I have told several people now that I will stop reading anything they send me until they stop externalizing the cost of their AI use to me. When something arrives with “just skim it” or “check before using,” that is where I stop, because it signals handing off work that I did not agree to.
Externalizing costs is a standard problem where others pay for the consequences of your actions: buying bigger cars to feel safe in a crash, who cares about pedestrians; forcing open the closing elevator door so you can save one minute, while others wait; extracting fossil fuels and leaving the cleanup to the next generation; or companies taking wild risks and letting others pay through bailouts or clean up after bankruptcy when it does not work out. In personal interactions, externalization is an asshole move, in economics it is the now normalized foundation of “privatizing profits and socializing losses,” letting society pay for cleaning up environmental damage or bailing out companies.
With AI, externalization is hitting my time and attention. Polished-looking text or code became cheap to produce, but reading it and providing feedback did not. I can no longer rely on the heuristic that producing text took effort as a sign that somebody is trying to communicate something of value.
The pattern
I see this pattern frustratingly often: Somebody asks an LLM a question, and sends me the answer, sometimes with a note that it is AI-generated so I should skim it or verify it before trusting it. I am sure they are trying to be helpful.
If you tell me to just skim it, why should I? Why is the work of finding the point in a long sloppy text suddenly mine, why is the checking on me? What value did you provide beyond the original prompt – if the value is in the prompt, give me that prompt – I have access to a model too.
The point is not that the text was written by AI. The problem is that it is unedited or poorly edited slop that takes effort to filter. It is a transfer of labor that a disclaimer will not turn into consent. Adding “this is AI-generated, double-check it” just makes the laziness explicit – it was apparently not worth your time. At the very least apologize for sending raw AI output.
Externalizing AI cost undermines review and feedback mechanics broadly
As an educator and researcher, I spend a lot of time giving feedback as part of my job. I usually have a lot of patience for helping others to improve their writing, their argument, their code, their research ideas, or their startup ideas. I am happy to provide feedback on a draft or review a paper or proposal. Creating the draft was usually a sign of trying, of putting effort in, and of wanting to improve.
Now that creating the initial artifact to be reviewed has become so cheap to send out a draft for feedback, why even think about it before sending? Why not just externalize the real remaining cost of the work to the reviewer? This has unfortunately become very common:
Open-source maintainers are getting buried by AI slop issues, security reports, and code contributions. In the past, testing and security researchers already got pushback when they submitted potential bugs found with fuzzing tools to open source projects, putting the cost of evaluating their research on those maintainers under the guise of helping them IRB review and without asking for consent (responsible researchers usually invested heavily in checking their reports before submitting them). Now anyone well-meaning can generate a mountain of plausible-looking issues and pull requests, burning out maintainers and shutting down bug bounty programs left and right.
I stopped offering free homework resubmissions, even though letting students revise and resubmit is good pedagogy (see specifications grading or grading for equity), because it started to get exploited: Just submit AI-generated solutions without checking and feed the instructor’s feedback into an LLM for generating a regrade request.
Academic peer review is probably the most obvious to suffer. With double-anonymous reviews, the only real penalty for submitting a weak paper is the time to wait for reviews to come back – so why not just submit half-baked AI papers for review and use the reviewers’ labor to improve the paper by feeding reviews back into the agent. Now the production costs are so low, why not do this with dozens of papers in parallel. This summer, I intentionally wrote a one-line review, because it felt exploitative (with more detailed comments only to other reviewers). I do not see how peer review survives much longer with this asymmetry.
Collateral damage – I’m sorry
I do not mind the use of AI to support writing, coding, or research, but I get very annoyed at the externalization of costs. Unfortunately, it is effectively impossible to distinguish a lazy shortcut with AI that externalizes costs from a genuine but failed attempt by somebody putting in real work with the intention of actually learning from feedback. In self-defense, the moment something reads as AI-generated and bloated, I now get suspicious and reject investing effort in feedback. I am aware that this creates collateral damage, but I see no alternative.
The retreat to small trust networks seems inevitable
Unfortunately, I think it is pretty bleak. We are destroying systems, building inefficiencies and friction (transaction costs) just to reduce free riding but harm everybody in the process. I suspect we will escape to personal trust networks and give up on many opportunities from a more open past.
I still vividly remember listening to the last chapter of Domingos’ book The Master Algorithm in 2016, which described his utopia of a future where my personal AI agent talks to your personal AI agent, negotiating on our behalf so that neither of us has to. I remember it because it read extremely dystopian to me back then, but unfortunately I think this is where we are actively heading now: We start using AI in self-defense to sort through AI slop and might lose any signal in the process. We use AI to summarize emails to their one-sentence point or use AI to do the skimming or fact checking that the sender was too lazy to do for us. We consider it for code review and even academic peer review, probably losing the value human feedback used to provide. This is what eroding trust looks like and low-trust societies are inefficient and no fun.
Given that what’s exploited is mostly attention and mental labor, traditional economic solutions to externalities like privatization and structures for maintaining commons are unlikely to help, and anonymity in peer review and open source also eliminates any control through reputation or shame (though I hope somebody will figure this out eventually). What’s left is disengagement, defensive AI use, and approaches that introduce cost and friction, such as capping feedback and revision opportunities, imposing costs per submission, or requiring deposits. All of these hit everybody, not just bad actors or well-intentioned actors who use AI without knowing better.
Where I think we can fight back is in long-term personal relationships, friends and family, colleagues, PhD students, collaborators, and open-source contributors we personally know. Here trust and reputation through repeated interactions matter, externalization can be called out, and labor-sharing arrangements can be discussed. Cleaning AI-generated text before sending can become a norm and additional negotiations about when and how to share things for feedback can happen, for example agreeing on sending a narrative in rough bullet points rather than verbose text generated from it. Here, sending raw AI output comes with an apology rather than a disclaimer. In personal networks, we learn who to trust and whose messages to ignore, where to be generous with feedback and where a brisk response or even an AI response suffices. Retreating to a small circle of trusted people is not a happy ending; it is giving up on the openness that made academia and open source great spaces to work in.
So use AI as much as you like, but do not make others pay for it. If you are reading this because I sent you this link, I hope you now know why.
— Christian Kästner
Your article finds me in a place where my student just send me her article entirely written by IA. Where the first sentence placing the capital of Tunisia, Tunis where her data came from, in the south of the country, whereas Tunis is at the very north. And she is living there. I asked her at least to read another time her text but the second version was worse as she used IA for the second shot.
I dont want to correct IA, that's what I said. But I realised she cannot write correctly in English, her third language. In fact, without IA, the tasks were also there to select people that are good at writing and have clear ideas. I still don't know what to do. Write the article myself ? We are really at a fork in the road.
I have a hard time feeling sympathetic to the downfall of some of these "goodwill" systems, especially academic publishing.
Even before AI, we were already playing a massive game of pretend. I’ve lost count of how many times I had to deliberately bloat my papers and make the content look way more convoluted than it actually was, just because human reviewers equate unnecessary complexity with rigor.
Academia has always rewarded human-generated slop and jargon-heavy prose. And don't even get me started on the reviewers themselves. In computer science and software engineering, we constantly have our practical, built-from-scratch systems judged by "quack" reviewers who have never written a single line of production code in their lives.
AI is just accelerating the collapse of a fundamentally broken, elitist framework that refused to hold itself accountable. If AI ends up killing traditional peer review in academic research, I will be the first to celebrate.