Same person, different role, different position
There is a familiar cycle on social media. During reviewing, people complain about the papers they have been assigned: what is this stuff, and why is it all autoresearch again? Once scores come out, authors complain that the reviewers are too mean and did not even understand their papers. The same person switches roles and ends up on the other side.
Then the discussion starts revolving around strange clues: the figures look too good, the prose is too smooth, the writing is too defensive—it must be AI. Bad figures and muddled prose? Also churned out by an automated pipeline. When opposite features can support the same conclusion, the method has no way to correct itself. Everyone is trying to identify how a paper was produced, while asking less and less about what it actually accomplished.
On the other side is the dismissal of papers altogether: even top conferences are full of trash, papers are toilet paper, and publishing them will not get you a job. That sentiment resonates easily. But whether a paper adds knowledge, demonstrates its author's ability, or leads to an opportunity are three different questions. We used to bundle those expectations together; now the cracks between them have widened. Saying papers are useless while rushing toward the next deadline is therefore not much of a contradiction.
A cycle that keeps speeding up
What interests me most is not whether AI can write papers, but the cycle taking shape between paper production, peer review, and career rewards.
Tools have lowered the cost of coding, experiments, and writing without reducing the demand for publications. Graduation, applications, hiring, and performance reviews still depend on those results. Time saved may not go into more thorough verification. It may simply go into the next submission.
For an author, one more submission means one more chance. For the system, it means several more people have to read, check, and discuss it. The benefit goes to the individual; the cost is spread across the community.
It gets worse once trust breaks down. If authors think reviewing is random, they have more reason to submit and resubmit, using more attempts to hedge against bad luck. More submissions make careful reviewing harder. A perfunctory review neither persuades the author nor reduces the next round of submissions. It only reinforces the belief that a different set of reviewers would do the trick.
The less people trust evaluation, the more attempts they want to make; the more attempts they make, the harder evaluation becomes.
This cycle does not require everyone to be flooding conferences with low-quality work. Serious researchers will also choose to submit more often in this environment. Blaming careless authors or overly junior reviewers does not explain the whole thing. The system itself rewards choices that make it more congested.
The awkwardness of credentials
So what does a top-conference paper still mean?
It can still bring work into public discussion, subject it to some scrutiny, and make it easier for others to assess you during applications and hiring. But the chain of inferences we casually attach to it has never followed automatically: the paper was accepted, so the work matters; the work matters, so the author is strong; the author is strong, so they deserve more resources. The more AI is involved, the more additional information that chain needs.
This creates an awkward situation: everyone thinks the metric is becoming less sufficient, yet everyone still needs to obtain it. With zero publications, saying papers do not matter sounds like an excuse for failure. Say it after publishing, and you have already played by those rules.
You have to get the credential first before people believe your criticism of it comes from judgment rather than disappointment.
However widely an evaluation system is doubted, it will not leave quietly as long as it still allocates opportunities.
Even if all sixty thousand papers were good
But there is a harder question beyond “too many bad papers.”
Suppose the authors of all sixty thousand submissions ran their experiments carefully, could explain why they did the work, took responsibility for their results, and contributed some genuine new finding. If the quality problem were actually solved, would the anxiety go away?
I do not think so.
Being correct, making a contribution, and deserving priority are three different thresholds. Jobs, funding, and opportunities to collaborate remain limited. For knowledge production, this might be a success. For people who make a living through research, it means fiercer competition. Doing good work and getting enough in return for it are becoming increasingly disconnected.
“At least I am careful and responsible” does not answer this problem. It rules out some things one should not do, but it does not establish why you deserve to be chosen over others who are just as careful and responsible.
Hands-on work, taste, and a clear conscience
So hands-on work, taste, and a clear conscience matter. But they cannot keep serving as the end of the discussion.
The value of hands-on work lies in encountering details, finding errors, and understanding results—not in the amount of manual labor. If a person working with tools still produces something less reliable than an automated system, the extra effort does not entitle them to a better evaluation. Hard work cannot change whether a conclusion is true.
Still, I insist on doing the work myself. Not because effort deserves a reward in its own right; it does not. But tuning a baseline, seeing unstable results across seeds, and encountering samples that do not fit the story tell me where my results are solid and where they are fragile. That understanding is what gives me the confidence to judge other work later, including work produced by models.
Learning and output also need to be considered separately. A newcomer who reproduces an experiment, checks data, or works through a proof is training their ability to judge future work, even when the exercise has no publication value. “AI does it faster” does not imply “humans should not learn it.” Worth learning, worth publishing, and worth continuing are three different questions. Keeping them separate helps us avoid demanding academic rewards for practice, or abandoning our own learning because a tool is faster.
Taste is similar. Pointing out a paper's weakness in a comment section and deciding what to pursue or abandon before the results are known are not equally difficult. If models can participate in those decisions, or even improve them, use them. I do not want to find some patch of territory that models can never enter. Today we say they cannot choose research questions, tomorrow that they lack experience, and the day after that they cannot persist over time. Retreating a little further each time only makes us more uneasy. What is more worth examining is what the decisions I make with these tools actually lead to.
Talk to people earlier
The really difficult judgments often come after the investment.
After several months, it is hard to accept that your work might not be that important. “They just did not understand it” becomes especially attractive. Sometimes that is entirely true, but it can also explain almost any failure. Careless authors, earnest authors who misjudged their work, and authors who were genuinely misunderstood can all tell the story of good work going unrecognized. The author's own conviction cannot distinguish the three.
Talking to people has been one of the most useful things I have done. I just used to ask the wrong questions. I kept asking, “How can I write this better?” when I should have asked, “Does this conclusion hold up?” or “Is this question still worth pursuing?” The earlier you ask, the easier it is to listen—and to change things.
Presentation remains a basic responsibility of the author. Readers are not obliged to untangle a confused argument, and reviewers should not replace checking the content with guessing from the prose. Clear writing helps people judge the work; beautiful writing cannot fill a gap in the evidence. These responsibilities do not conflict.
As for the cycle of reviewers and authors blaming each other, eventually we have to return to the specific dispute: what is wrong, what evidence is missing, and what result would change either side's mind? The loudest person online is not necessarily the one with the best judgment.
Everyone has Opus. Why do the same people still stand out?
Because using the same model does not mean working under the same research conditions.
Data, compute, real-world settings, collaborators, and time for trial and error determine whether an idea can be tested. It is easy for a model to make a suggestion. Running that suggestion in a real environment for several months takes a whole other set of conditions. A team that knows where the problem is, can organize the testing, and can mobilize resources can quickly turn model output into results. Without those conditions, what you get may mostly be more candidate ideas and more polished material.
Resources determine what is possible, judgment determines where those resources go, and reputation brings the next round of resources. They reinforce one another. Securing resources, organizing collaboration, and keeping things moving are abilities in their own right. “The strong get stronger” hides all these distinctions.
Voting with your feet has barriers too
When a single paper cannot tell us enough about a person's abilities, evaluation will naturally pay more attention to concrete work and a longer record: whether people use the code, whether results have been reproduced, what products someone has built, whether former collaborators want to work with them again, and whether their writing helps others understand a problem. Blogs, tutorials, tools, data, and infrastructure should also receive recognition proportionate to their contributions.
I think these signals are more honest than a publication count, and they deserve more attention. But they are not friendly to newcomers: before someone uses your work, they have to see it; before you have a reputation, you need a chance to collaborate. The public route offered by publication should not be discarded, though that is no reason to go back to looking only at papers.
Changing the metrics will not end the habit of optimizing for them. Once people start comparing stars, followers, and influence, new forms of packaging will appear. Promotion itself is not a problem. Getting work to the people who need it can expand its contribution. The question to ask yourself is: am I helping more people use it, or am I mainly trying to prove that it has impact?
Submissions, and the choices beyond them
All of this has changed how I see submitting papers. Submission is one way to share results, seek feedback, join a discussion, and pursue opportunities. It deserves serious preparation, but the value of a piece of work should not be reduced to one accept or reject. If it is rejected, distinguish the comments that identify real flaws from those based on misunderstanding, and decide whether the work is still worth pursuing. If it is accepted, that acceptance will not do the remaining verification for you.
We can be more straightforward about career choices too. Some people want to dig deeply into a question, some want to build a product, and others care more about income and a good team. I want money, real-world results, and a reputation. I do not think that makes research dirty. I just do not intend to let them decide everything for me.
Finally
At the start, I said that many things we used to take for granted are coming loose. Sometimes it really does feel as though everything is collapsing, with AI rolling over one thing after another. I do not know what it will all become, and I do not expect a reassuring new set of rules to arrive any time soon.
What I can do is probably still a few straightforward things: make the work solid, talk to people earlier, be honest about credit, and bring evidence when challenged. Keep learning, stay open to new things, and allow tomorrow to overturn today's judgment. Problems in the evaluation system need institutional solutions; individuals cannot absorb them all through self-discipline. But I want to hold up my own part.
And I do not want to become someone who has nothing except research. I like the plants on my windowsill, spending an afternoon in my usual café, browsing a tokusatsu shop for an hour and leaving without buying anything, and being with people I love. None of these things will make my papers more likely to be accepted. They do not need to.
Then I thought: nothing was ever going to stay the same. Conferences will change, evaluation will change, tools will change, and the things I believe today will change too. In that case, I do not have to wait for everything to settle down before I start living well.
When the world is shaking, these things help me know where I stand.