Generative AI has changed academic writing permanently. By the most conservative estimates, a significant minority of papers submitted to major conferences in 2024 and 2025 were written with substantial LLM assistance. By 2026, that figure has grown — and conference organisers are navigating a policy landscape that did not exist three years ago.
The Scale of the Problem (and Why "Problem" Is Contested)
Major conferences including NeurIPS, ICML, and ACL have publicly noted reviewing irregular patterns in submissions consistent with LLM-generated text. These include unusual uniformity of style across submissions, formulaic structure, and responses to reviewer feedback that do not engage substantively with the technical critique.
However, the research community is genuinely divided on where to draw the line:
- Using LLMs to improve grammar and clarity in a paper where the ideas, experiments, and analysis are the author's own: widely considered acceptable
- Using LLMs to generate literature review summaries: contested — some say it introduces hallucinated citations; others see it as a more efficient form of an existing practice (copying from existing surveys)
- Using LLMs to generate experiment descriptions or results interpretations: widely considered unacceptable — this is the area that most directly falsifies the research record
- Submitting a paper that is substantially generated by an LLM with minimal intellectual contribution from the named authors: almost universally condemned as a form of academic misconduct
Current Policy Landscape
Conference policies have evolved rapidly. The current spectrum includes:
- Prohibition on LLM authorship, not LLM use: The most common policy position, modelled on NeurIPS 2023's pioneering statement. LLMs cannot be named as authors. Use of LLMs for writing assistance must be disclosed. This does not ban LLM-assisted writing; it bans undisclosed LLM authorship.
- Mandatory disclosure: Many venues now require a checkbox at submission indicating whether generative AI tools were used, and a brief description of how. This creates a disclosure norm without banning use.
- AI-use policy on review portals: EasyChair, OpenReview, and other submission systems have added AI use disclosure fields at both the submission and review stages — reviewers are increasingly asked to confirm they have not used LLMs to write their reviews.
Detection: What Works and What Does Not
AI detection tools (Turnitin AI, GPTZero, and others) are widely deployed in universities for student work — and widely acknowledged to have significant false positive rates, particularly for non-native English speakers whose writing patterns can resemble LLM output. Most conference organisers do not rely on automated detection for this reason.
What programme chairs actually rely on:
- Reviewer signal: Experienced reviewers in a field notice when a paper lacks the texture of genuine expertise — vague methods sections, results that do not parse on careful reading, references that exist but do not say what the paper claims they say
- Rebuttal quality: A paper written by an LLM with minimal author understanding typically produces a weak rebuttal that cannot engage with specific technical objections — this is a strong signal
- Cross-submission patterns: Papers that appear to share structural elements across a batch of submissions from related accounts are flagged for human review
The Reviewing Crisis Underneath the AI Problem
A less-discussed issue is AI use in reviewing rather than writing. With paper submission volumes at major conferences growing 20–30% annually, reviewers are under unprecedented pressure. There is strong evidence that LLMs are being used to generate first-draft reviews at significant scale. This is considered a serious breach of reviewer responsibility at most venues, because it violates the expectation of expert human judgment that the peer review system rests on.
Conferences including ICLR 2025 ran experiments comparing the distribution of review language to LLM-generated text and found statistically significant signals of AI-assisted reviewing at rates that concerned the programme chairs publicly.
Where Organizers Are Settling
Pragmatically, the 2026 consensus is converging around a few principles:
- Disclosure is mandatory, prohibition is limited to authorship fraud and undisclosed use
- Responsibility for accuracy rests with the human authors regardless of what tools they used
- Reviewers who delegate their reviews to LLMs risk removal from the programme committee and may be publicly named
- Journals and conferences are developing shared infrastructure for tracking disclosed AI use across the research record
What This Means for Authors
Use AI tools transparently. If you used Claude, ChatGPT, or another LLM to help write sections of your paper, disclose it in your submission according to the venue's stated policy. The risk of undisclosed use — which is increasingly detectable and treated as a form of academic misconduct — far outweighs the reputational cost of transparent disclosure.
The research record is a long-run reputation system. Papers that are later found to have been generated without genuine intellectual contribution damage the author's reputation in ways that outlast the paper itself.