a concrete case
Often it's written too abstractly. "Be ethical" guides no choice; "never suggest hiding information from a client" does. Turn the principle into a rule that can be applied to a real situation.
If two of your values conflict
It happens, and it's normal: economy and sustainability, for example, sometimes pull in opposite directions. Tell the AI which one comes first when they clash, so it knows how to weigh them. Without an order of priority, it will decide at random.
If you notice a principle has changed
Values evolve. When one changes, go back to the instructions and update it: the AI will keep applying the old one as long as that's what's written. You recalibrate the compass yourself.
A tip from someone who really uses it
Write values as instructions for acting, not as a portrait of yourself. It's nice to describe yourself as "an honest person, attentive to others", but it doesn't tell the AI what to do when it advises you. Far more useful: "when you propose a solution, always discard those that harm someone, even if more convenient". A value, for the assistant, counts only if you translate it into a criterion it can follow when facing a concrete choice. Abstract principles stay nice on paper; operational rules change the answers.
Frequently asked questions
What's the difference between values and preferences?
Preferences concern tastes (what you like); values concern criteria (what's right for you). The former tune the practical proposals, the latter steer the decisions. It's worth keeping them in two distinct parts of the instructions.
Will the AI always respect my values?
It will keep them in mind if they're clearly written, but it remains a tool: when facing important choices, the final judgment is yours. Use its advice as support, not as a moral verdict.
Do I have to share personal beliefs with the AI?
Only to the extent it serves you to get aligned advice, and with the caution you'd use for any personal information. The instructions are yours: you decide how much detail to go into.
If I use the AI a lot, will it understand on its own what I believe in?
No, and it's the illusion that leads to advice that's always a bit off. The assistant doesn't deduce your principles by talking to you: at best the automatic memory grabs a fragment of them, randomly and incompletely. Until you explicitly write the criteria that should guide it, the AI reasons on the average of everyone, not on you. Values aren't transmitted by osmosis: they're declared. Whoever waits for the AI to "know them well enough" keeps receiving generic advice and is surprised it never quite reflects them.