I think it's fair to say the following rule is useful for guiding how we use it: when it comes to known knowledge, skill sets, and tool sets, AI is extremely good—it can even correct us if we're saying nonsense.
But for unknown knowledge—the kinds of things no one really has answers to—it acts more like transparent glass. It just polishes whatever it's been given, often by being agreeable. So if a human feeds it nonsense, it produces polished nonsense. But if someone provides something genuinely insightful, it returns polished insight. This is the usage pattern I have observed in 2025 (I will update this as time and the technology evolves). So far, it seems artificial intelligence tools don't replace theoretical physicists, but function as springboards for them to efficiently jump to higher understanding.
Here are three guiding rules I think are useful:
Use language models only for the following three purposes:
1) Updating knowledge and skills – For example, conducting literature reviews of papers you genuinely need, or refreshing concepts you’ve already learned but may need to revisit. But you must be able to cross-check every step of its reasoning.
2) Encyclopedic and manual labor help while creating – Use it for routine calculations after you've already learned how to do them and can do them, or to get unstuck on specific calculations or concepts you've struggled with for a reasonable amount of time. Again, you must be able to cross-check its reasoning (there should be no black boxes in your understanding).
3) Editing language and flow – Use it for improving clarity and tone, but only after you’ve completed writing the entire document to the best of your ability.
AI should not be used for brainstorming or as a collaborator; AI tools are, by design, meant to inflate the user's ego and not to be a true critic of the user's work. It should not be used for moral advice either—that'd be like using crutches instead of relying on your own moral axis, which is often closely linked to one's identity.
But for unknown knowledge—the kinds of things no one really has answers to—it acts more like transparent glass. It just polishes whatever it's been given, often by being agreeable. So if a human feeds it nonsense, it produces polished nonsense. But if someone provides something genuinely insightful, it returns polished insight. This is the usage pattern I have observed in 2025 (I will update this as time and the technology evolves). So far, it seems artificial intelligence tools don't replace theoretical physicists, but function as springboards for them to efficiently jump to higher understanding.
Here are three guiding rules I think are useful:
Use language models only for the following three purposes:
1) Updating knowledge and skills – For example, conducting literature reviews of papers you genuinely need, or refreshing concepts you’ve already learned but may need to revisit. But you must be able to cross-check every step of its reasoning.
2) Encyclopedic and manual labor help while creating – Use it for routine calculations after you've already learned how to do them and can do them, or to get unstuck on specific calculations or concepts you've struggled with for a reasonable amount of time. Again, you must be able to cross-check its reasoning (there should be no black boxes in your understanding).
3) Editing language and flow – Use it for improving clarity and tone, but only after you’ve completed writing the entire document to the best of your ability.
AI should not be used for brainstorming or as a collaborator; AI tools are, by design, meant to inflate the user's ego and not to be a true critic of the user's work. It should not be used for moral advice either—that'd be like using crutches instead of relying on your own moral axis, which is often closely linked to one's identity.