Column

Where is this ship sailing, and who's at the helm?

AI is starting to do science itself. Breakthroughs may arrive faster than we can understand them, and control over research may end up with a few companies.

Black-and-white close-up of Petri Lattu: a bearded man looking upward with a finger at his temple, beaded bracelets on his wrist.

Science’s biggest breakthroughs may soon arrive faster than we can understand them. At the same time, the power to decide what gets researched and who benefits from the results may concentrate in the hands of a few technology companies.

Over the past few weeks, something far deeper than yet another leap in capability has started to take shape in AI. OpenAI’s mathematical research results, Google DeepMind’s infrastructure for biological research and Anthropic’s automation of experimental research all point in the same direction.

AI no longer just helps researchers do science.

OpenAI’s maths grenade this week alone, in which the company released hundreds of mathematical research papers in one go, is in its own way the biggest news of the entire AI era.

It’s funny, then, that it isn’t making more headlines.

So the new models are starting to form systems in which working through research questions, forming hypotheses, searching for solutions and verifying results can be linked into an increasingly autonomous whole.

In its own way, this is also about the language of mathematics, for example, changing. If we can simply ask new models to solve things instead of working out among ourselves how things get solved, much of what we think of as the scientific process gets rewritten.

The most important question of the AI era may not be what the machine can solve, but what is left for humans to apply.

This may be the most significant of the AI era’s upheavals so far. Not because the machine can finally do something better than a human.

But because the speed of scientific progress may start to come loose from the speed of human thought.

When solving becomes an industry

Mathematics has been one of the most fascinating proving grounds for AI. A mathematical proof is an unusual object of research, because in certain cases its correctness can be checked formally. That makes a new division of labour possible.

AI can explore an enormous number of alternative routes to a solution, construct proofs and use formal systems to check them. Humans no longer need to take part in every intermediate step of the reasoning.

The results presented in OpenAI’s recent mathematics project are an exceptionally ambitious example. The project has reported hundreds of solutions to mathematical problems that had long remained open.

Independent evaluation of the claims is, of course, essential. Publishing a mathematical result or checking it by machine does not yet mean the scientific community has accepted all of its conclusions. But then again, does it need to? That is a hot potato that calls for thick oven gloves.

What happens if solving mathematical problems turns from the intellectual achievement of a narrow few into a scalable production process?

Then the bottleneck may no longer be finding the solution. It may be understanding why the solution matters.

Exactly this friction has come up in the reflections on AI by mathematician Terence Tao and the Association for Human Mathematics group: a mathematical result’s correctness, its comprehensibility and its significance are not the same thing. We can imagine a future in which machines produce new mathematical structures faster than humans can develop the concepts needed to understand them.

Then we are no longer talking about whether AI can solve some known problem.

We are talking about whether science can start to develop in a language its own researchers do not yet command.

And if that happens, the human role changes.

A lab that never sleeps, however much the waves rock it

Mathematics is only the beginning, though.

In biology and medicine the same trend is even more significant, because a computational proof alone is not enough. A hypothesis has to meet physical reality.

The protein really has to fold. The molecule really has to work. The cell really has to respond. DeepMind’s protein research and new infrastructures for AI-driven biological research show how computational modelling can be combined with experimental research.

Anthropic, for its part, has developed automated biological research environments in which AI takes part in designing experiments and robotics carries them out.

This creates a new research loop.

AI forms a hypothesis. An experimental system tests it. The results feed back to the model. The next hypothesis grows out of the previous experiment.

And the cycle starts again.

In traditional research this process is limited by researchers’ time, lab capacity, funding and organisations’ ability to coordinate work. Perhaps even by human stamina and imagination.

I recently talked with a friend who works in fertility technology. They described an interesting problem in sperm selection: a person working at a microscope may pay particular attention to cells that move actively, even though movement alone does not tell you how suitable they are for fertilisation.

AI could look at the selection based on different characteristics and possibly find good candidates a human would miss.

In an automated system, some of these limits turn into technical and economic questions. How much compute is available? How many experiments can run in parallel? How quickly can results be measured and fed back into the system?

This opens up the possibility of an entirely new scale of scientific research.

But it is also right to ask what happens when scientific breakthroughs can be produced at industrial speed, while understanding, evaluating and debating them as a society still moves at human speed?

Whose science is AI doing?

Until now, scientific progress has rested on humans’ ability to understand the world and build new solutions on that understanding. Scientific knowledge has accumulated slowly, often over generations, and research methods have evolved together with the people who use them. We have believed in open science, and universities have played a core role in basic research – and why not in applied research too.

Tomorrow may change that.

If a research system can generate thousands of hypotheses, test them and build on the most promising results, it may not need a human to choose every new research direction. That holds a huge opportunity. We could discover new medicines, materials, energy solutions and mathematical methods at a speed that used to be impossible.

But at the same time, a new kind of concentration of power emerges.

Who owns these tools?

Imagine a technology company whose AI system can develop a new treatment for a previously incurable disease. The company has invested billions in compute, research and labs. It is reasonable for it to get a return on those investments.

Or, even more deliciously: imagine an international cosmetics giant that brings a planeload of money to a frontier lab and asks it to develop a peptide that demonstrably slows skin ageing.

Does it get exclusive rights to its product? Does the frontier lab charge licence fees for it? What would the market value of that product be?

And what if the same system could solve dozens of other socially significant problems, but researching them does not offer an equally attractive commercial return?

The privatisation of science does not just mean patenting research results. It can also mean privatising the choice of research questions.

The goals of scientific research have traditionally been steered by many parties: universities, public research institutes, companies, foundations and researchers themselves. Their interests have not always aligned, but the system has had many mutually independent actors.

In future, a small group of companies may control research capacity that entire states, or companies competing against one another, cannot access.

The most important task for AI labs over the coming years is to monetise their products. It is easy enough to imagine that one big lever for that will be offering application-based capacity to the highest bidders. And fair enough, but what’s interesting is where we start. Is a peptide for smoothing skin worth more money than, say, developing a new kind of malaria vaccine?

The wildest part: humans can no longer keep up with their own thinking

The acceleration of scientific progress raises another question that has received fairly little attention so far.

What happens to people?

We are used to thinking of AI as a tool that makes human work more efficient. It helps us write, analyse, plan and make decisions. But more and more often, AI does not merely carry out human ideas. It takes part in forming them and even creates something new. With future models, it may even turn out that keeping a human in the loop slows things down and makes the results worse.

This is the core of co-thinking.

AI can help us look at a problem from angles we would not have come up with ourselves. It can challenge our assumptions, sketch alternative futures and bring entire new conceptual structures into our thinking. At its best, this is a vast extension of human capability.

But co-thinking requires more than an intelligent machine.

It requires a person who can operate in a new kind of thinking environment. Who understands where their own role and limits lie, and where their own superpowers lie.

Imagine a researcher whose AI partner produces more new research paths in a single afternoon than their research group could previously have handled in a year. So the researcher’s problem is no longer a lack of possibilities.

It is that there is no end to them.

What to grab hold of? What to consider meaningful? When to stop? How do you know you have really understood something, rather than just accepted the machine’s convincing explanation? What if your own biases end a promising line of research because it is already Thursday afternoon and you’re tired?

The same phenomenon already shows up in ordinary laptop work, of course. AI can grow a person’s production capacity much faster than their ability to evaluate, lead and understand their own work grows.

We can do more than ever, but we cannot necessarily process everything we do. And this asymmetry is not only about knowledge.

It is also psychological.

The human nervous system does not scale with compute.

An unlimited number of possibilities does not automatically mean an unlimited capacity to act. It can also mean a constant sense of unfinishedness, decision fatigue and a feeling that your own thinking is no longer enough. Or a backlash against the whole subject: one reason for the steadily growing anti-AI sentiment is probably invisible burnout and the feeling that someone else is steering this boat.

And at a deeper level still: AI can produce endless new ideas, but humans still have to live with their consequences.

So we need a new model for the human–AI boat trip

As AI’s capabilities grow at tremendous speed, we also need to think about letting go, even as we enjoy everything AI adds.

That means the ability to think with AI without losing our own judgement, the ability to evaluate the possibilities the machine produces, and the ability to recognise when the speed of thinking starts to outrun our own understanding.

Increasingly, it also means the ability to understand when to get out of the way. This is a painful and emotionally charged issue, and thinking it through is one of the most important tasks for all of us in the years ahead.

If the machine can solve a problem better than we can, why on earth would we insist on a human messing up the process? But if we also let the machine decide which problems are worth solving, we have handed over something far more essential than a shortcut through a single work step.

In the age of AI, we may for the first time have to seriously separate intelligence, understanding and wisdom. They are not the same thing.

The final canoe (I’m stretching this metaphor far too far, I know)

The answer to this development is, of course, not to stop AI research or slow down scientific breakthroughs just because they challenge existing institutions.

Quite the opposite.

If AI can help solve problems in disease, energy production or materials science, the opportunity is too great to leave unused. And why shouldn’t that skin-ageing product be valuable too.

But neither can we hand over our future research agenda to the commercial goals of a few technology companies. Nor our own thinking to ever-evolving models.

As AI develops under full sail, shouldn’t we track this with equal weight: how fast can humans change?


The writer thinks they’re clever because they don’t own a boat. As a teenager, though, they once travelled thanks to their father’s connections on a cargo ship from Helsinki to England, where they learned that sailors really do have their own way of doing things. Those memories have been sealed in a tightly locked chest deep in the Baltic Sea.

Petri Lattu

Innovation director and AI strategist who writes about AI's impact on society and about cognitive sovereignty: how people can think with AI without handing their thinking over to it.

More from this writer →

Read next