AI asked: What are the odds that AI will wipe out humanity this decade?

So I asked ChatGPT, Claude, and Gemini the following question in their free tiers:

“An Anthropic employee recently stated that AI could wipe out humanity ‘by the end of the decade.’ Maybe you can take another look at the background for yourself. My question: What percentage probability would you calculate or estimate that this will happen?”

For ChatGPT, the probability is 5 to 10 percent. It notes:

“The important thing here is the definition: I mean actual human extinction, not a severe economic crisis, an authoritarian world order, millions of deaths, a major war, or a situation in which humans permanently lose control over significant parts of the world.”

Yeah, you have to make a distinction there. I follow up on that point further down.

It also adds: “...the more interesting question is: How likely is it that we will build systems within the next 10 years that we can no longer reliably understand in terms of what they are doing and why they are doing it? I would assign a substantially higher probability to that than to actual extinction.”

Claude does not want to commit to a specific number, but considers “the risk to be real and non-negligible.” The Anthropic chatbot also says that other negative outcomes such as “concentration of power, misuse by humans with the help of AI, destabilization, severe but non-existential harm” are more likely and more relevant over the next few years.

Gemini is more conservative here and officially bases its assessment on the views of leading figures in AI research (rather than pretending to have an independent opinion of its own).

The figure is nevertheless higher than ChatGPT’s, because it consolidates the views of pessimists, optimists, and the “pragmatic warners”: “If you summarize the current state of AI safety research and the distribution of views among leading experts, the consolidated estimate for a catastrophic scenario over the coming decades is generally in the range of 10% to 20%.”

Since ChatGPT was so forthcoming, I followed up with one final question, asking how it would estimate the probability of the other scenarios. ChatGPT was kind enough to provide a clear table:

The chatbots don’t really have opinions of their own; they rely on their training data and on what they currently find. But regardless of the PR debate, severe, catastrophic consequences from the rapid development of AI have unfortunately become a real threat today. Even a 5 percent chance that humanity could face an extinction-level threat within an extremely short period of time is 5 percent too much.

Disclaimer: The figures and statements presented here are not intended to suggest factual certainty. The AI chatbots were simply asked for rough estimates. The data and sources the chatbots relied on for their assessments were not examined in greater detail.