A Discovery Is a Beginning

The most important word in a scientific announcement can be the one that seems to spoil the celebration.

Unknown.

On September 23, 2026, Anthropic reported that Claude had helped identify an enzyme system with features resembling those of CRISPR-related systems. The underlying reverse transcriptase had been noticed in earlier research; the company said the newly recognized features included an associated repeat array and an additional protein. The system’s function remained unknown. Human scientists directed the search and performed the laboratory work.¹

That is a finding worth being interested in. It is also a finding whose value is easily distorted.

One reader may see the reference to CRISPR and imagine a new treatment. Another may see that the announcement comes from an AI company and dismiss the work as advertising. Neither reaction establishes what the researchers have found. The first leaps past work still to be done. The second risks overlooking a promising result because of who reported it.

The proper response is curiosity with standards.

We can ask whether the observation holds up, whether independent researchers can inspect the evidence, and what further study reveals. We can welcome a new way to notice something while refusing to describe an unknown function as a demonstrated medical use. We can give the result its present name without denying it a future.

The discovery does not need a borrowed cure to matter.

Finding a better question

Popular accounts often make discovery look like a moment in which the answer suddenly appears. Much useful work happens before that moment. Someone has to find a question worth the effort of answering.

Imagine a vast shelf of records. Most entries look ordinary. Some appear different because the measurement is poor. A few differ in ways that could change what we understand. The challenge is not simply to locate an odd entry. It is to decide which kind of oddness deserves attention.

If a tool helps researchers search more records, compare patterns or connect findings from separate fields, it can change that early work. A promising use of AI is therefore not limited to producing final answers. It can help enlarge the set of questions that people are able to investigate.

But a larger set is not automatically a better set. A machine that produces a thousand attractive guesses can create a thousand new obligations to check. If those guesses are weak, the apparent productivity is a transfer of work to the people who must reject them. If they are strong, a scarce research team may gain a better choice of where to spend its time.

The relevant gain is the value of what survives careful checking, compared with the work required to obtain it.

This changes what we should want to know about a scientific system. How many suggestions were generated? How many were investigated? Which survived? How much human work was needed? What would the researchers have achieved with their existing tools? A striking success is evidence that something was possible. A record of the whole search is better evidence about how often it may be possible again.

There is a public interest in that fuller account. If only the winners are shown, a reader cannot see the cost of finding them. Nor can another research team judge whether it could use the method with its own budget and skills. A powerful demonstration may depend on support that never appears in the photograph.

To ask about that support is not to take credit away from the machine. It is to understand the machine’s contribution well enough to use it.

Science already has several forms of assistance. An instrument can detect what the eye cannot see. A database can preserve what no one person could remember. A mathematical method can reveal a pattern hidden in a mass of measurements. AI may add new forms of assistance, and some may be unusually broad. The worthwhile question is how they improve inquiry, rather than whether the final discovery can be assigned to a single heroic actor.

The language of a machine discovering something on its own can obscure the people who built the records, chose the question, supplied the tools and checked the result. It can also obscure a real advance in the amount of searching or reasoning the system performed without step-by-step direction. Both errors make evaluation harder. We should be able to describe the contribution precisely enough to recognize it.

That is especially important when the provider benefits from the story. A company has good reason to show what its technology can do. It also has an interest in how the demonstration affects confidence in its products. An interested source can supply true information. Its interests give us a reason to seek independent checks, rather than a reason to assume either truth or falsehood.

The shape and the life

AlphaFold offers a separate example of an AI capability that was tested against a demanding scientific task. In the CASP14 assessment, its predictions of protein structures achieved accuracy competitive with experimental structures in many cases, as reported in the 2021 Nature paper by John Jumper and colleagues. The assessment compared predictions with structures whose answers were not available to the predictors in advance.²

That matters. A test in which the answer is held back asks more of a system than a display of selected successes.

A protein’s shape can help scientists reason about its role. But a prediction of shape is a particular kind of information. It does not by itself establish every function of the protein, the safety of a possible treatment or the health outcome for a patient. EMBL-EBI’s training materials explain that structure predictions have limits, including their incomplete account of the changing behavior of molecules.³

Those limits do not erase the advance. They tell researchers how to use it.

There is a childish version of praise in which a tool must be able to do everything. If it has limits, someone has apparently exposed a fraud. Serious praise can be more exact. A tool may remove a difficult step, improve a weak one or make a question affordable to study. That is enough to change a field even while other work remains necessary.

Think of the difference between receiving a map and arriving at a place. A map can save enormous effort. It can show that the route you planned is impossible. It can reveal a route you had not considered. Its value does not depend on pretending that the journey has already happened.

The same discipline applies to medical hopes. A scientific result may suggest a direction worth pursuing. Later evidence may show that the idea works only in limited conditions, has unacceptable effects, or cannot be delivered at a cost people can bear. Sometimes the work opens another question instead of producing the use first imagined.

There is no honest timetable that can be inferred from the word breakthrough alone.

This is painful because the public does not encounter research only as a series of intellectual puzzles. Someone is waiting for help. A family hears a headline while living with a condition for which the available choices are inadequate. It is natural to want the steps between discovery and treatment to be shorter.

That desire should motivate better work. It should not become a license to announce that the steps have vanished.

People waiting for care are owed ambition and accuracy together. They should not be used as emotional proof that every proposed release is urgent. Nor should their need disappear from an account of risk that considers only what a new tool might do wrong. Their position is a reason to investigate how real help can arrive sooner, with evidence strong enough to deserve their trust.

Learning the languages of change

There is a connection here that excites me. A person writes code and a machine begins to do something it could not do before. A change in instructions becomes a change in the world: a record can be found, an image can be formed, a calculation can be completed.

What happens as we become better at understanding other kinds of instructions and structures?

DNA carries inherited information. Proteins take shapes that help determine what they can do. Materials have structures that affect how they behave. The comparison with software invites a powerful thought: some conditions we experience as fixed may contain possibilities we have not yet learned how to use.

A living cell answers through chemistry, structure and its changing surroundings. To carry the programmer’s ambition into biology, we have to learn that world on its own terms. Our models must answer to experiments.

That requirement makes the possibility more interesting. We are learning to connect representations with consequences, and to test whether those connections hold.

Google DeepMind’s AlphaEvolve provides one example in computing. Its researchers described a system that proposes changes to code and uses automated evaluation to seek better algorithms. The reported work includes mathematical and computing tasks. The combination of proposed ideas and checks is part of the method; a plausible piece of code alone is not the achievement.⁴

In materials research, the GNoME work reported in Nature in 2023 used machine learning and further calculations to identify candidate crystal structures predicted to be stable. That computational result enlarged a field of possibilities for research. It did not mean that every proposed material had been made, tested for a practical use or produced at an affordable scale.⁵

Software, proteins and crystals are different subjects. What connects these examples is a growing ability to search among possibilities, make a prediction and confront it with a test appropriate to the domain.

Imagine where that habit could take us. Better ways to store energy. Better tools for studying a disease. Materials that use fewer scarce resources. New means of restoring a lost ability. These are aims for inquiry, not a list of achievements we can announce in advance. Yet they are large enough to deserve courage.

I do not want us to look at that horizon and make shrinking our ambition the first expression of our ethics. The ethical task is to turn greater understanding toward the reduction of suffering, and to organize the work so that more people share the result.

There is an old temptation to dream of possessing the power of a god. The better question is what we would do with the power. If we could alter more of the conditions that injure and limit human life, would we use that ability to serve, or to place others beneath us? Would a discovery become a remedy, a weapon, a toll or a closely guarded privilege?

The answer will depend on institutions and choices as much as on intelligence. That is why the argument for development and the argument for responsibility belong together. We should become more capable of changing the world, and more exacting about what our changes are for.

A knife can prepare a meal or injure a person. The comparison is simple, and AI can create risks far less familiar than a knife’s. Still, the ordinary example exposes a moral mistake: the possibility of harm does not exhaust the purpose of a tool. We must govern the harmful use while developing the good we can do with it.

The larger our powers become, the less excuse we have for treating preventable helplessness as the natural order of things.

Where the queue moves

A simple illustration shows why a dramatic improvement in one stage can produce a smaller improvement in the whole process.

Suppose a piece of work takes ten weeks. Four weeks are spent on an early analysis and six on other tasks that must follow it. A new tool reduces the analysis from four weeks to one. That is a 75 percent reduction in the time spent on the early analysis. The whole job now takes seven weeks, a 30 percent reduction in total time.

Illustrative time savings: a faster first stage can shorten the whole process by a smaller proportion. These quantities are invented and the stages occur in sequence.

These are invented numbers, and the stages are assumed to occur one after another. Actual research can loop, overlap, fail and begin again. The example explains arithmetic, not the expected speed of a scientific field.

Even so, it reveals a mistake worth avoiding. A claim about one part of a process cannot be applied unchanged to the entire process.

The gain may still be large. Three weeks can matter. The right response is to keep the gain in its proper place and examine the other six weeks. Are they spent obtaining a scarce input? Waiting for equipment? Checking a result? Meeting a necessary standard? Repeating work because records cannot be shared?

Different causes call for different action.

Buying a faster analytical tool will not fix an institution that cannot recruit the staff needed for the next stage. Removing an essential check may shorten the schedule while destroying the reason to trust the outcome. A slow approval process may contain waste, a necessary safeguard, or both. To improve it, someone has to find out which.

The easy political story says that invention is fast and institutions are slow. Sometimes the institution is the obstacle. Sometimes it is the place where a claim finally meets a test it cannot pass. Treating every delay as obstruction rewards the person whose promise is least examined.

There is another possibility. Better AI may help with the checking itself. It may make records easier to examine, reveal errors or reduce routine work around a study. That possibility deserves investigation on its own evidence. We should not assume that progress applies only to producing claims while the methods for testing them remain fixed forever.

This is a more constructive ambition than demanding that everyone simply lower their standards. Improve the path to trustworthy knowledge. Improve the equipment, the staffing and the exchange of results. Measure where time is lost. Where a requirement adds little protection, change it. Where a check exposes serious failure, learn from it.

An experiment that disproves a beautiful idea has completed important work. Its failure to produce the hoped-for product is not a failure to serve the public.

The less visible workers in this process deserve attention too. A system needs maintained records, working equipment, skilled operators and people who can interpret an unexpected result. If new tools increase the demand for these roles, their absence becomes more costly. Money spent on the most celebrated stage may produce less benefit than money spent on the stage where work is now waiting.

Progress can depend on the person who repairs the instrument.

When the help reaches a person

There are cases in which AI-assisted research has moved beyond a predicted possibility to a demonstrated human use. They should be described with the same care.

A 2024 report in the New England Journal of Medicine described an implanted speech system used by one participant with amyotrophic lateral sclerosis, or ALS, and severe difficulty speaking. Machine-learning methods helped translate recorded brain activity into words and speech output. The study demonstrated a meaningful form of assisted communication for that person.⁶

It did not establish a cure for ALS. It did not show that a general chatbot could offer the same help. It did not prove that the system would work equally well for everyone who might need it. The study involved a specialized implanted system and a research team.

Within that boundary, the result remains deeply important. Communication is an ordinary part of human life whose loss can be profound. Restoring an ability to express a thought is a kind of progress that does not need an economy-wide statistic to make it worth pursuing.

The next questions concern reliability over time, who else can benefit, the demands placed on users and caregivers, and the support needed outside the research setting. These questions extend the ambition. They ask how an achievement can become something people can actually count on.

A technology can be technically available while its benefit remains practically unavailable. A specialist may be too far away. Maintenance may be uncertain. Payment may be unresolved. A device may depend on a company that can change its priorities. A person may be able to try a system but have little assurance that support will continue.

These are possible barriers to examine, not findings about the particular speech study. They show why access belongs inside the meaning of progress from the beginning. If delivery is considered only after the invention is finished, important design choices may already have been made without the people who will live with them.

We should ask who can maintain the tool, what skills the user needs and how the system will be repaired before dependence becomes unavoidable. Where a use requires a lasting relationship, the promise should include more than an impressive first day.

The future has a maintenance budget.

That phrase may seem less exciting than a promise of abundance. Yet an ability that can be relied on is more valuable than one that disappears when a demonstration ends. Reliability allows a person to make plans around the help. It changes possibility into practical freedom.

Enough honesty for hope

Some people fear that careful language will drain science of its wonder. I do not find that to be true. The honest account often reveals more to admire.

A discovery with an unknown function tells us that nature still contains possibilities we have not understood. A strong prediction shows what people can achieve when they build on records gathered by many others. A communication system shows that a difficult barrier in one person’s life can become a subject of patient, inventive work.

None of these stories requires a claim that every remaining problem is about to disappear.

What they require is a willingness to continue. Continue the research. Continue the checking. Continue the work of making the result usable, affordable and worth trusting. Continue asking whose needs are receiving attention and whose are left for later.

That last choice is easy to hide behind technical difficulty. A field may pursue what is commercially rewarding, what is easy to measure, what fits its available tools or what attracts prestige. Those incentives do not always point toward the most urgent unmet need. Public support can help widen the set of questions worth pursuing. So can giving affected people a part in deciding what success should mean.

This is the pro-science position I want us to adopt. Invest more. Investigate more. Make the checking faster and stronger. Expand the number of people able to contribute, and bring demonstrated help within reach. It can object to pointless delay without treating every protective rule as pointless. It can defend a promising inquiry before it has a commercial application, while insisting that a commercial claim wait for the evidence it needs.

That position refuses an odd bargain often offered to the public: believe too much today so that we can build something better tomorrow.

We can fund the question without pretending to possess the answer.

The enzyme system in the September announcement may turn out to have important uses. It may teach scientists something valuable without becoming a widely used tool. Its place in the history of science is not yet known. The proper next pages will be written by investigation.

For now, the beginning is enough to take seriously. It is enough to justify curiosity and further work. It is not enough to promise a cure. The responsibility is to move the work forward until we know what this possibility can become.

The workshop

The first thing on the bench was only light,
A pattern waiting for a hand to read;
Beyond the glass, the city held its night.

We asked the pattern what a life might need.
It answered nothing. We began to learn
Which questions made a test, which made a creed.

A thousand paths appeared at every turn.
Some led to nothing. Some returned us wrong.
The world, not praise, decided what we’d learn.

The work required a patience fierce and long,
A way to hear what matter would refuse,
And change the claim that first had sounded strong.

Yet who would call such labor something to lose?
A child beyond the window still had pain.
The dark had never asked what we would choose.

We chose to read, to test, to read again,
To carry questions farther into day,
And bring a little shelter through the rain

To those still waiting where the shadows lay.
The power on the bench was still a seed.
We gave it work, and taught our hands the way.

Let knowledge grow in answer to the need.

Sources and notes