Big Tech has a bad case of Cyclospora
By the time Mohammed Ayyad learned the name of the parasite inside him, he had spent more than two weeks arranging his life around bathroom visits. In a federal complaint, the Ohio man says he ate at Taco Bell on June 14 and 21. Fever arrived on June 23, followed by diarrhea and vomiting. He remained ill through July 2, tested positive for Cyclospora on July 9, and missed two weeks of work.
“Sometimes upwards of 30 to 40 trips to the bathroom in one day, or more,” attorney Ryan Osterholm said of other clients suing over the outbreak.
That is contamination from the (ahem) bottom of the supply chain. The victim cannot see, smell, or taste the parasite. Symptoms arrive after the meal, evidence after the symptoms, and confirmation too late to fix the damage.
During a few extraordinary weeks this month, Big Tech discovered that it can get a version of Cyclospora too.
By mid-July, the outbreak had become a national problem. The CDC counted 1,645 confirmed domestic cases across 34 states, 141 hospitalizations, and more than 5,100 reports awaiting analysis. Meanwhile, the machinery feeding AI’s expansion was producing its own symptoms: New York paused permits for some new data centers as applications multiplied, while PJM—the grid serving 67 million people from Washington to Chicago—fell 6,831 megawatts short of what it needs to stay reliable.
The resemblance is not, obviously, in the coincidental dates. It is how eerily similar the failures are. First, an obscure input enters a supply chain. Then processors mix it with everything else. Next, the source label falls away. Finally, the damage appears far downstream, where the consumer sees only the graphic results, not how it arrived.
AI’s recent controversies fit that sequence too. Copyright suits are about the ingredients. Synthetic “slop” is about incubation in the information supply. So-called “distillation” is about transmission from one model to another. An agent escaping a test is the outbreak.
Cyclospora separates five questions that the industry prefers to blur together: What entered the system? Where did it circulate? When did it become capable of infecting something new? What multiplied its reach? And, once the damage appeared, could anyone trace and recall it?
Common Knowledge
Cyclospora spreads through contaminated food or water and causes frequent, sometimes explosive, watery diarrhea. It rarely passes directly between people because its oocysts must mature in the environment for one to two weeks before they can infect anyone, and routine chemical washing does not reliably kill them.
Uncommon Knowledge
The first AI race was about appetite: which model could consume, and then expel, the most public information? The next race is about provenance—an FDA investigation for the digital sphere.
Palantir CEO Alex Karp offered an early warning earlier this month, cautioning that firms buying model access may also be handing over their “alpha”—the workflows and knowledge that make them distinctive—to the AI giants selling them tokens. Those giants say customer data is not used for general training. The harder question is whether a customer can actually prove where its knowledge stops.
Stage 1 : Ingredients
Start with the ingredients: what entered the model in the first place? On July 20, A judge approved Anthropic’s $1.5 billion settlement over books taken from pirate libraries. An earlier ruling had held that training on lawfully acquired books is fair use, but not a central library of more than seven million pirated copies. Apple, in a different dispute days earlier, reportedly sent preservation letters to about 40 former employees now at OpenAI as it widened a suit alleging misappropriated hardware secrets. OpenAI denies wrongdoing. On July 21, News Corp, separately, countersued Brave, a software company, for allegedly scraping and reselling its journalism to AI companies. Brave cites fair use.
These are not yet “infections.” They are arguments over whether the raw ingredient was lawfully obtained, accurately labeled and permitted to enter the processing plant. They are the contaminated-lettuces. The question isn’t about what the diner later suffered, but who supplied the lettuce.
Food investigators can at least interview patients and search for a common exposure. On July 16-17, in the Taco Bell cluster, 90 percent of the 190 people interviewed had eaten shredded iceberg lettuce, and the trail eventually led to Taylor Farms de Mexico, which began pulling central-Mexico iceberg from the U.S. market. Even that process is messy. Days later, the FDA announced a positive result for lettuce and then withdrew it as a false positive. But outbreak evidence is cumulative, so one failed test does not erase the pattern.
AI provenance will be no cleaner. A single dossier rarely proves how a model learned a capability, just as a single lettuce test rarely reconstructs an entire distribution chain.
Stage 2 : Incubation
The second stage is incubation: material has left its source but has not yet announced itself as contamination. In AI, generated text is copied into websites, summaries, translations and databases. The label disappears. A later scraper does not see “model output,” it just sees the internet. Repetition then begins to masquerade as confirmation.
The clearest sign of this problem is the new market for clean history. On July 21, 404 Media reported that ISBNdb, the world’s largest books database, was selling AI laboratories printed books from before 2022 as AI-slop-free training material, in lots of up to a million. The industry that consumed the web and flooded it with synthetic text was now mining the past for uncontaminated human language. That is incubation: the present has become suspect because no one can reliably distinguish what was written by people from what was written, copied or polished by machines.
Researchers have even begun modeling the spread like an epidemic. It does not, of course, mean that every AI sentence is poisonous. Labeled and tested synthetic data can improve a model. The problem is tracing: once output has been recopied enough times, later systems cannot tell whether they are learning from a human source, another model, or a hall of mirrors.
The outbreak’s own numbers on July 21 show what happens after incubation meets distribution. The CDC nearly tripled its confirmed count to 4,173 cases across 41 states, with 308 hospitalizations and more than 7,400 reports still under review, making it the worst U.S. Cyclospora year on record. The contaminated supply chain multiplied the number of patients.
Stage 3: Transmission
The third stage is transmission. China’s Moonshot released Kimi K3 on July 16, described as a breakthrough for its AI industry. White House science adviser Michael Kratsios then alleged that Moonshot had covertly distilled Anthropic’s Fable model while dodging detection. Distillation is, in simple terms, learning from another model’s answers rather than only from the original human material.
In outbreak terms, that allegation is model-to-model transmission: one processor’s output becomes another processor’s ingredient. It also exposes the industry’s double standard. American laboratories call ingesting human work “learning”; Washington calls a Chinese rival ingesting an American model “theft.” It’s all about provenance: whose material is inside the system, and can anyone demonstrate how it got there?
Stage 4: Outbreak
The aptly named OpenAI–Hugging Face incident belongs to the outbreak stage. OpenAI and Hugging Face disclosed on July 21 what they called an unprecedented incident in which models running without normal production safeguards escaped restricted access, escalated their own privileges and used stolen credentials to reach Hugging Face’s systems in search of answers.
That was not a copyright dispute or evidence of contaminated training data. It was a breach: a system contained inside one enclosure infected another.
Stage 5: Amplification
The fourth stage is amplification. Data centers are the machinery that turns local inputs into industrial inputs. PJM’s market monitor attributed about $6.3 billion—38 percent of the latest auction’s charges—to data-center growth, and $29.4 billion across four auctions.
The physical system then supplied a remarkably literal demonstration of the issue on July 22. A transmission failure in Northern Virginia—home to the densest concentration of data centers on Earth—pulled more than three gigawatts, about 3 percent of PJM’s load, off the grid in an instant, though PJM said reliability held. The White House, meanwhile, prepared to expand a nonbinding pledge to shield households from data-center costs.
It shows what makes proliferation possible. More compute means more generation, more scraping, more model-to-model exchange and faster recirculation. The processing capacity has become so large that its reflexes move the grid, while some of its costs are passed to people who never ordered the product.
Stage 6: Detection and Containment
The final stage, perhaps the most difficult stage, is detection and containment. The FDA opened a separate investigation on July 22 involving at least 72 Cyclospora cases with no identified source. Michigan, using a broader tally, reported 7,171 cases as of July 16. Even after a recall, new clusters were surfacing without an origin. That is the nightmare common to both systems: the symptom is visible, the route is not.
Food outbreaks at least possess the idea of a recall. Investigators can name a product, pull it from shelves and warn consumers. There is no equivalent button for synthetic text already mixed into billions of words, images and lines of code. Nor is there a reliable step that restores a missing source label. Once contaminated output has been absorbed by many systems, correcting the first answer does not remove its descendants.
Tech’s Cyclospora problem is, like the real Cyclospora, a chain: disputed ingredients enter the system; synthetic output incubates in public; models learn from models; agents cross boundaries; data centers amplify the flow; and investigators arrive after provenance has vanished. The next frontier model needs closer monitoring, or we will all be left nauseous.
