This week, there was no escaping it. A researcher who left Anthropic says the people building AI earnestly believe it could kill us all. His posts got tens of millions of views. Time, CNN, El País: everyone picked it up.

I build AI tools for small businesses. So, as you can imagine, people have asked me about it several times since. And I think it's a fair question. The risk is discussed very seriously by the people building these systems, and that deserves to be heard.

But I noticed something. Nobody asks the opposite question with the same seriousness. Not “AI will save the world”, the ad version. Rather: if we apply exactly the same honest reasoning, could AI save us?

I gave it a try.

The serious rescue isn't a movie either

The serious disaster scenario isn't Terminator. It's a highly capable system pursuing goals slightly different from the ones we intended.

The serious rescue scenario isn't a super-brain that fixes the climate over a weekend either. It's far less spectacular. It's concrete, often invisible gains that add up: a cancer caught earlier, a flood announced days ahead, a molecule found in weeks instead of years.

What is already saving lives

No promises here, only things that exist.

  • Proteins. AlphaFold, developed by Google DeepMind, predicted the structure of hundreds of millions of proteins. That work earned its creators the 2024 Nobel Prize in Chemistry. Researchers worldwide use it for vaccines, drugs and enzymes.
  • Breast cancer screening. In Sweden, the MASAI trial compared standard screening with AI-supported screening across more than 100,000 women. First results, published in 2023: about 20% more cancers detected, with 44% less reading work for radiologists. Radiologists still make the call. The AI shows them where to look.
  • Antibiotics. At MIT, researchers used AI to identify new molecules that work against drug-resistant bacteria, including a dangerous hospital superbug. These are bacteria we have almost nothing left against.
  • Floods. Google publishes river flood forecasts up to 7 days ahead in more than 100 countries, many of which had no warning system at all.
  • Wildfires. In California, more than 1,200 AI-equipped cameras watch the forests. They have spotted more than 900 fires before the first emergency call. And every alert is confirmed by a human before firefighters are sent.

Reasons to believe

  • AI speeds up science. Testing millions of leads before handing ten to a lab changes the pace of research.
  • It multiplies scarce experts. Radiologists, dermatologists and eye doctors are in short supply everywhere. A tool that saves them time means more patients seen.
  • It sees what we can't. Faint signals in data far too large for a human eye.

Reasons to keep a cool head

  • “Could save us” doesn't mean “will save us”. Just as “could kill us” doesn't mean “will kill us”. Many AI promises are marketing, and announced timelines are often wrong.
  • A tool doesn't fix root causes. Take the wildfires in the forests of Gascony, in south-west France, which hit close to home for me. A smart camera spots a fire earlier. It changes nothing about the fact that we planted a million hectares of the same tree on dry sand. That is a human, political and economic decision.
  • AI has never put out a fire. A firefighter does. An alert is useless if nobody can respond.
  • Benefits don't reach the right places on their own. A brilliant medical tool in a big hospital saves no one in a village without a doctor.

The real issue: it's the same question

While writing this, I realised something. “Could AI kill us?” and “could AI save us?” are not two opposite questions. They are the same one.

In both cases, the answer depends less on the technology than on three very human things:

  1. Who decides what gets built, and for whom.
  2. Who controls what the tool does, and can stop it.
  3. Who is accountable when things go wrong.

The researcher who resigned isn't saying the technology is bad in itself. He's saying companies are locked in a race, and nobody can slow down alone. The problem he describes isn't technical. It's a governance problem.

And it's exactly the same for the rescue side. In every positive example in this article, a human stays in charge. The radiologist validates. The researcher tests in the lab. The lookout confirms the alert, the firefighter responds. AI prepares, humans decide.

What I take from it in my own work

At my scale, I'm not building superintelligence. I build custom AI automation tools for businesses: sorting emails, preparing quotes, finding information across hundreds of documents.

And it all starts with clean data. That's the topic of my article on businesses that run on Excel files.

But the rule I hold myself to is the same one that makes the good examples above work. The repetitive work is automated, the decisions stay with you. Not everything can be automated, and I tell you when that's the case.

It isn't a cautious pose. It's what makes a tool genuinely useful instead of scary.

AI won't kill us on its own, and it won't save us on its own. The difference between the two is made by humans. That sounds like good news to me.

Sources