AI audit and disclosure

The record

Whichever engine does the thinking, the project keeps the trail:

• Conversations. Project → Your AI lists every conversation, including those a connected MCP client created — marked MCP with the client's name — with its tool calls, their results and the plan it kept. A conversation closes after 30 minutes without a call and stays readable.
• Activity. Every action an agent took is on the project's Activity tab under your account, next to the same actions taken by hand.
• Authorship. The manuscript's Authors panel colours the text by who wrote it and keeps that attribution after a tracked change is accepted, so the share written with AI help is visible in the document itself rather than only in a log.

The AI audit tab

• Platform AI. Every run of the platform's assistant (interactive, unattended, workflow) with status, spend and a link to its transcript, and model use by task class and model with calls, failures, tokens and spend. The record comes from the ledger that meters every call; nothing is inferred.
• Your own AI. Every judgement a connected client handed in through a submission contract — extracted evidence, screening votes, a risk-of-bias or GRADE proposal, claim dispositions, a critique, a readiness assessment, a reading note, a drafted section — with the client that submitted it, what it says it read (source versions, pages, works, the manuscript content hash), how much of the batch it covered, and, for a review, the mode it declared (self-review, external second opinion, isolated review, human expert review).
• A review mode is what the client said, not something Hikma measured: a second token, client, session or model proves attribution, not independence. The page says so, and never claims more.

Disclosure statement

Where platform AI is on, a statement is composed from that record — the tool, the models, what they were used for, the period and the number of calls — following the ICMJE recommendation on AI-assisted technology. Copy it into the acknowledgements or methods and adjust the wording to the journal's template.

Where you brought your own agent, the export names the client — Claude Code, Cursor, whatever your key is called — and how many proposals it made and how many you accepted, all from the record. The one thing the platform cannot see is which model your client ran, so the statement leaves that blank for you to complete; everything else is filled in. Where an administrator has switched the platform's own AI on, the models are named from the ledger and nothing is left blank.