I would like to promote the use of the following term: AI-safe task to mean the following:

A task which can be given to an AI, and for which it is universally trivial for an expert human to verify whether the task has been completed or not based on the changes made by the AI.

I’m not familiar with any other term used to mean the same thing, but if there is one - please let me know! I know that I’m just giving a name to something, but I think in some cases that can be very useful.

Examples of AI-safe tasks would be:

• Changing the color of a page in html.
• Improving the grammar or spelling in a sentence.
• Shortening (golfing) a mathematical proof in Lean.

Examples of non-AI-safe tasks would be:

• Coming up with a physics theory.
• Refactoring a whole code base.
• Formalizing a full academic paper in Lean.

This is not to say that one should not use AI for non-AI-safe tasks, just that one should be mindful that one either has to check the output carefully, or have a high degree of trust in the AI system.

(AI was used to do AI-safe task of fixing grammer and spelling of this post.)