Dr. Peter Andrew Jamieson — Academic in an information evolution

← Back to Meanderings, Opinions, and Articles

It’s Not AI Slop, It’s a Curation Problem

October 2026 · written with the help of Claude (Anthropic)

In the last two years, I have watched three fields that I care about change significantly because of large language models. Each time, the reaction has followed a similar pattern. First there is disbelief, then there is a shrug, and then there is a new word for the output we do not like.

The current word is slop. Merriam-Webster made it the 2025 Word of the Year and noted that generative AI helped spread it1. I think the word points at a real problem, but it names the wrong one. The problem is not that people are using machines to make things. The problem is that people working with machines are going to make far more of everything, and we do not have a good way to find the good parts.

That is a curation problem, and going forward, I believe it will be the crux of this whole shift.

Three waves, all made of language

Writing (roughly two years ago). The first space to tip was prose. An analysis by the SEO firm Graphite of roughly 65,000 web articles found that by November 2024, newly published AI-generated articles had overtaken human-written ones in volume2. Since then, the two have stayed at roughly a 50/50 split. Interestingly, AI-generated articles show up far less in Google results and in what ChatGPT and Perplexity cite3. Keep that detail in mind, since it is already a curation story.

Software (roughly December 2025). The second space was code. Late 2025 brought Claude Opus 4.5, Gemini 3, and GPT-5.2 in quick succession, and many people, including Andrej Karpathy, point to December 2025 as the moment coding agents crossed a threshold4. Karpathy’s framing was that agents went from brittle demos to completing long, multi-step tasks coherently. Developers who had used AI coding tools with mixed results for years started describing a different experience in which they ran several agents in parallel and effectively scaled themselves horizontally5.

Mathematics (roughly a month or two ago). The third space is math, and it moved the fastest. In May 2026, an internal OpenAI model resolved the unit distance conjecture, which had been open since 1946, during routine evaluation6. In August, OpenAI published ten results in pure mathematics and theoretical computer science with formal Lean proofs7. In early September, Axios reported that OpenAI’s GPT-6 Astra had produced results on long-standing open problems and that an unreleased research version of Claude made a major improvement on decades-old work related to the Riemann hypothesis8. Nature covered OpenAI’s claim of a resolution of the Navier-Stokes Millennium Problem the same week9. MIT’s Andrew Sutherland described the shift as AI doing mathematical research rather than solving problems that are handed to it8.

Writing, software, and math (prose, code, and proofs) have one thing in common. They are all language. Natural language, programming languages, and formal mathematical language. Language is what LLMs are best at.

We will say it was easy

Before going further, I want to address a reaction that I can already hear. Those fields were always going to fall. Writing is pattern matching, code is syntax, and competition math is search.

Richard Hamming anticipated this reaction thirty years ago. In The Art of Doing Science and Engineering, he discusses a geometry theorem-proving program that found an elegant proof its designers had not thought of, and Samuel’s checkers program that learned to beat a state champion. We could say the checkers program learned and the geometry program showed originality. However, as his lecture slides put it, once the program exists we regard it as “nothing but a rote routine”10. From that perspective, there is no way for a machine to demonstrate that it thinks, because the moment it does the task, we redefine the task11.

Chess was considered “real intelligence” until Deep Blue, and then it was “just search.” Writing was a uniquely human craft until it was not. I expect that within a year, someone will tell me the open problems that fell in 2026 were “the easy kind.” Hamming’s point is that this reaction is about our egos and not about the problems.

So, let us assume this trend continues. If it does, two questions matter.

Question 1: What else is defined by language?

The three waves so far hit fields that are obviously textual. However, many fields become language once they are written down carefully.

Hamming himself argued that programming is closer to novel writing than to classical engineering, and he predicted that by 2020 domain experts would write their own programs rather than handing them to programmers11. He was a few years early, but he was right in spirit. The boundary between describing a problem and solving it is collapsing because both are now done in language.

A note on robots. The robots are coming. Hardware taught me that language can describe the physical world, and robotics is where language meets the physical world with intelligence in the loop. A robot that can take an instruction in natural language and act on it turns the physical world into one more language-defined space. I believe physical space with intelligence is a space that will be targeted soon. The good news is that the complexity of the space and its dynamics make the problem wicked.

Question 2: Which language spaces change next?

If a field can be expressed in language and has some way to check the output (tests for code, proofs for math, assays for biology, courts for law), I expect it to change next. Verification is what took math from “LLMs are bad at arithmetic” to “LLMs are producing Lean proofs” in about eighteen months. Fields with fast and cheap checks will change first. Fields where checking is slow and contested, such as policy, medicine, and much of the humanities, will change more unevenly and probably more painfully.

My hypothesis: we are going to make a lot more

This is the core of my argument. As we learn to teach the next generation how to use agents to build things (artifacts, software, papers, analyses, and media), people working with these tools are going to create more. Much more.

Some people will call that slop. However, if volume is what makes something slop, then we should be honest that we have been making human slop for a long time. Hamming observed that knowledge has roughly doubled every 17 years since Newton, and that this applies not only to theorems but also to recordings of Beethoven’s Ninth and to the TV programs to watch or not watch11. Most of what humans produce has always been mediocre. Sturgeon’s Law was not written about AI.

The interesting question was never who made something. The interesting question is how do we find the good stuff?

How we curate today

We already have several curation systems. None of them is perfect for the new scale that is coming.

1. Curation by experts: peer review

In academia, our approach is peer review combined with notable venues where we would like our work to appear. We submit to a conference or journal, blind reviewers judge the work, and hopefully the best ideas are propagated. This is human-gated curation, and it is already struggling.

ICLR 2026 received 19,525 valid submissions, about 68% more than the previous year, and handled 76,139 reviews from 18,054 reviewers13. An analysis of those reviews estimated that about 21% were fully AI-generated13. ICLR’s response for 2027 was to cap authors at 20 submissions each13. That is not a solution; it is a dam. Human-gated review does not scale to the volume we can already see coming.

On a personal note, I am starting to question the value of publishing in the traditional sense. My next paper might simply be an arXiv release, and I am not sure I care whether it is gatekept. If the gate cannot keep up with the volume, and if the people I want to reach will find the work through other channels anyway, then the gate is serving a different purpose than it used to.

2. Curation by crowds: the algorithm

The second system is the recommender system, or the feed. When we talk about “the algorithm,” we mostly mean aggregated human actions such as how many people clicked, liked, watched, commented, or shared. This is the classic like and subscribe. Each action is a small vote, and ideas are propagated based on the tally. This approach scales well, but it optimizes for engagement, which is not the same as quality, and it tends to amplify what is already popular.

3. Curation by networks: word of mouth and cities

The third system is the oldest. Ideas spread from person to person. The Heath brothers’ Made to Stick examines what makes an idea memorable enough to pass on, and Jonah Berger’s Contagious examines why people share14.

Cities are a useful lens for how network size affects this spread. Bettencourt, West, and colleagues showed that measures of innovation, such as the number of new patents, grow superlinearly with city population (with a scaling exponent of roughly 1.2), so a city twice as big produces more than twice the innovation15. The proposed explanation is the network of human interactions, since a bigger city gives each person more contacts. Schläpfer and colleagues tested this directly using hundreds of millions of phone call records from Portugal and the UK16. They found that each doubling of a city’s population gives the average person about 12% more contacts, and that a resident of Lisbon accumulated roughly twice as many reciprocated contacts as a resident of a small rural town over the same period16. They also showed that this denser network makes it easier for things to spread through it. Geoffrey West brings this data together for a general audience in his book Scale17.

The implication for curation is twofold. Bigger and denser networks move ideas faster, which is good. However, the internet is effectively the largest city ever built, and in that city every person is connected to more ideas than they could ever evaluate. The network effect that makes cities productive also makes the curation problem harder.

The question I do not have an answer to

So we have database approaches (indexes, search, and venues) and population approaches (likes, feeds, and networks). What I am actually interested in is more personal.

How do I find my next ideas?

I have watched certain television and liked it. What is the next thing I will like? I have read certain papers that changed how I think. Which of the next ten thousand papers will do that again? When creating things becomes cheap, finding things becomes the scarce skill. What we each discover will matter almost as much as what we all create.

I believe curation will be the crux of all of this, and I expect AI approaches to be part of the answer. Agents could read on our behalf, know our tastes and our open questions, and check claims rather than count clicks. The Graphite data already hints at this, since the flood of AI-generated articles mostly does not surface in search and answer engines3. Some curation is already happening quietly through machines.

However, I do not want a curator that only tells me what is popular, and I do not want one that only tells me what I already like. I want one that finds the thing I did not know I was looking for. That is a hard problem, and it is a much more interesting one than complaining about slop.

It is not AI slop. It is a curation problem, and that is what I need one of you with these new superpowers to solve for me…selfish.

References

  1. AP, “Merriam-Webster’s 2025 word of the year is ‘slop’,” December 15, 2025.
  2. Graphite, “More Articles Are Now Created by AI Than Humans,” October 2025. Summary via Stacker.
  3. Axios, “AI writing hasn’t overwhelmed the web yet,” October 14, 2025.
  4. Latent Space, “[AINews] WTF Happened in December 2025?” See also Anthropic, “Introducing Claude Opus 4.5,” November 24, 2025.
  5. R. Jansen, “Claude Opus 4.5 changes everything,” DEV Community.
  6. A. Romero, “The Month AI Conquered Math: The Full Story,” The Algorithmic Bridge.
  7. Build Fast with AI, “AI News Today August 2 2026.”
  8. Axios, “AI is changing math,” September 8, 2026.
  9. Nature, “OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’,” September 8, 2026. doi:10.1038/d41586-026-02842-5.
  10. R. W. Hamming, “Learning to Learn, Session 7: Artificial Intelligence II” (lecture slides), Naval Postgraduate School.
  11. R. W. Hamming, The Art of Doing Science and Engineering: Learning to Learn (Gordon and Breach, 1997; Stripe Press, 2020), Chs. 1, 4, and 6 to 8.
  12. P. Jamieson and K. B. Kent, “Odin II: An Open-source Verilog HDL Synthesis Tool for CAD Research,” IEEE FCCM, 2010. Odin II documentation.
  13. Pebblous, “ICLR 2026 AI Peer Review Crisis,” and on the 2027 submission cap, “ICLR 2027 caps author submissions at 20, sparks debate”.
  14. C. Heath and D. Heath, Made to Stick (Random House, 2007); J. Berger, Contagious: Why Things Catch On (Simon & Schuster, 2013).
  15. L. M. A. Bettencourt, J. Lobo, D. Helbing, C. Kühnert, and G. B. West, “Growth, innovation, scaling, and the pace of life in cities,” PNAS 104(17):7301–7306, 2007.
  16. M. Schläpfer, L. M. A. Bettencourt, S. Grauwin, et al., “The scaling of human interactions with city size,” Journal of the Royal Society Interface, 2014 (arXiv:1210.5215).
  17. G. West, Scale: The Universal Laws of Life, Growth, and Death in Organisms, Cities, and Companies (Penguin Press, 2017).

This post was written by me, with the help of Claude (Anthropic) for editing and for tracking down the citations.

Peter Jamieson is an Associate Professor of Electrical and Computer Engineering at Miami University. More of his work, publications, and projects are at drpeterjamieson.com.