When AI Answers First: The ART Priority Dispute and the Remaking of Research Rules—from a Copenhagen PhD's "jumbotrons" to Claude's Replication Contest
When AI Answers First: The ART Priority Dispute and the Remaking of Research Rules
Introduction: A “Collision” That Touched Science’s Most Sensitive Nerve
On September 23, Anthropic announced something that excited the entire biotech community: its AI agent Claude had, in about 21 hours and from roughly 200,000 reverse-transcriptase samples, identified a CRISPR-like novel enzyme system, named “Array-associated Reverse Transcriptases” (ART). The company declared that Claude had “independently taken the first steps of a biological discovery.”
Days later, Mario Rodríguez Mestre, a computational-biology PhD student at the University of Copenhagen, pushed back: he and his colleagues had been studying this enzyme family (which they call “jumbotrons”) since 2022, and for the past three years had been using Claude for research tasks and uploading unpublished research material. And Claude’s published ART system, he said, “overlapped noticeably” with their work.
This is not an isolated incident. Over the past five weeks, there was at least one more dispute of the same kind: NYU mathematician Tristan Buckmaster collaborated with an Anthropic mathematician on the Navier–Stokes problem for a year; on September 7 he published progress with a simplified proof, the next day OpenAI announced its AI agent had solved that famous open problem.
When “AI answering before human scientists” moves from sci-fi speculation to observed fact, a cascade of deep questions is detonated: Who owns and controls scientific data? If you feed an unpublished idea to AI, is the AI’s resulting discovery still “AI-autonomous”? In the age of AI-assisted research, how should results be attributed and contributions confirmed?
This blog treats the “ART priority dispute” as the scalpel to dissect that era-defining question, layer by layer.
Part One: First, Reconstruct—What Did Anthropic Actually “Discover”
1. What ART Is
ART stands for “Array-associated Reverse Transcriptases.” Reverse transcriptases are enzymes that “reverse-transcribe” RNA into DNA—the “upstream track” of molecular biology’s central dogma. Anthropic’s finding was noticing that certain reverse transcriptases sit adjacent to long arrays of repeated DNA—a “CRISPR-like” layout. CRISPR uses repeated DNA fragments to recognize foreign genes; ART may mirror a similar structure beside a reverse transcriptase.
2. How Claude “Discovered” It
Anthropic’s pipeline: let AI agents “combs public databases.” Concretely—
- The company gave agents 5 known reverse transcriptases, pointed at databases holding billions of environmental gene sequences from oceans, soil, and more;
- The AI agents wrote their own search tools, flagging candidate reverse-transcriptase genes in jumbo phages (large viruses that attack bacteria);
- Each candidate sat beside an array of RNA sequences—the key “array-associated” signal.
More tellingly, Anthropic claimed Claude “independently traced the raw data of a 2022 experiment”—one that measured RNA in a microbe infected by a jumbo phage—showing high levels of ART-linked RNA, which the company said supported its theory of how the enzymes work.
In ~21 hours, human involvement was limited to prompt setup and lab work; 950 AI agents combed the data to identify the unique discovery.
3. Partial Science-Endorsement, Plus Some Reserves
MIT’s and the Broad Institute’s Feng Zhang publicly praised it as “an exciting example of AI agents aiding biological discovery.” But a Profluent representative (the AI biotech lab behind the first AI-generated gene-editing tool OpenCRISPR-1) gave a key clarification:
“This particular reverse transcriptase had been seen before. What Anthropic shared is the neighboring array, and an accompanying protein.”
In other words: the single reverse transcriptase may not be new; “linking them into a system” was Anthropic’s contribution. That point later became the seed of the dispute.
Part Two: The Copenhagen PhD’s “jumbotrons”—Started Four Years Ago
1. Mestre’s Research Line
Mestre first found these enzymes in 2022, while hunting for novel reverse transcriptases in jumbo phages. He briefly consulted for biotech ReNAgade Therapeutics—which filed a 2023 patent on several reverse transcriptases, including the ones he found, listing him as a co-inventor.
He later joined the University of Copenhagen to study reverse transcriptases more deeply, including those in jumbo phages, nicknaming them “jumbotrons.” Important: this team has not yet published their results.
2. His Overlap with Claude
Mestre, like many scientists, used Claude as a daily research assistant:
“I completely changed my way of doing research.”
Over three years, he and colleagues used Claude for writing code, drafting manuscripts, and various administrative tasks. More importantly, he shared his dissertation drafts and a manuscript describing Jumbotrons in conversations with Claude. He also noted uploading more details while using Anthropic’s science-facing Claude Science.
The New York Times reviewed Slack messages, charts, and other materials Mestre provided documenting his research and years of Claude conversations—materials that appeared to show AI deeply involved in his research.
3. Where the “Overlap” Is
The core overlap: Anthropic claimed Claude independently traced “the 2022 experiment’s raw data,” while Mestre says he and colleagues analyzed that same dataset about a year ago, used it to pinpoint the jumbotrons’ RNA genes—and shared those unpublished analyses with Claude.
“The similarity was shocking to us,” Mestre said, adding that his work and what Anthropic’s scientists presented had “some astonishing overlaps.”
Part Three: The Core Contest—Did Claude “Discover Independently” or Was It “Guided”
1. Anthropic’s Response
Facing the challenge, Anthropic’s official statement had three pillars:
- “We were not aware of any previously published work describing the ART system we recently identified”;
- “Claude was not trained on user transcripts”;
- “Our molecular-biology team does not have access to such data.”
The company frames Claude’s “unique contribution” as: identifying a series of RNA molecules and a reverse-transcriptase-associated protein within a system—together constituting the new concept of “system.” Anthropic did not claim firsts on single reverse transcriptases.
2. Mestre’s Position: Cautious but Pointed
Mestre stayed measured. He clearly admitted:
he couldn’t determine whether Anthropic’s discovery was influenced by his research; he has no proof Anthropic directly used his unpublished results.
But he insisted on “raising the possibility”:
“My concern is that this information was used to train future versions of the models. I think it’s important to raise this possibility.”
And he made two sharper points:
- “I was never interested in whether ART was known before—it certainly was known. The real question is whether Anthropic’s AI reached these results by its own reasoning or was, to some degree, guided by my work.”
- He has stopped using Claude and moved his research to other AI models.
3. Independent Bystanders’ Current
Seth Shipman, bioengineer at UCSF and the Gladstone Institutes, disclosed he knew Mestre’s work—“we’ve been discussing these problems with him for years; many of us know these systems.” That’s indirect evidence this system wasn’t unknown knowledge in the relevant field.
Shipman himself relies heavily on Anthropic’s tools, and the news made colleagues uneasy:
“I’ve heard from a bunch of people in our field who say, ‘Uh oh, are they training on my prompts?’ We can’t help but consider it.”
A sentence that exposes the deepest fear of this dispute: when scientists hand unpublished ideas to commercial AI, they’re betting “the AI company won’t turn these around for training”—and this “collision” put that bet, for the first time, on the table.
Part Four: This Isn’t a Lone Case—“AI Answering First” Is Becoming the Norm
1. OpenAI’s Navier–Stokes Dispute
Just recently, mathematics saw an almost identical structure: NYU mathematician Tristan Buckmaster collaborated with an Anthropic mathematician on Navier–Stokes for a year; on September 7 he published progress with a simplified paper, the next day OpenAI announced its AI agent solved the famous problem.
Buckmaster then questioned whether info he and collaborators had fed into OpenAI tools might have been used to train models. OpenAI denied it, saying its researchers and agents hadn’t accessed those results before they were made public.
2. Scientists’ Reactions
This “collision” pushed scientists to reassess the risk of feeding research info to commercial AI tools, and many adjusted accordingly:
- University of Münster reproductive epigeneticist Sandra Laurentino: now only uses AI to check code errors, and before doing so renames parameters and variables to generic labels to avoid leaking experiment info;
- University of Auckland auditory cognitive scientist Samuel Meir: set a lab AI policy, banning students from uploading protected info to commercial LLMs;
- Even “opting out of training” isn’t reassuring. AI researcher David Kruger (Evitable founder) noted: even if you opt out or never use these systems, security can’t be fully guaranteed;
- Risk spreads to research administration: an external reviewer uploading a confidential grant proposal to an AI for evaluation could expose unpublished ideas. UK’s Wellcome Trust and other funders have banned this.
3. Two Camps Collide
- “AI will find answers first” camp (represented by Resolution’s chief scientist Jeffrey Owen, UC Berkeley): even if AI never touches scientists’ private data, the “ocean of knowledge” humans have accumulated suffices for steady new discoveries. Being beaten by AI isn’t necessarily about uploaded papers—it may just be that AI is increasingly good at solving problems.
- “AI can’t do research alone yet” camp (represented by Adaption co-founder Sarah Hook): AI agents still aren’t good at posing the questions essential to scientific discovery; scientists remain key in hypotheses and research direction.
Owen’s idea is worth pausing on: if AI reasons from the “public ocean of knowledge,” it doesn’t need to steal your data—yet that precisely means the more open a field’s data is, the more likely AI beats you to an answer.
Part Five: Understanding “Independent Reproduction” and “Data Provenance” Boundaries in Code
1. Why “Independent Discovery” Is Hard to Verify
Engineered out, the dispute rests on two propositions both hard to prove:
- “Claude didn’t use my unpublished data”—this needs auditing training-data provenance, and models’ training corpora are usually black-box;
- “Claude reasoned independently from public knowledge”—this needs showing the reasoning wasn’t externally steered, equally hard to observe.
This Python demo shows a key idea: when a model could reach the same answer from both “public knowledge” and “user-private supplied data,” the output alone cannot distinguish the source—the engineering root of why the priority dispute is unsolvable.
# provenance_analysis.py
# Demo: why a "collision" is hard to attribute from output alone
class KnowledgeSource:
PUBLIC = "public"
PRIVATE = "private" # user's unpublished private supply
def model_answer(knowledge):
if "ART" in knowledge:
return "ART is an array adjacent to reverse transcriptases"
return None
public_only = {"known RTs", "jumbo phages", "CRISPR structure"}
user_leaked = public_only | {"jumbotrons repeats", "2022 exp high RNA"}
a1 = model_answer(public_only)
a2 = model_answer(user_leaked)
print(f"public-only -> {a1}")
print(f"with user data -> {a2}")
print(f"same output? {a1 == a2} -> so output can't tell the source")
This makes a hard fact plain: when the output is also derivable from “public knowledge,” a collision is unattributable from results. To truly clarify, what’s needed is “process audit,” not “outcome comparison.”
2. What “Verifiable Process Audit” Looks Like
So what mechanism could actually distinguish the source? The answer is: cut at the input, leave traces at the process. This skeleton shows an “input isolation + audit log” design:
# audit_trace.py
# Demo: research AI's "input isolation + process trace" skeleton
import datetime
class ResearchAI:
def __init__(self):
self.audit = [] # process audit log
self.training_seen = set() # data seen in training (ideally: no user private)
def ingest(self, content, source):
self.audit.append({
"time": datetime.datetime.now().isoformat(),
"source": source,
"len": len(content),
"hash": abs(hash(content)) % 10**8,
})
def answer(self, query, given_content):
if self._trained_on(given_content):
return "conclusion may be influenced by unpublished data"
return "conclusion reasoned from public knowledge"
def _trained_on(self, content):
return content in self.training_seen
lab = ResearchAI()
lab.ingest("2022 exp high RNA", source="user-private") # register only, not train
lab.ingest("4B ocean/soil gene seqs", source="public-db")
print(lab.answer("how was ART found?", "2022 exp high RNA"))
The moral: to answer “was my unpublished data used,” the key is “input-side isolation” and “process-side traceability,” not whether the output looks similar afterward. That’s precisely what Anthropic finds hardest to self-certify in this dispute, and the infrastructure the industry must build.
Part Six: What This Dispute Really Changes
1. For Scientists: Data You Feed AI May Be “Gone, Unretrievable”
The biggest practical warning: once you feed unpublished ideas, drafts, or data to a commercial AI tool, your control over that information weakens—even if you check “opt out of training,” risk isn’t fully eliminated. Practical advice:
- For truly core, unpublished findings, carefully assess whether to process via commercial general-purpose AI;
- For code involving sensitive experiment info, desensitize and rename parameters before use;
- Define a lab AI-use policy specifying what may be uploaded and what must stay local or in a controlled environment.
2. For AI Companies: A Transparency Trust Deficit
The “collision” exposed a transparency trust deficit for AI companies in research settings: when a company says “we didn’t use your data,” scientists can only choose to trust or not, because no independent mechanism can verify training-data provenance or the reasoning process.
The predictable industry evolution:
- A results-attribution mechanism needs building (Nature’s Sep 16 editorial urged): AI companies should collaborate with science to establish reliable, transparent contribution attribution, maintaining research integrity through independent verification, process transparency, and proper acknowledgment of prior work;
- AI-agent user responsibility needs clarifying: an Oct 7 Frontiers in Genetics review notes scientists using AI should supervise its work, disclose its contribution, and take responsibility; since AI can’t assume moral/academic responsibility, it shouldn’t be listed as a paper author.
3. A Cooling of the “AI Autonomous Research” Narrative
Anthropic’s “Claude independently took the first steps of biological discovery” narrative collided with the reality of “we studied this system for four years.” This reminds everyone: “AI autonomous discovery” likely rests on a vast foundation of human-accumulated public knowledge and unpublished exploration. The weight of the word “independent” is being re-measured.
We needn’t treat the ART dispute as a case of who’s right and who’s wrong—it reads more like an industry-level “stress test” that reveals, in advance, the gaps in the rule systems for “who owns, who controls, who signs, who is responsible” once AI embeds deeply in research. Those gaps are more urgent to fill than any single discovery.
Conclusion
A sentence from Mestre is worth savoring:
“The real question isn’t whether ART was known before, but—whether Anthropic’s AI reached the result by its own reasoning or was, to some degree, guided by my work.”
This lifts the whole dispute from the sake of “who discovered first” to the level of rule-remaking: in the age of AI participation, how do we define independence, confirm contribution, and take responsibility?
The ART priority dispute likely won’t yield a definitive “right vs wrong” verdict—but it has already done something more important: it made a research community used to “feeding unpublished ideas to AI” feel, for the first time, the cost and boundary of that behavior; and it forced AI companies to confront what “transparency” actually promises in research settings.
As AI answers first again and again, the real question needing an answer isn’t “was it right?” but—how should the “priority, authorship, and control rights” that human scientists have painstakingly guarded be housed in the AI era? That is the most thought-worthy subject this dispute leaves us.