Roughly 72% of AI researchers believe the technology poses serious long-term risks to humanity — and yet the industry is projected to spend over $300 billion on AI development in 2026 alone. That contradiction is not a bug. It is the entire story. Readers writing to The Guardian put it plainly: we can run ethics panels, publish white papers, hold Senate hearings — and nothing actually slows down. Google DeepMind sits at the center of that contradiction more than almost any other organization on the planet right now.
Why does Google DeepMind keep pushing forward if the risks are real?
Here is the honest answer: because stopping unilaterally would just hand the lead to someone else. DeepMind’s researchers are not oblivious. These are some of the most technically literate people alive. Demis Hassabis has spoken openly about existential risk. The organization publishes safety research. It runs alignment teams. And then it ships increasingly powerful models on an accelerating schedule anyway.
That is not hypocrisy in the traditional sense. It is a collective action problem dressed up as corporate strategy. If DeepMind pauses, OpenAI does not. If OpenAI pauses, xAI does not. If every Western lab pauses, Chinese state-backed labs certainly do not. The logic of the race forces participation even among those who are most afraid of where the race ends. Google DeepMind is building dangerous technology because the alternative, in the current geopolitical and commercial environment, is simply losing.
The ethics debate is real. The structural incentives are just stronger.
What has Google DeepMind actually released in 2026 that we should be watching?
Gemini Ultra’s latest iteration now performs at or above human expert level on a majority of professional benchmarks — medicine, law, advanced coding, scientific reasoning. AlphaFold’s successors are being used to design novel proteins for drug discovery at a pace that would have taken decades of lab work. DeepMind’s robotics division is producing systems that can generalize physical tasks across environments they have never seen before.
Each of these is genuinely useful. Cancer researchers are finding targets faster. Rare disease patients may see treatments in years rather than lifetimes. That is not spin — those outcomes are real and they matter. But the same underlying capability that finds a cancer target can optimize disinformation at scale, automate surveillance systems, or be embedded in autonomous weapons. The molecule and the weapon are downstream of the same model.
Google DeepMind’s safety team is working on the alignment problem in good faith. The commercial division is shipping products on a quarterly cycle. Both things are true simultaneously, and the commercial cycle is winning.
Is this actually different from any other powerful technology humanity has built?
People reach for nuclear analogies constantly, and there is something to it — but the nuclear parallel breaks down fast. Nuclear weapons required nation-state resources, physical infrastructure, rare materials, and years of development time. A sufficiently capable AI model requires a data center, a large but not astronomical budget, and a team of engineers. The barrier to entry drops every eighteen months. The knowledge proliferates. You cannot un-invent the transformer architecture.
The more honest parallel might be social media — a technology that every serious researcher flagged as psychologically and socially dangerous, that generated voluminous internal documentation confirming those harms, and that nonetheless scaled to billions of users because the growth metrics were extraordinary and the harms were diffuse. We already know election officials are bracing for AI-driven cyberattacks as a baseline expectation now, not a fringe scenario. That is not preparation for a hypothetical. That is managing a current reality.
What separates AI from social media is the speed and the depth. Social media changed how people communicate and perceive reality. AI changes who — or what — makes decisions.
What would actually changing course look like, and why isn’t it happening?
Genuine course correction would require binding international agreements with real enforcement mechanisms, similar in ambition to nuclear non-proliferation but far harder to verify. It would require compute governance — controls on who can train frontier models and with what hardware. It would require liability frameworks that make labs financially responsible for downstream harms, which right now they almost entirely are not.
None of that is politically close. The US government has largely chosen a posture of competitive acceleration. The EU has passed the AI Act, which is meaningful on paper and early in its enforcement phase. China has implemented AI regulations that are primarily focused on controlling political speech rather than limiting capability development. There is no global body with teeth.
The economic pressure compounds the political paralysis. Automation is already threatening 69% of jobs in India according to World Bank data — and that displacement creates its own political volatility, which in turn pushes governments toward AI investment as a supposed economic hedge. The same technology causing disruption is being sold as the solution to the disruption it causes. That is a loop that is very difficult to exit.
Google DeepMind will publish another safety paper this quarter. It will also ship another model. Demis Hassabis will give another thoughtful interview about the importance of getting this right. And somewhere in a data center humming at full capacity, the next version is already training — drawing power equivalent to a mid-sized city, reading everything humanity has ever written, learning to be more capable than the last one.
That is where the ethics debate actually lives. Not in the letters pages, and not in the conference rooms. In the electricity bill.