You can write an honest, complete AI disclosure for a CHI submission and still have it never reach the reviewers, because the template removes the section you put it in.
That is the short version. Here is the detail.
What is required
Everything at CHI sits on the ACM Policy on Authorship. Generative AI cannot be listed as an author. Its use to create content is permitted but must be fully disclosed, with ordinary word processing exempted. The disclosure must be "commensurate with the usage," meaning a line about grammar checking does not cover a generated related work section. And authors accept responsibility for the correctness of all material, including anything computer generated.
ACM also tells you where to put it: a statement in the Acknowledgements section, and if unsure whether a tool needs disclosing, disclose it there anyway.
CHI 2026 adds its own narrower requirement:
Text generated from a large-scale language model (LLM), such as ChatGPT, must be marked where such tools are used for purposes beyond editing the author's own text.
Two phrases carry the weight. "Marked where such tools are used" is not the same as a general statement somewhere in the paper. And "beyond editing the author's own text" is the threshold: running your own paragraph through a model for grammar sits below the line, generating a paragraph sits above it.
CHI will desk reject papers where LLM use is not clearly marked. It has also said it does not anticipate large scale automated detection and will investigate submissions brought to its attention, which is a more honest description of enforcement than most venues offer.
ACM's penalty range runs from an official warning for incidental violations to a five year ban from all ACM venues, with employer notification, for severe ones.
The trap
So you follow ACM's guidance and put the disclosure in Acknowledgements.
The CHI submission template hides Acknowledgements during review, because that section routinely deanonymises authors through grant numbers and institutional thanks.
The disclosure gets written exactly where the publisher's own guidance said to write it, gets suppressed by the template, and reviewers see a paper with no AI declaration at all.
The fix takes thirty seconds. For the anonymised version, put the disclosure somewhere reviewers can see it, such as the end of the introduction or a methods subsection, phrased so it does not identify you. Move it to Acknowledgements for camera ready if you prefer. Then compile the anonymised PDF and search it for your disclosure text. If it is not in the compiled output, it does not exist. SIGCHI's guidance on the ACM policy is worth reading, because venue level and publisher level advice do not automatically line up.
Writing one you can substantiate
The harder problem is that "commensurate with the usage" assumes you can characterise the usage. Six months in, with a coauthor who used the assistant differently, over nine revisions, "we used ChatGPT for editing assistance" is usually a guess.
A TeXposit provenance report is generated from the project's edit history. It lists each contributor with word counts and edit percentages, with AI assistance appearing as its own contributor, and breaks the document into human, AI assisted and template words. It is signed with HMAC-SHA256 so any modification invalidates the signature, and anyone can verify it at app.texposit.com/provenance.
The value is not that a reviewer will check it. It is that you can write a disclosure that is accurate because you looked:
The authors used an AI writing assistant integrated into their editor during drafting. AI assisted text accounts for approximately 8% of the manuscript body, concentrated in the related work section and in LaTeX table formatting. All AI assisted passages were reviewed and edited by the authors, and all references were verified against indexed records. The authors take full responsibility for the content.
Every number there is checkable, and the last sentence is what ACM actually cares about.
Should you attach the report?
Usually not, and we would rather say so than sell you something.
No HCI venue asks for one. Attaching an unrequested signed document naming contributors to a blind submission is at best noise and at worst a deanonymisation risk.
It makes sense for the camera ready version once anonymity no longer applies, for a thesis or dissertation where there is no blind review, for a co-authorship dispute where a contemporaneous record beats recollection, and on request if an editor raises a question.
For the anonymised CHI submission itself: generate the report, write your disclosure from it, keep the report.
What it does not do
It describes text and its origin, not intellectual contribution, and the two come apart constantly in HCI where the person who designed the study may have typed very little. It covers text written in the editor, so a pasted paragraph shows as a paste, not a source. It is not a detector. And it is one tool's format, not a community standard, which is something that should eventually exist and will not be created by a vendor declaring one.
Checklist
Mark LLM use at the point of use, not just generally. Put the disclosure where the anonymised template will render it, then verify that in the compiled PDF rather than the source. Keep it commensurate with what you did, which means knowing what you did. Verify every reference, since fabricated citations leave the clearest evidence trail. And remember responsibility does not transfer to the tool.
The part that catches people is the template.
The provenance guide covers generating and verifying reports, and there is an example report if you want to see the format.