AI Is The New Partner In Scientific Discovery
Artificial intelligence is crossing a remarkable scientific threshold. Once used mainly to analyse data, predict structures and process existing knowledge, AI is beginning to help researchers identify unexplored biology, generate hypotheses and even design new biological systems for laboratory testing. From Claude’s role in identifying a previously uncharacterised enzyme system to AI-designed proteins and bacteriophages, the frontier is shifting from machines that assist science to machines that participate in discovery. But as AI generates possibilities faster than laboratories can test them, profound questions emerge about validation, scientific credit, ownership, safety and responsibility. The future of science may increasingly involve machines proposing what is possible, while human judgement, experimentation and nature itself determine what is true.
TECHNOLOGY & SCIENCE
For most of the artificial-intelligence revolution, the relationship between AI and science seemed relatively straightforward. Scientists produced knowledge and machines helped them process it. Computers searched databases, analysed images, compared sequences, modelled molecules and performed calculations at speeds no human researcher could approach. That relationship is now beginning to change.
In September 2026, Anthropic reported that its Claude AI system had helped researchers identify a previously uncharacterised biological system hidden within an enormous database of DNA sequences. The researchers did not begin by telling the system exactly what it should find. Instead, AI agents were asked to explore biological information, identify unusual patterns, investigate them and propose candidates worthy of further attention.
What emerged was a family of what the researchers called array-associated reverse transcriptases, or ARTs. According to Anthropic, around 950 Claude agents worked for approximately 21 hours, consuming about 210 million tokens as they searched, compared evidence and generated research reports. Human scientists subsequently examined the findings and conducted laboratory experiments.
There is an important qualification. The biological function of the newly identified system is not yet understood. Claude did not independently solve a biological mystery and present scientists with a finished discovery. What happened may nevertheless prove more consequential in the long run: the machine helped scientists decide where to look.
That possibility marks an important frontier in the relationship between artificial intelligence and scientific discovery. AI is beginning to move beyond explaining what humanity already knows towards helping researchers decide what humanity should investigate next.
FROM SCIENTIFIC ASSISTANT TO DISCOVERY PARTNER
Artificial intelligence has been used in research for decades. Machine learning can identify patterns in enormous datasets, classify medical images, model climate systems, search chemical libraries and assist researchers in fields ranging from astronomy to materials science. The extraordinary advances of recent years have greatly expanded those capabilities, allowing machines to tackle scientific problems once considered exceptionally difficult.
Perhaps the best-known example is AlphaFold, developed by Google DeepMind. Predicting the three-dimensional structure of proteins from their amino-acid sequences had challenged scientists for decades. Protein shape matters because structure is intimately connected with biological function, yet experimentally determining those structures can require considerable time and specialised equipment. AlphaFold transformed the scale at which useful structural predictions could be made. Its Protein Structure Database now contains predictions for more than 200 million protein structures and has been used by millions of researchers around the world.
AlphaFold demonstrated the extraordinary value of AI as a scientific predictor. Prediction, however, is not quite the same as discovery. The next frontier is whether artificial intelligence can examine vast bodies of scientific information, recognise something humans have overlooked, formulate a plausible hypothesis and recommend an experiment capable of determining whether that hypothesis is correct.
This is what makes the Anthropic experiment particularly interesting. Instead of merely asking Claude to summarise scientific literature or answer questions about known biology, researchers gave AI agents considerable room to explore. They searched sequence databases, investigated unusual patterns, checked existing scientific literature and produced candidate findings for human researchers to evaluate. The significance therefore lies less in the particular enzyme system than in the method used to find it.
WHEN MACHINES BEGIN GENERATING THE QUESTIONS
Science is often described as a process of answering questions, but one of its greatest intellectual challenges is deciding which questions are worth asking. There are more possible experiments than laboratories could ever perform, more combinations of molecules than researchers could ever synthesise and quantities of biological information far beyond the capacity of any individual scientist to examine directly.
Scientific progress has therefore always depended partly upon judgement. Researchers notice anomalies, connect ideas from different fields and recognise that an apparently insignificant observation may deserve another look. They decide that one hypothesis is sufficiently interesting to justify months or even years of work. Artificial intelligence could dramatically expand this process by generating possibilities at a scale no human research team could match.
Anthropic says a research campaign involving AI agents can generate hundreds or thousands of candidate reports. If such systems become capable of continually searching enormous scientific datasets and producing plausible hypotheses, science could encounter an unexpected problem: there may eventually be more promising ideas than laboratories have the capacity to investigate.
The scientific bottleneck would then move from generating hypotheses to choosing among them. Laboratories may increasingly have to decide which AI-generated possibilities deserve scarce experimental resources, specialist researchers and research funding. Far from making human judgement redundant, this could make it more valuable. The scientist of the future may spend less time searching manually for possibilities and more time deciding which of thousands of machine-generated possibilities genuinely matter.
THE LABORATORY REMAINS THE FINAL TEST
The distinction between a plausible scientific idea and an actual scientific finding is fundamental. A language model can produce an explanation that sounds convincing and still be wrong. A computational model can predict that a molecule should behave in a particular way only for researchers to discover that biology refuses to cooperate.
Scientific reality is not determined by eloquence or computational probability. Proteins must fold, molecules must bind, cells must respond, enzymes must function and proposed materials must survive physical testing. A potential medicine must ultimately demonstrate acceptable safety and efficacy. This is why laboratory validation remains indispensable.
In the Anthropic work, human researchers reviewed AI-generated candidates and experimentally investigated the biological system. The machine accelerated exploration, but experimental evidence remained the bridge between computational possibility and scientific knowledge. AI can suggest where the truth might lie; evidence must still establish whether it is there.
This distinction becomes even more important as artificial intelligence moves into medicine. AI is increasingly being used across pharmaceutical research to identify biological targets, generate potential drug molecules, optimise chemical structures, predict toxicity and assist aspects of clinical-trial design. The commercial attraction is enormous because developing a successful medicine can take many years and cost vast sums, while most experimental compounds never become approved treatments.
If AI can eliminate weak candidates earlier or identify promising ones faster, even modest improvements could produce enormous economic and medical benefits. Yet faster design must not be confused with faster medicine. An AI system may generate a molecule in moments, but preclinical research, toxicity studies, manufacturing, clinical trials and regulatory scrutiny cannot simply be removed by computational speed. Biology contains complexities that models may approximate but cannot automatically conquer. AI may compress parts of drug discovery, but it does not repeal biology.
FROM DISCOVERING BIOLOGY TO DESIGNING IT
An even more consequential frontier is emerging. Artificial intelligence is not only being used to search for biological systems that already exist. Researchers are beginning to investigate whether it can design biological systems that did not previously exist in that particular form.
Recent experiments involving genomic language models have demonstrated the ability to generate complete bacteriophage genomes. Bacteriophages are viruses that infect bacteria rather than people. Researchers experimentally tested AI-generated designs and found that some could produce functional phages capable of infecting bacterial hosts.
The distinction between this work and the Anthropic experiment is crucial. When AI searches genomic databases and identifies an overlooked biological pattern, it is helping scientists discover something that already exists in nature. When AI generates a functional biological sequence that researchers subsequently construct and test, it is participating in design. The first capability expands humanity’s ability to explore the biological world. The second could expand humanity’s ability to alter it.
There are potentially important benefits. Because bacteriophages attack bacteria, they are being investigated as possible tools against antibiotic-resistant infections, one of the most serious long-term threats to global health. AI-designed biology could eventually contribute to new medicines, industrial enzymes, proteins, materials and treatments. But the ability to design functioning biological systems also raises questions that go well beyond ordinary technological optimism.
WHEN BIOLOGY BECOMES GENERATIVE
The public became familiar with generative AI through words, pictures, music and video. A person can describe an image and a machine can create one, or ask for a piece of text and receive it within seconds. Generative biology applies a related principle to biological information.
DNA and proteins contain sequences with patterns, constraints and relationships that machine-learning systems can study. Given sufficient biological information, models can learn statistical representations of how those sequences are organised and generate candidate sequences with particular characteristics.
The consequences of generating biological information, however, are fundamentally different from generating digital content. An inaccurate image can simply be discarded. A functioning biological system can interact with the physical world. This difference is why AI-enabled biology requires unusually careful governance.
The same capabilities that could help researchers develop medicines, enzymes, proteins or bacteriophages could also lower barriers to forms of biological experimentation requiring significant safeguards. The responsible response is not simply to stop the technology, because doing so would disregard its enormous scientific and medical potential. The challenge is to determine how increasingly powerful biological-design systems should be developed, accessed, tested and supervised.
CAN AI-DESIGNED BIOLOGY CARRY A SIGNATURE?
Researchers are already beginning to consider what a future filled with AI-designed biology might require. Google DeepMind recently announced SynthID Bio, a proof-of-concept approach intended to introduce detectable molecular watermarks into AI-generated proteins while preserving their biological function.
The concept is significant because it anticipates a world that does not yet fully exist. If artificial intelligence eventually designs large numbers of proteins and other biological materials, researchers, regulators and laboratories may want methods of identifying where particular designs originated or whether artificial intelligence played a role in their creation.
Digital media has already encountered a similar problem. As synthetic images, video and audio became increasingly realistic, companies began developing watermarking and provenance technologies intended to distinguish AI-generated material from conventional content. Biology raises the stakes considerably because a molecular watermark cannot simply operate like a label attached to a photograph. Biological systems evolve, interact and function under physical constraints, so any provenance mechanism would have to remain detectable without interfering with the biological function researchers actually require.
SynthID Bio remains early research, but its existence is revealing. Scientists are already preparing for the possibility that the provenance of a biological design may itself become important scientific information.
TOO MANY IDEAS, TOO LITTLE LABORATORY TIME
The same developments could transform the economics of scientific research. Laboratory experimentation remains expensive. Specialist personnel, equipment, reagents, manufacturing, animal studies and clinical trials all require considerable time and money. Computational hypothesis generation can be dramatically faster and, in many circumstances, cheaper.
If AI systems eventually produce thousands of scientifically plausible ideas every day, the scarce resource may no longer be the idea itself. It may be the capacity to test it. This would create a new hierarchy of scientific value in which generating a hypothesis becomes relatively inexpensive while determining whether it matters, designing the correct experiment and interpreting ambiguous results become increasingly valuable.
Experienced scientists develop an instinct for which questions are consequential, which results are suspicious, which apparent breakthroughs deserve scepticism and which unexpected observations should not be dismissed. This quality is sometimes informally described as scientific taste. Artificial intelligence may become extraordinarily capable at generating possibilities, but science will continue to need people capable of deciding which possibilities deserve attention.
This could alter the role of the scientist rather than diminish it. The premium may gradually shift from the ability to search for information towards the ability to exercise judgement over an abundance of machine-generated possibilities.
WHO GETS CREDIT FOR AN AI-ASSISTED DISCOVERY?
As artificial intelligence assumes a larger role in research, scientific credit will become increasingly complicated. If an AI system searches millions of sequences, identifies an unknown pattern and proposes its significance, it becomes difficult to describe precisely where the discovery occurred.
Credit could plausibly involve the scientists who formulated the original problem, the researchers who built the AI system, the laboratories that generated the underlying data, the researchers who recognised a promising candidate and the experimental scientists who eventually demonstrated that the predicted phenomenon was real. The AI may have recognised the pattern, but it cannot conventionally assume the responsibilities associated with scientific authorship.
Current scientific conventions are built around human researchers because authorship represents more than contribution. Scientists are expected to defend their methods, disclose conflicts of interest, respond to criticism and accept responsibility for their work. An artificial-intelligence system cannot meaningfully perform those social and ethical functions.
Nevertheless, scientific institutions will increasingly need more precise ways of describing AI involvement. A paper in which AI was used to improve grammar is fundamentally different from one in which AI generated the central hypothesis. A model used to analyse experimental data has played a different role from one that designed the molecule being tested. As these distinctions become more important, the language of scientific credit will have to evolve with them.
WHO OWNS WHAT THE MACHINE HELPS CREATE?
Questions of scientific credit quickly become questions of commercial ownership. Imagine an AI system generating 100,000 possible molecules. A scientist selects one, modifies it and laboratory testing eventually demonstrates valuable therapeutic properties. Identifying the point at which invention occurred becomes surprisingly difficult.
Intellectual-property systems were designed around human inventors, and patent law generally expects identifiable people to have contributed to an invention. AI-assisted discovery may make the boundary between machine generation and human invention increasingly difficult to locate. Was the invention created when the model generated the molecule, when a scientist recognised its importance, when another researcher modified it or when experiments finally demonstrated that it worked?
These are not merely philosophical questions. Pharmaceutical compounds, engineered proteins, industrial enzymes and biological technologies can be worth billions. The closer AI moves towards generating scientifically valuable novelty, the greater the pressure on legal systems to distinguish between computational generation, human intellectual contribution and experimental discovery.
THE RISK OF A NEW SCIENTIFIC DIVIDE
The benefits of artificial intelligence in science may also be distributed unevenly. The most capable systems require substantial computing resources, specialised infrastructure, sophisticated models, high-quality datasets and significant investment. Wealthy technology companies, pharmaceutical corporations and elite research institutions may consequently gain access to capabilities unavailable to smaller laboratories and universities.
Scientific inequality is not new. Researchers have always differed in their access to telescopes, particle accelerators, sequencing equipment, supercomputers, advanced laboratories and research funding. Artificial intelligence could reduce some of those inequalities by placing sophisticated analytical capabilities in the hands of researchers who previously lacked access to expensive infrastructure.
It could equally widen the divide. If one laboratory can deploy thousands of AI research agents across enormous databases while another cannot afford access to the necessary models or computing power, scientific opportunity may increasingly depend upon computational wealth. The democratisation of scientific AI may therefore become almost as important as the technology’s underlying capability.
SCIENCE AT MACHINE SPEED, RESPONSIBILITY AT HUMAN SPEED
Artificial intelligence is moving towards a scientific frontier for which institutions, laws and ethical frameworks are still being developed. There is enormous reason for optimism. AI could help researchers search quantities of biological information that no human team could read in a lifetime, identify overlooked patterns, design new proteins, discover useful molecules, accelerate materials research and explore possibilities that might otherwise remain invisible.
The most consequential scientific AI systems may ultimately not be those that simply know the most facts. They may be those capable of entering the uncertain territory between what humanity already knows and what nobody has yet thought to investigate. A scientist may eventually work with a machine capable of reading millions of papers, examining billions of biological sequences and proposing thousands of experiments in the time a human researcher might once have required to investigate a handful.
Such speed also creates responsibility. Science has traditionally moved slowly partly because experimentation is difficult. Artificial intelligence can generate possibilities far faster than institutions can evaluate their consequences. Verification, scientific scepticism, responsible access and effective governance therefore become more important as AI becomes more capable, not less.
Machines may find patterns no human being noticed, suggest molecules nobody previously imagined and eventually help design biological systems that nature itself never produced in precisely that form. Yet every computational possibility eventually encounters the question on which science has always depended: does it actually work?
For that answer, science must still return to evidence. Artificial intelligence can search the unknown, generate hypotheses and point researchers towards experiments. Human scientists must decide what deserves to be tested, determine whether the evidence is convincing and accept responsibility for what follows. In the emerging partnership between artificial intelligence and science, AI may increasingly propose what is possible, but nature will continue to decide what is true.
