You share a project with two colleagues. One must add reference files and invite a third person; the other should only read and use what is there. How should access be set?

The principle — a permission model is applied by matching each person's actual task to the smallest level that covers it, and the way to reason about it is to read what each level grants rather than to rank them by feel.

Why the key is correct — OpenAI states that invitees get two levels of access, edit or chat. Edit access allows members to update instructions, upload or remove files, and invite others, though not remove existing members. Chat access lets members see and interact with the project's chats, files and instructions, but not invite others. The first colleague must upload files and invite someone, which is the edit level exactly; the second only reads and uses what is there, which is the chat level exactly.

Why the others are wrong — giving both the higher level assumes chat access cannot reach the files, and it can; it simply cannot change them or invite. Keeping both low and doing the uploads yourself works but makes you a bottleneck for a job the model already covers. And owner is not required to invite, since inviting is included at the edit level; owner additionally removes members and deletes the project, which is more than anyone here needs.

Remember this — two questions settle almost every sharing decision: does this person need to change anything, and do they need to bring anyone else in. Both are yes at the edit level and both are no at the chat level. Note also that permissions granted to a group and to an individual are not simply additive — where they differ, the higher of the two applies, so a group invited with edit access makes every joining member an editor.

Sources — OpenAI's Projects in ChatGPT article.