In November 1987 the first NIPS was held in Denver. About six hundred people came, and the proceedings held around ninety papers. A conference was a simple thing then: the people working in a field met once a year and told each other what they were doing. The proceedings were a by-product of the meeting.
In 2025 the NeurIPS main track received 21,575 valid submissions and accepted 5,290. In September 2026, abstract registrations for ICLR 2027 passed sixty thousand, against 19,525 valid submissions the year before. By one count, that is more than every previous ICLR combined.
My sense over the last two years is that there are more conferences, each one is bigger, and getting a paper accepted means less and less. This essay is an attempt to make that feeling precise: what conferences have lost, what they still have, and what they will turn into. The conclusion is already in the title.
Conferences no longer do the publishing
A conference used to do two things for us: get work out, and pick out the good work.
Dissemination moved to arXiv long ago. A paper waits more than four months from submission to decision and more than half a year to the actual meeting. One rejection and resubmission makes it a year. In this field a result often stays fresh for only a few months. By the time the poster goes up, the people who needed to read it have read it, and sometimes two rounds of follow-up work are already out.
Filtering was crushed by scale. Tens of thousands of submissions need tens of thousands of reviewers, so the reviewer pool keeps getting diluted. In 2021 NeurIPS ran a consistency experiment, sending a tenth of the submissions to two independent committees. The finding was that if the whole review were run again, about half of the accepted papers would be replaced. That year there were just over nine thousand submissions.
So as a publishing mechanism, one of the conference's two functions has been replaced and the other is failing.
Certification is running on inertia
None of this has made conferences unimportant. Graduation, hiring, promotion, and visa applications still count top-conference papers.
The reason is that certification is a derivative of filtering, and there is a lag between the two. The people who use the certificate cannot see the quality of the filter, only its output. Reviewers know how random reviewing is. Hiring committees and departments may not. So insiders have mostly stopped believing in it while outsiders still rely on it.
There is a second reason. Even if filtering is close to a lottery, "five accepted papers" still says one thing: this person can keep producing work that clears a basic bar. It certifies output, not quality.
When this kind of certification fails, it does not become void overnight. It inflates. From "has a top-conference paper" to "how many first-author papers" to "any orals", and then on to citations, which group the work came from, and recommendation letters. The users keep stacking secondary filters on top. It is the same mechanism as degree inflation, and it has already started.
AI will use up the inertia
AI-generated research entering this system is inevitable. It comes in two forms, and they break different things.
The first is filler: it looks like a paper, it can pass review, and it contributes nothing. It removes the ceiling on submission volume, and no mechanism that spends human attention per paper can survive that. This is no longer a prediction. ICLR 2027 caps every author at 20 submissions, and the organizers say plainly that the aim is to keep the conference from being flooded with low-quality papers. At the same time, the conference has begun offering authors AI-generated feedback on their drafts. Both the writing end and the reviewing end are being handed to machines.
The second form goes deeper: the conclusion is right, the experiments are real, and no human did the work. Nothing is wrong with the paper, but it no longer says anything about its author. I said above that a top-conference paper at least certifies the ability to produce. That layer goes too.
At that point output is no longer scarce. Three things are: the taste to decide what is worth doing, the verification that a result is true, and someone taking responsibility for the conclusion. Authorship will stop meaning "I did this" and start meaning "I vouch for this".
Why there is no replacement
If conference certification is unreliable, why not evaluate differently, for example by whether a piece of work actually gets used?
Because that is very hard to assess. Usage takes two or three years to show, and the young researchers who most need to be evaluated do not have that time. Citations, stars, and downloads can all be gamed, and a citation does not distinguish between a passing mention in related work and an entire project built on your method. More fundamentally, impact is a counterfactual question: without your paper, would the later work have happened anyway? Someone reads your work, changes direction, and never cites you. You show that a path is a dead end and save ten groups half a year. None of that leaves a trace.
All of these difficulties come down to one tension:
Any metric that scales can be gamed, and judgment that can be trusted does not scale.
What can actually evaluate a piece of work is the judgment of someone who knows the area and has read it carefully. That judgment is expensive, slow, and tied to specific people. Peer review was an attempt to scale it. It held at a few hundred papers and collapsed at tens of thousands. The devaluation of conferences and the inflation of credentials are both expressions of this one tension. AI has only pushed it to the limit.
Back to the beginning
Take away publishing and certification, and what is left of a conference? What it was at the start: a group of people meeting.
That is not a bad thing. The more public information floods in, the more valuable the things exchanged face to face become: who is working on what that has not been released, which result does not really reproduce, which direction people have quietly given up on. None of that was ever in the papers. My guess is that the valuable meetings of the future will be small ones among people who trust each other, while the big conferences look more like job fairs and trade shows.
But a circle has an inside and an outside. However random double-blind review is, it gave people without connections a way in: a student from an unknown school could be noticed for a single paper. Once filtering fails, if recommendation letters, advisors, and word of mouth fill the gap, the people who lose most are the ones outside the circle.
This has also already begun. One of the new ICLR 2027 rules says that if no author on a paper is qualified to review, each of those authors may be on only one such paper. Unqualified roughly means never having published at a major conference. The rule has its reasons: last year a fifth of submissions had no qualified reviewer among their authors, and those papers were accepted at about half the rate of the rest. But the effect is that a newcomer either submits one paper or finds an insider to put their name on it. The organizers themselves acknowledge that this may encourage gift authorship. "Do you know someone inside the circle" has been written into the submission rules for the first time.
The same holds for researchers looking at their own work. Getting into a top conference used to be at least a clear signal that the thing had been accomplished. Once that signal is distorted, how you judge whether several years of your own work meant anything becomes a more private question.
So I have no answer, only a question: can the next filtering mechanism be both hard to fake and open to people without connections?
Until it arrives, conferences will keep being held and papers will keep being submitted. We will all know that the things that matter are happening in the hallway.