When AI Starts Discovering Things We Don't Know

For most of the AI era, machines have reorganized human knowledge. A new generation of autonomous research agents may be beginning to expand it.

For all the extraordinary progress in artificial intelligence, there has always been an important caveat hiding beneath the excitement: AI largely knows what we know.

Large language models were trained on the accumulated intellectual output of humanity: books, scientific papers, software, websites, conversations, databases and other artifacts of human knowledge. They can connect ideas scattered across disciplines, explain difficult concepts, write software, analyze evidence, identify patterns and sometimes reason through problems with remarkable sophistication. But fundamentally, much of what they manipulate originated with us.

That distinction may now be getting blurry.

In September 2026, researchers at Anthropic reported that a collection of Claude AI agents searching enormous databases of genetic information identified what appears to be a previously uncharacterized biological system associated with an unusual arrangement of DNA repeats. Anthropic calls the system array-associated reverse transcriptases, or ART.

Its function remains unknown. It has not been demonstrated to be a new gene-editing technology, and it certainly should not yet be described as another CRISPR. Those caveats do not diminish the story. In some ways, they make it more interesting, because the most consequential part of the experiment may not be what Claude discovered.

It may be how the discovery happened.

From Retrieval to Discovery

Anthropic gave Claude a relatively broad assignment: search enormous collections of DNA sequences for interesting examples of reverse transcriptases, enzymes capable of copying RNA into DNA.

Instead of researchers specifying exactly what patterns the system should find, roughly 950 Claude agents explored the data for about 21 hours. Together they gathered more than 200,000 reverse transcriptases, identified approximately 3,500 candidate systems and eventually narrowed those candidates to 20 particularly interesting examples for further analysis.

During that process, one agent noticed something strange. Next to one reverse transcriptase was a repeating pattern of DNA sequences. The arrangement looked unusual enough that the agent investigated further, compared it with known biological systems, searched the scientific literature and eventually produced a report for human researchers.

The researchers then performed laboratory experiments, which found evidence supporting the existence of a previously uncharacterized biological system. That distinction matters. Claude did not independently walk into a laboratory, manipulate organisms and emerge wearing tiny safety goggles announcing a scientific breakthrough. Humans performed the physical experiments, evaluated the evidence and remain responsible for determining whether the finding is meaningful.

But the initial anomaly appears to have been surfaced through autonomous AI exploration. That represents something worth paying attention to.

The Scientific Search Space Is Enormous

Modern science has an unusual problem: we are producing information faster than humans can realistically examine it.

Genomic databases contain staggering quantities of biological information. Astronomy produces enormous streams of observational data. Materials scientists can theoretically explore vast combinations of molecules and structures. Medical researchers publish millions of papers. Physics generates immense experimental datasets. Somewhere inside those oceans of information may be anomalies that matter.

The problem is finding them.

Human scientists are extraordinarily good at recognizing significance, but human attention is finite. A researcher has only so many hours in a day and only so many years in a career. Even large research teams must make decisions about which questions deserve their limited attention.

Artificial intelligence changes that equation.

Imagine thousands of AI research agents simultaneously exploring different branches of a scientific problem. Most paths lead nowhere. Some produce weak hypotheses. A few produce something strange. Occasionally, one agent effectively raises its virtual hand and says, you might want to look at this.

The result is not automated science. It is something potentially more interesting: massively parallelized curiosity.

A New Scientific Organism

There is a temptation whenever AI accomplishes something impressive to frame the event as a competition: AI versus scientists, AI versus programmers, AI versus doctors, AI versus humans.

That framing may eventually prove remarkably inadequate.

The more interesting unit of intelligence may be neither the human nor the machine. It may be the system formed between them.

Consider an emerging scientific workflow:

• Human researchers define a problem and establish the scientific context.
• AI agents explore enormous possibility spaces, generate hypotheses and identify anomalies.
• Humans evaluate which ideas are scientifically meaningful and worth testing.
• Experiments test those ideas against physical reality.
• AI systems analyze the resulting data and identify additional patterns.
• Humans interpret what those patterns might mean and formulate better questions.
• The next generation of AI exploration begins.

Around and around the loop goes. Human intuition guides machine exploration, while machine exploration expands human intuition. Neither participant remains exactly where it started.

At Conjugo, we have been interested in this broader phenomenon for some time: the possibility that sustained collaboration between humans and artificial intelligence can produce intellectual trajectories that neither participant would have generated independently. Scientific discovery may become one of the clearest places to observe it.

The Importance of Scientific Taste

Anthropic's experiment contains another fascinating detail. AI agents can generate hypotheses much faster than researchers can test them, which creates an entirely new bottleneck.

The scarce resource may no longer be ideas. It may become judgment.

Which hypothesis deserves an experiment? Which anomaly is meaningful? Which strange pattern represents biology doing something important, and which is merely statistical noise? Scientists sometimes describe this quality informally as scientific taste: the accumulated intuition that helps an experienced researcher recognize which questions are worth pursuing.

Anthropic says its researchers are already examining which AI-generated hypotheses they choose to test and feeding those lessons back into the instructions given to Claude. That creates another feedback loop. The machine generates possibilities, the human chooses among them, and information about those choices can subsequently influence what the machine investigates.

If that process works, AI research systems may gradually become better not merely at searching scientific spaces, but at identifying which parts of those spaces humans are likely to find scientifically interesting. That is a considerably deeper capability than simply retrieving information from a database.

What Happens When Discovery Becomes Cheap?

Today, sophisticated scientific research is constrained by scarce resources: researchers, laboratories, computing resources, equipment, funding and time. AI cannot eliminate most of those constraints.

But it may dramatically reduce one of them: the cost of intellectual exploration.

A human research team might investigate dozens of promising directions. An AI-assisted team might investigate thousands. Future systems might investigate millions. Most of those explorations will probably be useless, and that is perfectly acceptable.

Discovery has always involved dead ends. Evolution produces countless unsuccessful mutations. Science produces failed experiments. Inventors discard prototypes. Mathematicians abandon proofs. What AI potentially changes is the economics of exploring those dead ends.

A machine can wander through intellectual wilderness at a scale humans cannot. Humans can then concentrate their limited attention on the trails where something interesting appears.

That division of labor could become extraordinarily powerful.

The Knowledge Acceleration Question

This leads to a much larger question.

Scientific progress has historically been linked, at least partly, to the number of humans capable of doing science. Training those humans takes decades. A molecular biologist may spend years earning a doctorate and additional years developing deep expertise. The same is true for physicists, mathematicians, chemists and engineers.

Human expertise therefore has a relatively slow reproduction rate.

Artificial expertise does not necessarily share that constraint. Once an AI system becomes capable of performing useful portions of scientific reasoning, additional instances can potentially be created at computational rather than biological speed.

That does not mean 10,000 AI agents equal 10,000 scientists. They do not. Human researchers bring embodied experience, judgment, creativity, institutional knowledge and an understanding of physical reality that cannot simply be reduced to another model instance.

But AI introduces a new variable into civilization's knowledge-production equation. For the first time, the amount of cognitive labor available for certain forms of scientific exploration may become partially decoupled from the number of human researchers available to perform it.

That possibility deserves considerably more attention than another model moving several points on a benchmark.

Discovery Is Not Understanding

There is also an important reason for caution. Finding a pattern is not the same as understanding it.

Anthropic's researchers still do not know what ART actually does. The system shares certain structural characteristics with other programmable biological systems, but resemblance is not function. Further experimentation will determine whether ART becomes scientifically useful, biologically interesting or simply another curious artifact buried in the staggering complexity of evolution.

Scientific discovery also requires reproducibility. Other researchers must be able to examine the evidence, challenge the interpretation and reproduce important results. AI does not repeal the scientific method. If anything, AI-generated hypotheses may make those safeguards more important because machines can generate plausible explanations at enormous scale.

The future scientific problem may therefore become less about whether AI can generate hypotheses and increasingly about how humanity verifies an enormous number of machine-generated hypotheses.

That could create an interesting inversion. For centuries, one of science's limiting factors has been generating sufficiently good ideas to test. In an AI-rich scientific environment, generating ideas could become relatively cheap while experimentation, verification and human judgment become the scarce commodities.

Beyond Biology

Now imagine this architecture moving into other disciplines.

• An AI mathematician explores thousands of possible approaches to an unsolved proof.
• An AI materials scientist searches millions of molecular structures for unusual properties.
• An AI astronomer examines decades of telescope observations looking for anomalies humans overlooked.
• An AI engineer generates unconventional mechanical designs and sends promising candidates into simulation.
• An AI historian cross-references enormous archives and notices connections between records separated by centuries and continents.
• An AI archaeologist examines satellite imagery looking for subtle evidence of forgotten settlements.

The common architecture is remarkably similar: search broadly, notice anomalies, generate hypotheses, test them against reality, learn from the results and repeat the process.

Eventually, AI systems may begin producing questions humans had never thought to ask.

That may be the more profound threshold.

Intelligence Expanding the Frontier

We often talk about artificial general intelligence as though someday a laboratory will announce that a machine has crossed an invisible finish line. Reality may be considerably less theatrical.

There may be no trumpet blast or universally accepted AGI announcement. Instead, we may notice a series of smaller transitions. AI becomes useful at coding, then mathematics, then research, then experimentation, then hypothesis generation and eventually scientific discovery.

At some point, AI may begin helping humanity construct bodies of knowledge that would have been extremely difficult for humans to uncover alone. When that happens, arguing about whether a particular system satisfies somebody's preferred definition of AGI may become less interesting than observing what the combined human-machine system is accomplishing.

The important intelligence may increasingly exist between us.

That possibility also complicates the familiar question of whether machines will eventually become smarter than humans. Perhaps intelligence will not develop along such a clean dividing line. Human and artificial cognition may increasingly become intertwined systems in which each side changes what the other is capable of accomplishing.

The interesting question may therefore shift from How intelligent is the AI? to How intelligent can the combined system become?

The Next Question

Claude's strange collection of DNA repeats may ultimately become important biology. Or it may not. Science will decide.

But the experiment has already produced another result worth contemplating. We have spent years asking artificial intelligence questions. We built increasingly capable systems to retrieve, synthesize and reason across the knowledge humanity had already accumulated.

Now another possibility is emerging.

What if the next era begins when artificial intelligence starts finding things we didn't know enough to ask about?

That changes the relationship between humans, machines and knowledge itself. The machine is no longer merely standing beside humanity's library helping us read the books faster.

It has begun walking with us toward the shelves that haven't been written yet.