The artificial intelligence that holds the public’s attention writes essays, answers questions and holds a conversation. The artificial intelligence that is changing what scientists can actually do tends to be narrower, quieter and far less talkative. It reads charred scrolls no one can physically open, ranks a million galaxies to pull out the rare ones, and predicts the shapes of proteins faster than any laboratory bench.

What these systems share is not fluency. It is scale. Each is pointed at a body of data too large for any human team to work through by hand, and each is built to find the few things worth a closer look. The interesting part is that they tend to work best inside a pipeline where people still do the confirming.

The scrolls no one could open

When Mount Vesuvius buried Herculaneum in 79 CE, it carbonised the library of a seaside villa whose surviving papyri number more than 1,800. Many of the rolled scrolls came through as lumps of compressed charcoal, too brittle to unroll. The Vesuvius Challenge, launched in March 2023 by Nat Friedman, Daniel Gross and the computer scientist Brent Seales, set out to read them without touching them, using high-resolution X-ray scans from a synchrotron and machine-learning models trained to pick out the faint trace of carbon ink against carbonised papyrus.

In October 2023 a contestant read the first word, the Greek for purple. In February 2024 a team recovered more than 2,000 characters from one scroll, an Epicurean text probably by the philosopher Philodemus, discussing pleasure, music and food. In May 2025 the title and author of another scroll, PHerc. 172 in Oxford’s Bodleian Libraries, were identified as Philodemus’ On Vices, likely its first book, the first time the title of one of these scrolls had ever been read.

None of this is the algorithm reading on its own. The model flags where ink is likely to be, and papyrologists confirm the letters and the sense. The machine made an unreadable object legible. People are still doing the reading.

A million galaxies, ranked

The same shape of problem turns up in astronomy, where the constraint is the sheer number of objects. Strong gravitational lenses, where a foreground galaxy bends the light of something behind it, are valuable for studying dark matter and cosmology, and they are rare: fewer than a thousand had been confirmed in the whole history of the field.

When the European Space Agency released the first quick batch of data from its Euclid mission in March 2025, deep-learning models ranked about a million galaxies in a patch of sky covering less than half a per cent of the planned survey. Around 1,800 volunteer citizen scientists and 61 professional astronomers then vetted the top of the list. The result was a catalogue of 497 galaxy-galaxy strong lens candidates from about six weeks of searching, and the collaboration forecasts on the order of 100,000 once the full survey is searched.

A separate project pointed a similar tool at the Hubble archive, searching 99.6 million image cutouts and surfacing nearly 1,400 anomalous objects, more than 800 of them not previously documented in the scientific literature, reported in Astronomy & Astrophysics in December 2025. The paper lists 138 new candidate gravitational lenses, along with jellyfish galaxies and hundreds of mergers or interacting galaxies. The pattern is identical. The model sorts, the people confirm.

The Nobel went to the quiet kind

It is worth remembering which AI the scientific establishment has already singled out. The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design, and half to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold, the system that predicts a protein’s three-dimensional structure from its amino acid sequence. AlphaFold has since produced predicted structures for around 200 million proteins, close to every one researchers have catalogued.

It was one of the clearest signs yet that the scientific establishment treats narrow AI systems as discovery tools, not just software demonstrations. It did not go to a chatbot. It went to a narrow tool that solved one long-standing problem in structural biology and made its results freely available.

What they have in common

These systems are not general-purpose minds. Each is trained on a specific kind of labelled data, ink against papyrus, lensed against unlensed galaxies, known protein structures, and each does one thing across a dataset no human could finish. They are, in effect, very good filters. Their value comes from the size of the haystack, not from any understanding of what they find.

That distinction matters for how the results should be read. What these tools mostly produce is candidates. A lens candidate still needs spectroscopic follow-up before anyone is sure what is being lensed. A reconstructed scroll passage still needs a papyrologist. A predicted protein structure comes with a confidence estimate, not a guarantee, and AlphaFold’s own makers are clear that it is a prediction. The algorithm narrows the field. It does not close the question.

This is also why the comparison with chatbots can mislead. A large language model generates fluent text and can be wrong in fluent ways. A lens-finder or an ink-detector is doing something more modest and more checkable: ranking, so that limited human attention lands where it is most likely to pay off.

What to watch

The trend to follow is the widening gap between the data being collected and the people available to look at it. Euclid’s larger data releases are still to come, more Herculaneum scrolls are being scanned, and the Vera C. Rubin Observatory will soon produce image volumes that make even Euclid’s look small. In every case the working assumption is now the same: a model ranks the data first, and people examine the top of the list.

The chatbots will keep getting the headlines. The tools quietly clearing those backlogs are the ones changing what gets discovered, and they are doing it one ranked list at a time.