The rot in AI research has moved beyond a few bad actors, it has become a systemic failure that threatens the credibility of the entire field. The testimony from this conference attendee is damning: presenters who cannot explain their own methods, papers submitted identically to multiple venues, and GitHub repositories left open with LLM chat histories that reveal fabricated data. This is not a side effect of rapid progress; it is a crisis of integrity that demands immediate, structural intervention. We agree with the call to reject AI-generated papers at the desk, but the problem runs deeper than a single policy change.
Consider what happens when a researcher who spent months designing experiments and nights debugging code submits their work for review. They compete against someone who prompted a language model, copied the output, and never read the result. The honest researcher faces a double loss: their paper may be rejected by a system already overwhelmed with low-effort submissions, and the field they trusted to advance is now publishing "garbage," as the post bluntly puts it. This dynamic actively punishes rigor. When we see stories like the data journalists who exposed a fake pollster by tracing digital footprints, we recognize the investigative tools exist to catch fraud. The question is why conference program committees are not applying similar scrutiny before accepting papers.
The solution cannot rest solely on "hard desk rejects" for AI-generated text. Detection tools are unreliable and easily circumvented. Instead, the community must change what counts as a valid submission. Require authors to submit a brief, recorded video explanation of their core contribution before the review process begins. Mandate that all code and data be checked for reproducibility by a separate committee. These are not radical ideas; they are basic quality gates that other scientific disciplines have used for decades. The goal is not to slow down progress but to ensure that designing LLM chat experiences that work for everyone includes designing review processes that filter out submissions the author themselves cannot defend.
A specific consequence to watch: if top conferences do not act, the signal-to-noise ratio will collapse entirely. Researchers will stop trusting any published result, and the most talented people will stop submitting their best work to these venues. The bounded canvas for AI agents that nimble teams are building means nothing if the foundation of academic validation is sand. The next step is not a plea for better behavior, it is a demand for verification. If a presenter cannot answer how they represented a knowledge graph, their paper should not have been on the program in the first place.