Skip to content
IAMB.AI
Documentation

How AI works in IAMB.AI

What the AI does and does not do, how every AI step is recorded, how the paper discloses it, and how your data is kept safe.

Last verified 2 min read

On this page
  1. What the AI can do
  2. What it does not do
  3. Every AI step is on the record
  4. The AI-use statement
  5. Choosing models
  6. Safety and privacy

What the AI can do

  • Screen records as an advisor or as an independent reviewer.
  • Discuss its votes in conflict threads, and change its mind with reasons.
  • Find open-access PDFs on the web (never shadow libraries).
  • Propose an extraction form from your research questions.
  • Extract data, always with a verbatim quote from the study.
  • Rate prior surveys for the related-work table.
  • Draft and review manuscript sections from your review's own data and your notes.
  • Look at every page of the built PDF for problems rules cannot see.

What it does not do

  • It does not decide alone unless you choose a mode that says so (auto-exclude, the autopilot, AI extraction), and the paper then reports exactly that.
  • It does not write from memory: sections are drafted from what your review recorded.
  • It is never listed as an author. AI co-authors are team members in the workflow only.

Every AI step is on the record

Each AI judgement is stored with its model, its prompt, its confidence and its reasons. Each job records the model and effort it ran with, the AI calls it made and what they used (see What the AI work used). The paper names the models its jobs actually used, so changing the site's default later does not change what an earlier review reports.

The AI-use statement

IEEE, ACM and APA all require disclosure of generative-AI use. The paper's AI-use statement is generated from the job log: which models did which tasks, and how people oversaw them. Only work a model actually did counts: open-access lookups and the PDF build are not AI. See The AI-use statement.

Choosing models

Each review, task and AI reviewer can use its own model; otherwise they inherit the site's default. Two AI reviewers on two different models read each record independently, which makes their agreement meaningful. A review whose protocol names a model should set it in its AI tab, so later work does not follow a new default.

Safety and privacy

  • The AI runs with no tools for structured tasks: a prompt cannot make it run commands, read files or browse, whatever a paper or a person wrote.
  • Text from papers and from people is always marked as material to read, never as instructions.
  • Your review's records and files stay on the site.

The interface always says "AI". The models a review used are named where it matters: in the AI tab and in the paper.

Tags ai ai-use statement models privacy safety transparency

Was this helpful?