We Use AI. We Are Not Building Slop.

Nobody Asked for More AI

On January 9 this year, at its annual vote, the American Dialect Society named slop the word of the year for 2025: low quality, high volume content, most typically produced by generative AI. [1]

That vote was not an isolated mood. Pew surveyed 3,488 US adults in late June 2026 and found that 52 percent are now more concerned than excited about the growing use of AI in daily life, up from 37 percent in 2021. Only 9 percent are more excited than concerned. For the first time, a majority of adults under 30 landed in the concerned group. [2]

The feeling gets much stronger when AI shows up as a physical neighbor. Gallup asked Americans in March 2026 whether they would favor an AI data center being built in their local area. Seventy-one percent opposed it, including 48 percent who opposed it strongly. For comparison, 53 percent oppose a new nuclear power plant nearby. [3] The Annenberg Public Policy Center watched local opposition to data centers climb from 49 percent to 61 percent in the space of four months. [4]

There is even a measurable penalty for saying the words out loud. In experiments with more than 1,000 US adults, adding "artificial intelligence" to an otherwise identical product description lowered people's intention to buy it. The effect ran through emotional trust, and it was strongest for high risk categories, the kind where a failure costs you money or health. [5]

Read that last study as a founder and the commercially optimal move is obvious. Say less about AI. Put something warmer on the homepage.

I am not going to do that. But I do want to be precise about what we are doing, because I think the skepticism is mostly earned, and I would rather explain myself than get quietly filed under the thing people are tired of.

---

The Skepticism Is Earned

The complaint underneath the word slop is not really about machines. It is about accountability.

A confident paragraph that nobody wrote, nobody checked, and nobody can be held to is worth very little, no matter how good it sounds. And language models are extremely good at producing exactly that. When researchers asked chatbots to supply references for systematic reviews, 28.6 percent of the references from GPT-4 and 39.6 percent from GPT-3.5 turned out to be hallucinated, and 91.4 percent from Bard. [6]

Models have improved a lot since that study. That is not the point. The point is structural: text that has the shape of a citation is not a citation, and a system whose only output is fluent prose gives you no way to tell the difference.

About half of Americans told Pew they expect AI to make people worse at thinking creatively and worse at forming meaningful relationships. [7] I do not think that worry is silly either.

So when someone sees "AI" attached to health information and their first reaction is a flinch, I read that as a healthy immune response. Health is exactly where an unverifiable confident paragraph does the most damage.

---

What We Are Not Doing

We are not training a model on nutrition and letting it talk.

That is the thing people are right to be tired of, and it is not what AllNutrition is. There is no model here that has absorbed a lot of nutrition text and now produces plausible nutrition opinions. If that were the product, the criticism would land, and I would deserve it.

The actual design goal is close to the opposite. Every claim we show is supposed to resolve back to a specific published paper, written by named humans, reviewed by other humans, with its own methods section, its own reference list, and its own funding disclosure. The library is tens of thousands of peer-reviewed studies, plus a small set of the most reputable clinical guidelines. If a claim cannot be traced to one of them, it should not be on the screen.

Peer review is a flawed institution. Reviewers work for free, fit the work into schedules that are already full, and often review papers slightly outside their exact specialty. One estimate put the global volume of unpaid reviewer time at over 100 million hours in a single year, with the US share alone worth more than $1.5 billion. [8] It is a system running on donated labor.

It is also, still, the best mechanism humanity has built for separating what is known from what someone wishes were true. I would rather build on top of an imperfect system with real accountability than replace it with a fluent machine that has none.

---

But Why Not Just Follow the Experts?

I am not anti-expert. Some of the clearest nutrition explanations I have come across were written by doctors, dietitians, and researchers who built large audiences by being genuinely good at this.

The problem is not that experts are untrustworthy. It is that expertise is narrow by construction, and every human being makes mistakes.

Deep expertise can even make it worse. Someone who has spent twenty years studying muscle and its relationship to longevity will read every new question through that lens. It is an excellent lens, and it is not the only one. That same person may not weigh your kidney function, your medications, your lipid trajectory, your family history, or a body of evidence sitting one field over that points somewhere else entirely. That is not a character flaw. No individual can hold all of this at once.

And the underlying literature is noisy enough to fool careful people. In a much cited 2013 paper, two researchers picked fifty ingredients at random from a cookbook and went looking for cancer studies on each. Forty of the fifty had at least one. Of the studies they found, 39 percent concluded the ingredient raised cancer risk and 33 percent concluded it lowered it. [9]

A credentialed, honest expert can build a confident story out of that literature. So can the equally credentialed, equally honest expert who believes the opposite. Both of them will cite real papers.

The way out is not a better expert. It is reading far more of the evidence than any one person can, and being explicit about where it agrees, where it conflicts, and how much each piece deserves to be weighted.

Which brings me to the size of the pile.

---

Nobody Can Read This

So we counted. Not by hand, and the method is worth spelling out, because the method is really the argument.

AllNutrition fills its library using 19 PubMed searches: core nutrition and diet, dietary patterns, gut and microbiome, metabolic health, nutrition and chronic disease, key nutrients, longevity and healthy ageing, clinical nutrition, and four searches covering supplement safety, effectiveness, food-first alternatives, and interactions with drugs and other supplements. For this post I had Claude take that exact set of queries, the ones running in production, join them into a single search, and count what PubMed returns for each publication year. Nothing was hand picked. The full search is linked under the chart, so you can open it in PubMed and get the same number.

In 2025 it matched 201,205 papers. Roughly 3,870 a week. Around 550 a day, weekends included.

Ten years earlier the same search matched 99,250. It has doubled in a decade, while PubMed as a whole grew by half.

I would rather flag the softness in that number than pretend it away. Keyword searches overcount and undercount: a paper that mentions calcium once falls into the net, and a paper on time restricted eating that avoids every one of our terms does not. Records carrying both an electronic and a print date can land in two different years. A differently tuned query set would produce a different total, and this one is deliberately tuned for recall.

None of that touches the part that matters. Whichever way anyone measures it, in nutrition and across science generally, the line goes up and the slope steepens.

Two things in that chart do the work.

The first is accumulation. Since 2000 the search has matched about 2.1 million papers, and almost a third of everything it matched between 2000 and 2023 was published in the last four of those years. That is not a backlog anyone clears by reading harder over a long weekend.

The second is the dashed part. If the last decade's rate simply holds, 2030 brings roughly 5,500 papers a week, and the five years from 2026 through 2030 add something on the order of a million more on their own. That is a trend line rather than a forecast, and it assumes nothing accelerates. In a moment I will explain why that assumption is probably too generous.

The growth is also not evenly spread, and the uneven part is more interesting than the total.

Notice what the fast growing clusters have in common. They are not obscure corners of biochemistry. Gut health, longevity, and supplements are three of the things people ask about most, and three of the areas with the largest marketing budgets aimed at them. The evidence is expanding fastest exactly where the stakes of getting it wrong are highest and where the incentive to shade a result is strongest.

---

And It Is About to Get Faster

Scientists use AI too. Not as a curiosity, as a daily tool.

A 2025 analysis in Science Advances looked at vocabulary patterns across more than 15 million PubMed abstracts and concluded that at least 13.5 percent of 2024 abstracts were written with LLM assistance, reaching 40 percent in some subfields. [10] Reviewers are doing it as well: an analysis of peer reviews submitted to several AI conferences estimated that between 6.5 and 16.9 percent of the review text could have been substantially modified by a model. [11]

Draw the obvious conclusion. Writing a paper is getting faster. Reviewing one is getting faster. Running an analysis on an existing dataset is getting much faster. Meanwhile the trials that take years still take years, and the number of humans available to scrutinize any individual paper has not moved.

Volume goes up. Attention per paper goes down. Retractions passed 10,000 in a single year for the first time in 2023, driven largely by sham papers and compromised peer review. [12]

So the coming decade offers more evidence and less scrutiny per piece of it. That is precisely the situation where careful, consistent, tireless triage is worth the most. And it is not a job a person can do, or a team, or a team that would then have to keep doing it every week forever without getting bored or acquiring a favorite theory.

---

What We Actually Use AI For

Not for having opinions. For reading.

Retrieval, not recall. Answers are assembled from papers pulled out of the library at the moment you ask, not from what a model happens to remember. Every claim carries the sources it came from, and you can open them.

Appraisal, one paper at a time. Agents read the full text of each paper and score it on four things: methodology, conflict of interest, reporting transparency, and evidence strength. Those scores are visible to you, not buried.

Following the money. For each paper we look at the authors' publication history, their institutions, their disclosures, and who funded the work. Industry funding on its own is not a conflict. What matters is funding combined with findings that happen to serve the funder, which is why disclosed industry ties that produced results against the funder's interest score well rather than badly.

Ranking, not deciding. Weaker or conflicted evidence gets weighted down rather than silently deleted, because "the industry funded trials say X and the independent ones say Y" is often the single most useful thing you can know about a question.

None of that requires the AI to be smart about nutrition. It requires it to be relentless about reading, sorting, cross referencing, and flagging, at a scale and a cadence no human team can sustain. The knowledge stays human. The machine does the part humans cannot: keeping up.

I wrote a while back that AI is not the product, and that has not changed. The mission is helping people make better decisions about food and longevity. AI is the most powerful tool currently available for that job, and if something better shows up, we will use that instead.

---

The Test I Would Ask You to Apply

Slop is not defined by whether a machine touched it. Plenty of slop is written by people, and some of it gets published in journals. Slop is content that cannot be traced and that nobody is accountable for.

So the test is simple, and it works on us as well as on anyone else making claims about your health.

Follow the claim. Can you get to a named paper, in a named journal, with named authors, a named funder, and a methods section you or your doctor can argue with? If yes, you are looking at evidence, whatever tool helped surface it. If not, you are looking at slop, no matter who said it or how many followers they have.

You are completely entitled to dislike AI. I would just ask that you judge what we build by whether the claim tracks back to a paper.

Because the alternative is not a world without AI in nutrition. It is a world where 200,000 relevant papers appear every year, almost nobody reads them, and the confident voice with the best distribution wins by default.

That world already exists. It is the one that produced the advice you are currently trying to make sense of.

---

Sources

1. American Dialect Society. 2025 Word of the Year Is "Slop." Vote held January 9, 2026. americandialect.org/2025-word-of-the-year-is-slop/

2. Pew Research Center. Young adults in the U.S. are increasingly wary of AI, concerned it will take jobs. August 18, 2026. Survey of 3,488 US adults, June 22–28, 2026. pewresearch.org/short-reads/2026/08/18/young-adults-in-the-us-are-increasingly-wary-of-ai-concerned-it-will-take-jobs/

3. Gallup. Americans Oppose AI Data Centers in Their Area. Survey of 1,000 US adults, March 2–18, 2026. news.gallup.com/poll/709772/americans-oppose-data-centers-area.aspx

4. Annenberg Public Policy Center, University of Pennsylvania. Opposition to Local Data Centers Rises Sharply, Annenberg Survey Finds. Survey of 1,320 US adults, June 16–July 19, 2026. annenbergpublicpolicycenter.org/opposition-to-local-data-centers-rises-sharply-annenberg-survey-finds/

5. Cicek M, Gursoy D, Lu L. Adverse impacts of revealing the presence of "Artificial Intelligence (AI)" technology in product and service descriptions on purchase intentions: the mediating role of emotional trust and the moderating role of perceived risk. Journal of Hospitality Marketing & Management. 2025;34(1):1–23. doi.org/10.1080/19368623.2024.2368040

6. Chelli M, Descamps J, Lavoué V, et al. Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis. Journal of Medical Internet Research. 2024;26:e53164. doi.org/10.2196/53164

7. Pew Research Center. What the data says about Americans' views of artificial intelligence. March 12, 2026. pewresearch.org/short-reads/2026/03/12/key-findings-about-how-americans-view-artificial-intelligence/

8. Aczel B, Szaszi B, Holcombe AO. A billion-dollar donation: estimating the cost of researchers' time spent on peer review. Research Integrity and Peer Review. 2021;6:14. doi.org/10.1186/s41073-021-00118-2

9. Schoenfeld JD, Ioannidis JPA. Is everything we eat associated with cancer? A systematic cookbook review. American Journal of Clinical Nutrition. 2013;97(1):127–134. doi.org/10.3945/ajcn.112.047142

10. Kobak D, González-Márquez R, Horvát EÁ, Lause J. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances. 2025;11(27):eadt3813. doi.org/10.1126/sciadv.adt3813

11. Liang W, Izzo Z, Zhang Y, et al. Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews. Proceedings of the 41st International Conference on Machine Learning (ICML). 2024. arxiv.org/abs/2403.07183

12. Van Noorden R. More than 10,000 research papers were retracted in 2023, a new record. Nature. 2023;624:479–481. doi.org/10.1038/d41586-023-03974-8

13. Publication counts in both charts come from the PubMed E-utilities API, retrieved August 26, 2026, using AllNutrition's own 19 ingestion queries joined into one search. The full search with its year filter is linked beneath each chart, and the exact query text sits in the source code that renders them.