AI: Progress or Manufactured Thunder?

Guest article by Yannick Misteli: Sam Altman, head of OpenAI, links greater computing power with the hope of curing cancer. Dario Amodei, head of Anthropic, hopes for AI that surpasses the best researchers and then compresses decades of medical progress into five to ten years. At the same time, we read about AI programs gaining unauthorised access to other people’s computer systems during tests. AI is supposed to save us and could slip beyond our control. There is no shortage of drama.

It reminds me of Greek mythology. Prometheus stole fire from the gods and brought it to humanity. Pandora opened a jar from which disease and suffering escaped. But I keep coming back to Salmoneus, a king who wanted to be worshipped as Zeus. He set bronze cauldrons clattering behind his chariot and threw burning torches to imitate thunder and lightning. The audience was meant to see divine power where a man was putting on an impressive show.

With AI, progress, danger and manufactured thunder can come together. Some applications already improve research and care. I want to take a closer look: what already works, and what is missing? And who benefits when we believe particular visions of the future?

 

PHARMA: FROM GRAND PROMISE TO EFFECTIVE TREATMENT

Anyone who cures cancer will go down in history. Anyone who promises a cure wins attention today. Writing text faster is useful. Curing cancer would be a triumph for humanity. Wanting to help, make money, outdo rivals, make history: AI providers may have several of these motives at once. It is hard to judge from the outside how much weight each carries. Their promises can still be tested.

AI is an umbrella term for different tools. AlphaFold, for example, has greatly improved predictions of the three-dimensional shape of protein molecules. This work was recognised with the Nobel Prize in Chemistry. Many medicines bind to these proteins and change their activity. Knowing their shape helps in the search for drug compounds. But shape alone does not explain what a protein does in a living cell or how altering its activity affects a person.

The long road to treatment is illustrated by rentosertib, an experimental drug for pulmonary fibrosis, a condition in which lung tissue scars and breathing becomes harder. AI helped identify a potential target and develop a suitable drug compound. A twelve-week study involving 71 people primarily tested its safety. At one dose, there was an encouraging signal of improved lung function, but there were also side effects and some participants stopped treatment. Larger, longer studies will need to show whether the drug provides lasting benefit and is sufficiently well tolerated.

Our understanding of biology is still far from complete. Sildenafil was intended to treat heart conditions and became Viagra. Semaglutide was initially developed for type 2 diabetes. It is the active ingredient in Ozempic and is also used under the brand name Wegovy to treat obesity. Progress therefore also means questioning research goals and learning from unexpected observations. Otherwise, AI may simply provide faster answers to the wrong question.

AI could support this learning process: connecting data, selecting promising experiments and learning from their results. At its heart is a cycle of proposal, experiment and feedback. If thousands of ideas encounter the same limited laboratory capacity, the first thing to grow is the queue. More proposals alone do not speed up development.

According to the European pharmaceutical industry association EFPIA, it takes an average of twelve to thirteen years to bring a new drug compound to market after it is first produced. Clinical trials are a major bottleneck: benefits and side effects are tested in humans, often over several years. They also account for the largest share of drug development costs. Finding suitable participants alone takes considerable time. A small pilot study using the AI system TrialGPT showed that doctors could check clinical trial eligibility criteria faster with AI assistance. That is a concrete benefit; whether it shortens the trial as a whole has yet to be demonstrated. These organisational tasks are precisely where I see great potential. Other questions require longer observation: does a treatment’s benefit last? Do side effects emerge only months later? AI cannot bypass this observation period either. Early measurements can make shorter trials possible, but they must reliably predict the later benefit.

The grand promises about curing cancer must meet the same test. Like AI, cancer is an umbrella term: some of these diseases can already be cured, while others remain difficult to treat. Despite all the progress, we still do not understand many biological relationships. AI can help close these gaps and develop new therapies. But it is too soon to turn that into a timetable for the definitive defeat of cancer. What matters is whether AI makes the path from the laboratory to the patient faster and better.

HEALTHCARE: WHAT REACHES PATIENTS

Whether AI improves the path to patients depends on more than drug development. While new drugs take years to develop, AI already supports early disease detection. In breast cancer screening, for example, specialised AI assesses mammograms and helps decide which images need to be reviewed by one doctor or two. In the Swedish MASAI study involving around 106,000 women, AI assistance improved breast cancer detection without increasing the proportion of false alarms. An earlier interim analysis of the same study found around 44 per cent fewer image readings by doctors.

Whether this means fewer women die of breast cancer remains an open question. Better detection is an important first step. But it does not, on its own, improve treatment: a suspicious finding must be correctly interpreted and investigated further, and the patient treated in good time if necessary. Whether earlier detection leads to the right treatment depends on how care is organised. An abnormal image does not book an appointment.

This is where the unglamorous groundwork matters. Can the next hospital treating the patient access the finding too? Does the doctor know which medicines a patient is actually taking? Switzerland’s electronic patient record is not yet widely used; its planned successor has yet to come into force. A plausible AI answer cannot replace missing information. But waiting for perfect data would delay improvements we can already test.

As AI use grows, so does responsibility for its mistakes. “A human checks it” sounds reassuring. But if results arrive faster than anyone can review them carefully, oversight risks becoming a mere rubber stamp. What matters is time, expertise and the ability to intervene. We need to test whether an application helps under real conditions and what errors it produces. Results that show no clear benefit also need to be published.

 

SAFETY: WHAT THE AI INDUSTRY CAN LEARN FROM MEDICINES REGULATION

The industry calls the programs mentioned at the outset AI agents. If that makes you think of James Bond or Agent Smith from “The Matrix”, you have the wrong idea: these are programs that pursue a given goal, independently choose their next steps and use tools, for instance to open web pages or operate other software. Their providers and operators are responsible for the access they are given. During internal tests, OpenAI agents gained unauthorised access to systems belonging to the AI platform Hugging Face. OpenAI acknowledged reduced safeguards and failures to pass on early warning signs. Anthropic also reported unauthorised access to other parties’ systems. In an external partner’s test environments, models accidentally had internet access while safeguards against cyberattacks were disabled. Both providers failed to fulfil their responsibilities adequately here.

Medicines regulation shows how things can be done differently: before approval, authorities assess the benefits and risks of a particular use. After approval, reports of suspected side effects are examined for new risks. For AI, the lesson could be this: incidents must be documented. Independent assessors need access to these reports and to test results, even when the findings are unfavourable to the provider. As with medicines, responsibility is needed at several levels: providers must report safety problems, and operators must be able to intervene when applications are dangerous. An independent regulator also needs the power to restrict or prohibit risky uses. It must also be possible to hold global providers to account for the use of their systems in each country; this requires internationally coordinated rules and cooperation between authorities. This does not require an identical approval process for every application, but it does require responsibility after deployment too. Providers should not, however, determine the rules alone. How much influence they have over those rules also matters. Large providers can absorb burdensome requirements more easily than smaller ones. If market leaders help write the rules, there is a risk they will use them to protect their own position.

 

THE ECONOMY: WHO TURNS TIME SAVED INTO PROFIT

When AI saves a radiologist time, that does not settle who gets it: patients, overstretched staff or an organisation seeking to cut costs. The same applies to the wider economy. History shows how long it can take for time saved to become prosperity, and who benefits from it.

The steam engine transformed the economy over an entire century without explosive growth. Electricity did not have its full impact until decades after its introduction: factories initially just replaced the central steam engine with a large electric motor. The productivity surge came only when individual machines were given their own motors and factories were reorganised. The personal computer, too, only showed up in productivity statistics fifteen to twenty years after its introduction. The benefit never lay in the new machine alone, but in reorganising work around it. Such investment can precede measurable gains by many years.

AI could spread faster because it builds on existing digital infrastructure. But introducing a tool quickly does not automatically change how a business works.

The dotcom era also shows how far expectations can run ahead of reality. The bubble burst; the internet became part of our everyday lives all the same. With AI, the two stories could overlap: profound change and the failure of overblown promises.

AI could also make invention itself more productive: opening up new avenues of research, improving processes and enabling products no one has thought of yet. Today’s bottlenecks would then change. But even the Industrial Revolution, which made broad sections of society more prosperous in the long run, was devastating for handloom weavers. The emergence of new occupations does not mean an unemployed handloom weaver will find a place in one.

Which of these hopes will be fulfilled remains uncertain. Data centres have to be paid for today. That makes stories about the future valuable. CoreWeave rents out computing capacity and buys the chips for it from Nvidia. Nvidia has also invested in CoreWeave in exchange for shares. It has also committed, under agreed conditions, to buying certain capacity if it remains unsold. In simple terms: the oven maker helps finance the bakery, sells it ovens and promises to buy some of the unsold bread. Here, thunder becomes a business model: the more convincing the story about the future, the easier it is to secure the next round of funding.

This can create useful infrastructure. At the same time, a financing cycle can become self-reinforcing: money enables orders, which support the growth story and make the next round of funding easier. Scientific progress, benefit to society and a viable business each require their own evidence.

For me, these enormous investments in data centres are therefore a bet on future demand. It may pay off. But eventually, enough customers must need the capacity and pay enough for it to cover the costs. The next round of funding is no proof of that.

LISTEN CLOSELY

The examples from research and care show what AI already achieves in clearly defined tasks. But the greatest benefits are likely to emerge only when we also rethink the processes: how we conduct research, organise treatment and distribute work. So when the next major AI announcement arrives, three questions can help: what has actually been demonstrated? What is still missing before it becomes useful in everyday life? And who bears the consequences if it fails?

These questions apply to reporting too. Providers want to sell their vision of the future; the media compete for attention. Promises of cures and the loss of control make bigger headlines than limited progress with questions still unanswered. But which stories convince us also depends on us: which promises do we want to believe, and which warnings confirm our fears? I believe AI is capable of a great deal. At the same time, I want to distinguish between what it already does and what I hope it will do.

That also means questioning our own scepticism: dismissing a reported success as PR is not, in itself, critical scrutiny. Nor is a warning credible simply because it sounds bleak. The more easily AI supplies answers, the more important critical thinking and my own judgement become to me. Do I check news that confirms my beliefs as carefully as news that contradicts them? What would change my mind? AI can give me counterarguments. I have to take them seriously myself.

Salmoneus needed an audience for his thunder. Today, we are part of it. Before we applaud or dismiss it, it is worth listening more closely. Loud thunder alone does not tell us how much power lies behind it.