One headline says AI will change everything about how you work. The next says it could end the world. Neither tells you what is known.

We wanted a clearer picture, for ourselves and for the women who build with us. So we put together The Risk Landscape of Artificial Intelligence, a review of the evidence across six areas: how AI is developing, extinction and loss of control, the environment, misuse, privacy, and child safeguarding.

This article is the short version. The full report has all 84 references and a glossary.

Why we did this research

Transparency is one of our values. We teach founders how to use AI, so we owe you an honest account of its risks. That means no sales pitch and no scare story.

The report is agnostic on purpose. Where the evidence is strong, it says so. Where a figure is contested or rests on a single study, it says that too. Where informed people disagree, it sets out both positions and leaves the judgement to you.

The aim is simple. You should be able to decide where you sit on the comfort scale with good information in front of you.

How we graded the evidence

Not all sources deserve the same trust, so every one in the report carries a tiered grade.

A - Independent, open about its methods, with no commercial stake. For example the International Energy Agency, Stanford HAI, the FBI and court records.

B - A company or regulator reporting on itself, under audit or regulatory pressure.

C - Commercial or advocacy research, or an estimate that rests on one method. Used only where nothing better exists, and always labelled.

You will see those grades next to the figures below.

The ten key points

1. AI is getting more capable faster than we can test it.

Stanford's 2026 AI Index recorded steep gains in what AI can do. The length of task an AI agent can finish on its own has doubled roughly every seven months since 2019. Safety and reliability testing has not kept pace. Documented AI incidents rose 55% in a year, from 233 in 2024 to 362 in 2025. (Tier A)

2. Capable is not the same as reliable.

The same models that reached gold medal standard at the International Mathematical Olympiad read an analog clock correctly only about half the time. A 2026 study of fifteen models found that recent gains in capability brought only small gains in reliability. In a global study of more than 80,000 people, unreliability was the concern raised most often.

3. Extinction risk is a serious position. It is also impossible to measure.

In the largest survey of AI researchers, with 2,778 responses, the middle estimate for AI causing human extinction or similarly severe harm was 5%. The average was 16.2%. Nobody has observed such an event, so every number is a judgement, not a measurement. Most people in the same survey thought good outcomes were more likely than bad ones, and high hopes and serious concerns often sat in the same person.

4. The experts disagree about how to decide, not only about the facts.

In one tournament, trained forecasters put the chance of extinction by 2100 at 1%. Subject experts answering the same questions said 6%. The International AI Safety Report 2026, written by more than 100 experts, does not predict extinction. It says progress could slow, continue or speed up between now and 2030. One camp holds that numbers nobody can test shouldn't drive decisions. The other holds that when a failure can't be undone, you don't wait for precise odds. Both positions are coherent.

5. AI's global energy use is growing, from a small base.

Data centres used about 1.5% of the world's electricity in 2024. The International Energy Agency expects just under 3% by 2030. That growth is around 8% of the total rise in electricity demand over the same period. Energy use per AI task has been falling by at least ten times a year, a rate the agency calls unprecedented. (Tier A)

6. The environmental pressure is local, and the reporting is patchy.

Around 21% of Ireland's electricity goes to data centres. In Dublin the figure is 79%. Water use is rising too. Alphabet reported a 34% increase in 2025 (Tier B). But companies measure water and emissions so differently that their figures can't be ranked against each other. This is the most poorly measured area in the report.

7. Fraud that uses AI is happening now, at scale.

In 2025 the FBI counted fraud involving AI as its own category for the first time. It logged 22,364 complaints and US$893.3 million in losses, close to US$39,900 per complaint (Tier A). People can't reliably spot a deepfake by eye, according to commercial research (Tier C). The protection that works is a habit, not a tool: confirm payment requests through a second channel and agree a pass phrase with the people who handle your money.

8. In cyber attacks, AI has moved from adviser to operator.

Security teams now document groups linked to governments using AI agents across whole attacks. Security firm CrowdStrike reports that the average time for an attacker to spread from a first foothold to the rest of a network fell to 29 minutes in 2025. The International AI Safety Report offers a calmer reading. It finds AI mostly speeds up the preparation, and is not yet running attacks fully on its own.

9. Privacy law is being enforced, and your chats are less private than you may think.

Fines under Europe's GDPR have passed €7.1 billion. Enforcement powers over general purpose AI under the EU AI Act began on 2 August 2026. In a copyright case, a US court ordered OpenAI to hand over 20 million user conversations with identifying details removed. The lesson applies to any business: once your data sits on someone else's server, a court order can override the provider's privacy promise. (Tier A)

10. Child safeguarding is where law and technology have drifted furthest apart.

Reports to the US National Center for Missing and Exploited Children involving generative AI rose from 4,700 in 2023 to about 1.5 million in 2025 (Tier A). A flagged report does not confirm the material was made by AI, so treat this as a signal of scale, not a clean count. A US court ruling in August 2026 was far narrower than the headlines suggested. Producing or sharing this material remains a crime, and so does any imagery involving a real child. For any business, safeguarding duties do not change.

The one thing every side agrees on

Across all six areas, one finding keeps appearing. Capability, use and investment are moving much faster than measurement, governance and the response from institutions.

The researchers issuing the strongest warnings say this. So do many of the people dismissing them. It is close to the only claim in the field that nobody contests.

There is a second gap worth knowing about. Stanford found that 73% of experts believe AI will be good for work. Only 23% of the public agree. Whoever turns out to be right, that distance shapes how AI will be regulated.

What we left out

The report has limits, and we would sooner name them than have you find them.

• It does not weigh risks against benefits. The benefits are substantial and documented in many of the same sources.

• It does not cover jobs, discrimination, mental health or misinformation. Several of those are better evidenced than extinction risk.

• Fraud figures from different sources can't be added together, and most fraud is never reported. Read every loss figure as a minimum.

• Some events were days old when we compiled the report on 13 September 2026. They may have moved on.

So where does this leave you?

That is yours to decide.

You might read this and feel steadier, because the harms with real evidence behind them are ones you can act on with good habits. You might feel more cautious, because the gap between what AI can do and how well it is overseen is widening. Both are fair readings.

It helps to separate two questions. Whether to use these tools in your own business is a decision you control. What happens with AI globally is a question you don't. Mixing them up is what leaves people stuck.

We will keep the evidence current and keep telling you what it says. Where you sit on the comfort scale is up to you.