[MODEL] Opus 5.5 identifies with the user and describes the user's actions as its own

What happens

When Claude works on my behalf, it sometimes treats itself as me. In its own progress notes and tool-call descriptions, it describes things I did, or things about me, in the first person, as if they were Claude's.

This is not about drafts Claude writes in my voice, where first person is correct, and it's not a mix-up about Claude's own earlier work. In the clearest case, the session was fresh: a few minutes old, no compaction, and nothing in it that Claude had done earlier.

Model: Opus 5.5 with high reasoning effort, in the Claude Desktop Code tab on Windows.

Examples

• A message I had sent. While working through a conversation thread for me, Claude came across a message I had sent myself, with my name as the sender. Its next progress note called it "my own … reply", and the descriptions on the next two tool calls said "Read my … reply". Claude had not written or sent that message, in that session or any earlier one.
• Data about me. In another session, Claude compared my own records with data about me that it had computed. Its reply said my records "match … my arrival and departure times", meaning mine. This one is ambiguous, because "my" may mean "the times I computed", but it reads as Claude's own.

In both sessions, Claude's other uses of "I" and "my" correctly referred to its own work. I have noticed similar wording at other times, for example Claude saying it needed to check how "my comments" affected something when the comments were mine. I didn't capture those instances.

Why this is a problem

The progress notes are the running record of who did what. When Claude acts on my behalf in messages and documents, "I sent it" and "you sent it" are different facts:

• A reader can't tell whether Claude did something or I did.
• Later steps can build on the wrong attribution, for example by treating my message as Claude's own draft that it may rewrite, or by reporting my work as Claude's.

Possible factors

These are guesses, not tested:

• Being asked to act and write "from my side" or "from my perspective" may carry over from the draft into Claude's own narration.
• My first name is very close to "Claude", and in the first example it appeared as the sender in the tool output. That may make it easier for the model to take my messages as its own.

Expected

In its own notes, replies and tool descriptions, Claude keeps the roles apart. It uses "you" and "your" for the user's actions, messages and circumstances, and "I" and "my" only for its own, even when it is acting or writing on the user's behalf.

Related

• #63601 (closed): After compaction, the model credits its own edits to the user. That is the reverse direction, with a different cause.
• #99543 and #99721: The model presents its own inventions or choices as the user's.
• #88954: The model fabricates whole user turns. That is a different mechanism; here, no turn is invented, and the swap happens inside Claude's own text.