Climate change and artificial intelligence pose very different threats, but uncertainty about the scale of catastrophe is no excuse for abandoning precaution
By
Around ten years ago, I happened to attend one of the earliest meetings of the Future of Life Institute – expressly founded to prevent humanity’s extinction – held at a private home near Boston. Its focus was AI-only.
In front of a couple of Nobel laureates, several MIT and Harvard professors and a score of young scientists, I dared to suggest that climate change was a much bigger risk. They dismissed my take.
In the meantime, the climatic menace has advanced, say, three steps forward. Following the 2022 breakthrough of large language model chatbots, AI’s development has leapt 100, or maybe more. Were they right?
With the recent public scare, p(doom) – the estimated probability of human extinction caused by AI – has been popularised. Max Tegmark (the Future of Life Institute’s chair) puts his p(doom) at over 90 per cent – or more than 0.9 in statistical notation.
In 2024, Nobel Prize winner Geoffrey Hinton estimated the risk of extinction within 30 years at between 0.1 and 0.2. Nvidia’s Jensen Huang puts the chance of AI destroying the world by 2030 at zero – which is too optimistic. These estimates involve different assumptions and timeframes, so they aren’t directly comparable, but even Hinton’s lower number is frightening enough.
Would you board an aeroplane that has a ten per cent chance of crashing?
Climatologists don’t gamble with figures. Based on policies currently enacted, a median warming of around 2.7°C is projected by 2100, a threshold that makes severe sea-level rise, widespread ecosystem collapse and food-chain disruptions – in the prudent forecast of the IPCC – ‘very likely’. Terrible consequences are expected well before the century’s end.
The climate crisis began a long time ago and, while it makes headlines – as in the case of the dreadful flash flood in Nepal last August – it remains a slow-moving physical phenomenon that guarantees a high p(disaster) value in a rather distant future.
By contrast, AI has been improving by the week, and both tech CEOs and whistleblowers claim that, with recursive self-improvement (RSI), algorithmic models are approaching superintelligence and thus the prospect of outcompeting humankind.
RSI is the process by which an AI system modifies its own code and architecture, and uses its enhanced intelligence to make itself exponentially better. However, while AI companies are tinkering with automated machine learning to optimise their models, at present, no model is autonomously redesigning its core architecture and rewriting its neural weights.
It shouldn’t be forgotten that the rogue hacking events everyone is talking about were underpinned by a deliberate removal of behavioural guardrails and inadequate ‘sandboxing’. In other words, humans were well and truly in the loop. It’s true that AI and the climate are separate and incomparable risks, but there’s something that links the two: the precautionary principle. This is a risk-management concept used in public policy, science and ethics: if an action, scheme or technology is likely to cause severe, irreversible harm, the burden of proof falls on those proposing it.
Enjoying this article? Check out our related reads…
Climate change is a cumulative and well-documented crisis. The harm is already occurring and compounding – researchers project that the current super El Niño could cause around 451,000 additional heat-related deaths between June 2026 and February 2027 (deaths from associated floods, droughts and wildfires are excluded from the count).
AI risks are largely probabilistic and speculative, yet moving at an exponential speed. Many argue that, once a system achieves RSI, it may be impossible to ‘uninvent’ or shut it down.
In a New York Times interview, Huang said that OpenAI, Anthropic, Google and SpaceX shouldn’t release models deemed unsafe, and that they will carry the burden of responsibility under current laws and regulations. Yet fossil fuel companies (well aware of the severe effects of carbon pollution since the 1970s) are barely held accountable.
Obviously, applying precautionary measures is hard. Halting emissions that cross dangerous warming thresholds clashes with immediate economic returns. Slowing down the scaling of AI capability until safety is guaranteed clashes with an intense geopolitical and corporate race. However, forgetting the toll of our carbon emissions on future generations is patently shameful; deploying recursive self-improvement in AI models without safeguards may be plainly criminal – no matter which of these two self-inflicted risks is closer, or nastier.
