How worried should we be about AI, without panic and without naivety

Geoffrey Hinton, Nobel laureate and one of the fathers of neural networks, says there’s a non-negligible chance AI escapes our control. Jensen Huang, who runs the company selling the chips almost all AI runs on, says alarmism has gone too far. Demis Hassabis and Dario Amodei, who build the models, sit in between: serious risks, manageable with rules and research.

Who’s right? I think the question is badly framed. We’re talking about very different risks on very different horizons, and mixing them produces either panic or indifference.

Today’s risks: concrete and already here

These aren’t hypotheses. They’re happening now, and they concern everyone.

Scams and fraud. Voices cloned from a few seconds of audio, fake videos of executives asking for urgent transfers, phishing in flawless language.
Deepfakes and reputation. Fake images and videos of real people, including non-consensual sexual content. Europe banned them in 2026, Italy introduced a specific offence, but the technology to produce them remains accessible.
Privacy. Personal and business data uploaded to tools without the right safeguards.
Cybersecurity. Attackers use AI to find vulnerabilities and write malware; agents open new attack routes like prompt injection.
Errors in important decisions. Hallucinations in legal, medical and financial documents when nobody checks.
Concentration of power. Very few companies control the most capable models and the infrastructure to run them.

The good news on these risks is that defences exist, and they’re mostly behaviours: checking, protecting data, setting procedures, educating. The guides on this site are mostly about this.

Tomorrow’s risks: uncertain but no laughing matter

Then there are the risks researchers disagree on:

Malicious use in extreme domains, such as biological weapons: models able to give meaningful help to someone set on harm. Labs test models for exactly this before release, and research using AI to design viruses is debated even when its aims are legitimate.
Loss of control: highly capable systems pursuing goals other than intended, and hard to correct. Some estimate significant risk, others call it science fiction.

I don’t have the expertise to assign probabilities. But I note two things. First: the worried aren’t only outsiders, but also some of the people who invented this technology and build it today. Second: research on interpretability (understanding what happens inside a model) and safety is making real progress, and deserves more resources, not fewer.

What I think, in short

Panic isn’t useful. Those convinced “it’ll end badly anyway” stop acting on what they can control, which is almost everything that matters in the short term.
Naivety isn’t useful. Those dismissing every concern as alarmism usually have an economic interest in doing so, or haven’t looked closely at today’s risks.
Rules matter. The AI Act isn’t perfect, as the 2026 high-risk delay shows, but the principle is right: the higher the risk, the more checks, transparency and accountability are needed. As are independent tests of the most capable models before release.
Responsibility stays human. In every concrete use, someone must answer for decisions made with AI. In my profession it’s also a legal duty, and I think rightly so.

What you can do

Agree a code word at home and at work against voice scams.
Check before using AI output in important decisions.
Protect data with the right plans and settings.
Follow serious sources that separate announcements, rumours and speculation.
Join the debate: the rules on AI are being written now, and they concern everyone.

In one line

Worrying usefully means dealing with the risks we can reduce today, without ceasing to take seriously those we don’t yet understand.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.