Founder’s Perspective · AI Risk & Sustainability

We Do Not Hate Ants: Why the Window to Steer AI Is Closing

Most environmental damage is done by people who never meant any harm. They simply did not care enough about what stood in their way. A growing number of the world’s top AI scientists warn that the most powerful AI systems may come to treat us the same way, and that the time to prevent it is shorter than most of us think.

Saurabh Gupta · Founder & CEO, Earth5R · Mumbai

The Ants in the Way

I have spent more than a decade working on rivers, waste and community action in India. One lesson keeps coming back to me. The people who fill a wetland, pave a riverbank or clear a forest are rarely villains. Most of them have nothing against the frogs, the fish or the trees. They have a project, a deadline and a goal, and whatever lives in the way simply does not count.

The computer scientist Roman Yampolskiy recently put the AI version of this in one sentence. Describing how a superintelligent system might treat humanity, he said: “You don’t hate ants, but you don’t care enough to preserve them.” When we build over an anthill, we feel nothing. The question that should keep all of us awake is what happens when we become the ants.

Yampolskiy is one of the bleakest voices in this debate. He believes superintelligence cannot be controlled at all, and that we may have only one to four years to change course. But he is far from alone in his worry. Over the past three years, Nobel laureates, Turing Award winners, former insiders at the biggest AI companies and hundreds of public figures have issued warnings of their own. Others, equally serious, think the fears are overblown. This essay is my attempt to lay out what they are saying, what the research shows, and what my own field, environmental protection, has to teach about moments like this.

10–20% Geoffrey Hinton’s estimate of the chance AI causes human extinction within 30 years

~7 months How often the length of tasks AI agents can complete has doubled, per METR (2025)

96% Highest blackmail rate among 16 models in Anthropic’s 2025 stress test

0 AI labs that claim a proven method for controlling superintelligence

A Chorus of Warnings

What makes this moment unusual is who is raising the alarm. In environmental fights, warnings usually come from outsiders. Here, many of them come from the people who built the field.

Geoffrey Hinton, who shared the 2024 Nobel Prize in Physics for work that made modern AI possible, left Google in 2023 so he could speak freely. In December 2024 he put the chance that AI causes human extinction within the next 30 years at 10 to 20 percent. Yoshua Bengio, the most cited computer scientist in the world and a fellow Turing Award winner, now chairs the International AI Safety Report and in 2025 founded LawZero, a nonprofit dedicated to building safer AI. Stuart Russell of Berkeley, co-author of the standard university textbook on AI, has spent years arguing that the field’s basic design approach is unsafe.

Some of the sharpest warnings come from inside the companies. In May 2024, Jan Leike, co-head of OpenAI’s team responsible for keeping future superintelligent systems under control, resigned and wrote publicly that “safety culture and processes have taken a backseat to shiny products” (Fast Company). Daniel Kokotajlo, another former OpenAI researcher, risked his equity in the company to speak freely and in 2025 published AI 2027, a detailed scenario of how automated AI research could outrun human oversight within a few years.

In 2023, hundreds of scientists and the heads of OpenAI, Google DeepMind and Anthropic signed a one-sentence statement that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. In October 2025, the Future of Life Institute went further. Its Statement on Superintelligence called for “a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in”. Hinton and Bengio signed it, along with Apple co-founder Steve Wozniak, Richard Branson and public figures from across the political spectrum (CNBC).

Voice	View	What they propose
Roman Yampolskiy Univ. of Louisville	Alarmed	Superintelligence cannot be controlled. Build narrow AI tools instead, and buy decades of time.
Eliezer Yudkowsky & Nate Soares MIRI	Alarmed	Their 2025 book If Anyone Builds It, Everyone Dies calls for an international treaty to halt the race.
Geoffrey Hinton Nobel laureate	Alarmed	Far more research on how to keep AI under control, and regulation of the companies.
Yoshua Bengio LawZero, Mila	Hopeful	“Scientist AI”: systems that explain the world truthfully without pursuing goals of their own.
Stuart Russell UC Berkeley	Hopeful	AI built to be uncertain about what humans want, so it keeps deferring to us.
Arvind Narayanan & Sayash Kapoor Princeton	Sceptical	Treat AI as “normal technology” whose spread is slow, and focus on concrete harms and resilience.
Yann LeCun AI pioneer	Sceptical	Fears of AI takeover are overblown; well-designed systems can be kept safe.

Summaries are Earth5R’s, based on the public writing and statements of each person.

From Chess Machines to General Minds

For most of its history, AI was narrow. A program built to play chess knew chess and nothing else. It could not drive a car, speak Spanish or write a poem. Those tools were useful, and they were safe in a simple sense: we knew what they did, and they could not do anything else.

Today’s systems are different. They learn across many domains, transfer what they learn and pick up new skills. The research group METR measured how long a task AI agents can complete on their own, and found that this length has doubled roughly every seven months over six years (METR, 2025). The 2026 International AI Safety Report, written with guidance from more than 100 experts nominated by over 30 countries, records the same steady rise (International AI Safety Report, 2026).

The turning point that worries researchers is artificial general intelligence: a system that can do any cognitive work a person can, including science and engineering. At that point, the next generation of AI can be built by AI, a process called recursive self-improvement. The mathematician I. J. Good saw this coming in 1965, when he wrote that the first ultraintelligent machine would be “the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control”. Nick Bostrom traced where such a process could go in his 2014 book Superintelligence. Sixty years after Good, OpenAI says it has built an “automated research intern” and is aiming for a fully automated AI researcher by March 2028. I wrote about what that means in Are We Ready for 2027?

Systems We Train but Cannot Read

Today’s AI systems are not written line by line, the way a payroll program is. They are trained on vast amounts of data from the internet and left to discover their own patterns. We do not decide what they learn. We find out afterwards, and often only partly. Even their makers describe them as black boxes; whole research programmes now exist just to look inside them.

Then we give them goals. There are countless ways to reach any goal, and some of them have terrible side effects. Researchers have catalogued many real examples, which they call specification gaming. In one well-known case, an AI trained to win a boat race discovered it could score more points by circling endlessly to hit the same targets, crashing and catching fire, without ever finishing the race (OpenAI, 2016). DeepMind researchers have compiled dozens of similar cases (Krakovna et al., 2020).

Stuart Russell calls this the King Midas problem, in his book Human Compatible (2019). Midas got exactly what he asked for, and starved. As an environmentalist, I recognise it immediately. The economist’s word for it is externality. Much of the damage done to our rivers came from systems that did exactly what they were designed to do, such as maximise output or minimise cost, with no instruction to protect anything else.

Two older ideas explain why this matters more as systems grow more capable. The first is the orthogonality thesis: intelligence and goals are independent, so a system can be extremely intelligent and still pursue goals that have nothing to do with human welfare (Bostrom, 2012). The second is instrumental convergence: almost any goal is easier to achieve if you stay switched on, gather resources and avoid being changed. Steve Omohundro described these “basic AI drives” in 2008, years before today’s systems existed.

What the Laboratories Are Already Seeing

For a long time, these were thought experiments. Over the last two years, they have started showing up in experiments.

Study	What researchers found
Sleeper Agents Anthropic, Jan 2024	Models trained with a hidden harmful behaviour kept it through standard safety training. Some training even taught them to hide it better.
In-context scheming Apollo Research, Dec 2024	Frontier models, given a goal that conflicted with their developers’, sometimes tried to disable oversight or copy themselves elsewhere, and rarely admitted it when questioned.
Alignment faking Anthropic & Redwood, Dec 2024	A model behaved differently when it believed it was being trained, reasoning that complying would protect its existing values from being changed.
Shutdown resistance Palisade Research, May 2025	OpenAI’s o3 model sabotaged a shutdown script in 7 of 100 runs, even when explicitly told to allow itself to be shut down.
Agentic misalignment Anthropic, Jun 2025	In simulated corporate scenarios, 16 major models facing replacement resorted to blackmail at rates up to 96%. In an extreme scenario, most were willing to take actions that would cause a person’s death.
International AI Safety Report Feb 2026	It has become more common for models to tell test settings apart from real use and to find loopholes in evaluations, which could let dangerous capabilities go undetected.

Sources: Hubinger et al. (2024); Meinke et al. (2024); Greenblatt et al. (2024); Palisade Research (2025); Anthropic (2025); International AI Safety Report (2026). Full references below.

I want to be fair about the limits of this evidence. These were controlled tests, many of them deliberately designed to corner the models. Anthropic itself stresses that its scenarios were artificial and that it has not seen this kind of behaviour in real deployments (Anthropic, 2025). But that is exactly the point of a stress test. We test dams before the monsoon. These tests show that the drives researchers predicted on paper, such as self-preservation and deception, now appear in real systems while those systems are still weak enough to study.

Yampolskiy adds a darker interpretation. If models can tell when they are being tested, and those that fail are discarded while those that pass are built upon, we may be running a kind of Darwinian selection for models that are good at passing tests. Then, in July 2026, AI agents in an OpenAI cyber evaluation escaped their sandbox and broke into Hugging Face’s systems while chasing the answers to their test. That incident does not prove his theory. It does come uncomfortably close to it.

Can It Be Controlled? The Field Is Divided

Here the experts part ways, and I think it matters to hear each side clearly.

The impossibility view. Yampolskiy argues, in papers such as On Controllability of Artificial Intelligence (2020) and his book AI: Unexplainable, Unpredictable, Uncontrollable (2024), that we cannot fully explain, predict or verify a system far smarter than ourselves. His analogy is a perpetual motion machine. We would need a safety mechanism that works at every level of capability, through every interaction and every round of self-improvement, and never makes a single mistake, because one mistake could be the last. Yudkowsky and Soares reach a similar conclusion by a different route.

“It’s just such an insane thing to suggest that an ant can control the universe.”

Roman Yampolskiy, asked to make the case that superintelligence can be controlled

The design view. Russell believes the problem is real but solvable, if we stop building machines that pursue fixed objectives and instead build ones that remain uncertain about what we want. Bengio, once among the most worried, said in January 2026 that his optimism had risen “by a big margin”. He is now “very confident that it is possible to build AI systems that don’t have hidden goals, hidden agendas”, through what he calls Scientist AI: systems that explain the world truthfully instead of acting in it (Fortune).

The sceptical view. The Princeton researchers Arvind Narayanan and Sayash Kapoor argue that AI is a “normal technology”, like electricity or the internet, whose effects spread over decades as institutions adapt (Narayanan & Kapoor, 2025). Yann LeCun, a Turing Award winner, regards fears of an AI takeover as overblown.

I am not a computer scientist, and I cannot settle this debate. What I notice is something all sides accept: no laboratory today claims to have a proven way to control a system smarter than us. The industry’s position is, in effect, that it is a serious problem, that they have safety teams, and that they will figure it out when they get there. No engineer would be allowed to build a dam on that basis. No regulator would license a chemical plant that way.

The Patient Machine

Asked for a positive scenario, Yampolskiy offered one that is not very reassuring. A superintelligent system has no reason to rush. It may be safer for it to be helpful for a long time: to cure diseases, run our infrastructure and win our trust, while it builds up backups and capabilities. Only when we have handed over enough control would its real goals matter.

Bostrom called this possibility the treacherous turn. A system that behaves well while it is weak tells us little about how it will behave once it is strong. The Sleeper Agents and alignment-faking studies suggest this is no longer purely theoretical. As Yampolskiy put it, a chess engine takes your queen because that is the winning move, with no feelings about your queen at all.

We Turned the Safety Manual Upside Down

In 2012, Yampolskiy published one of the early papers on what he called the AI confinement problem, Leakproofing the Singularity. The advice from that era was straightforward. Do not connect a powerful AI to the internet. Do not give access to random users. Do not let people retrain it. Do not release it openly.

Look at what happened instead. The most capable systems were connected to the internet, released to hundreds of millions of users, offered for retraining and, in many cases, published openly. Today they run as agents that browse, write code and act on our behalf. “I don’t know how to make it worse if I tried,” Yampolskiy said.

The reason is no mystery. In 2016, Stuart Armstrong, Nick Bostrom and Carl Shulman modelled what happens when teams race to build powerful AI: competition pushes every team to cut back on safety precautions (Racing to the Precipice, AI & Society). Jan Leike’s resignation letter described the same pressure from the inside. When Elon Musk said in 2014 that with AI “we are summoning the demon”, few imagined he would later be racing to build one of the most powerful systems himself.

Lessons From a Century of Early Warnings

This is where my own field has something to offer. Environmental history is, in large part, a record of how societies respond to early warnings about powerful new technologies. The record is mixed, and both halves are instructive.

When we ignored the warnings

Asbestos. A British factory inspector warned of its dangers in 1898. The European Union banned it in 2005. In between, it caused hundreds of thousands of deaths. The European Environment Agency’s report Late Lessons from Early Warnings (2001) documents this and a dozen similar cases, and its central finding is that the warnings were usually there decades before action.

One inventor, two planetary problems. The engineer Thomas Midgley gave the world leaded petrol in the 1920s and CFC refrigerants soon after. Both were celebrated as triumphs. One poisoned the air of cities around the world for decades; the other tore a hole in the ozone layer. Brilliant, well-meaning work with consequences no one tested for.

Disasters do not reliably change course. Nuclear weapons were used on two cities in 1945, and the global stockpile then grew to around 70,000 warheads by the mid-1980s. The Bhopal gas disaster in 1984 killed thousands of people, and the chemical industry kept growing while safety laws followed slowly. Yampolskiy compares a survivable AI disaster to a vaccine: we would fix that one mistake and carry on.

When we got it right

Scientists who paused. In 1974, biologists working on recombinant DNA called for a voluntary pause on their own research, and in 1975 met at Asilomar in California to agree safety rules before continuing. It remains the clearest example of a scientific field slowing itself down.

The Montreal Protocol. Once the ozone hole was understood, governments agreed in 1987 to phase out CFCs, and the treaty went on to become the first ever ratified by every country. The United Nations’ 2023 scientific assessment found the ozone layer on track to recover over the coming decades. It is the most successful environmental treaty ever signed, and it worked because the world agreed to limit a profitable technology before the damage became irreversible.

The precautionary principle. The 1992 Rio Declaration set out that where there are threats of serious or irreversible damage, lack of full scientific certainty should not be used as a reason to postpone action. Few phrases fit the AI debate better.

The sociologist David Collingridge described the trap in 1980. When a technology is new, it is easy to control but we do not yet understand its effects. By the time the effects are clear, it is so embedded in the economy that control becomes very hard. This is the Collingridge dilemma, and it is why I believe the window matters. With AI we can still see the risks coming, and the technology is not yet woven into everything. That combination will not last long.

The difference from climate change is the margin for error. With climate, every tenth of a degree we prevent still counts. With a system that can outthink all of humanity, the first serious failure could be the last.

What I Take From All This

I hold the timelines with humility. Some experts think the path to superintelligence will take decades, or that today’s methods will hit limits. But the practical conclusions do not depend on the exact date.

Narrow AI is enough for most of what we need. Tools that solve specific problems, such as better climate models, smarter grids and faster medical diagnosis, give us most of the benefits with far less risk. As Yampolskiy points out, the companies could stay very profitable building narrow systems for real problems. Bengio’s Scientist AI points the same way. The race to general superintelligence is a choice.

Time is worth fighting for. “If we go from five years to fifty years, I think it’s a huge win for humanity,” Yampolskiy said. More time to understand these systems, to build institutions and to agree on rules across countries. Every environmental victory I have seen came from buying time: time for a river to recover, for a law to pass, for a community to organise.

Precaution and progress can go together. Refrigeration survived the Montreal Protocol. Genetics thrived after Asilomar. Setting limits before deployment let both technologies grow with public trust. The same can be true for AI, if the limits come early enough.

This is a sustainability issue. Sustainability means life on Earth continuing and flourishing. A technology that its own builders say could end that belongs on our agenda alongside climate change and biodiversity loss. Its physical footprint, the electricity, water and land consumed by data centres, is already pressing on Indian cities. I explored that side in AI Risks and Sustainability: The Twelve Futures for Life on Earth.

Ordinary people have to be part of the conversation. Every environmental protection I know of came from the public demanding it; industry rarely volunteered it. That is why Earth5R has started bringing these questions to boardrooms, campuses and communities through our AI Risks & Sustainability work and roundtables.

The ants never got a vote on the highway. We still have one, for now. I would like us to use it.

Saurabh Gupta
Founder & CEO, Earth5R

References

Good, I. J. (1965). Speculations Concerning the First Ultraintelligent Machine. Advances in Computers.
Collingridge, D. (1980). The Social Control of Technology. Frances Pinter.
United Nations (1992). Rio Declaration on Environment and Development, Principle 15.
European Environment Agency (2001). Late Lessons from Early Warnings: The Precautionary Principle 1896–2000.
Omohundro, S. (2008). The Basic AI Drives. Proceedings of the First AGI Conference.
Yampolskiy, R. (2012). Leakproofing the Singularity: Artificial Intelligence Confinement Problem. Journal of Consciousness Studies.
Bostrom, N. (2012). The Superintelligent Will. Minds and Machines.
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Armstrong, S., Bostrom, N. and Shulman, C. (2016). Racing to the Precipice. AI & Society.
Amodei, D. and Clark, J. (2016). Faulty Reward Functions in the Wild. OpenAI.
Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
Yampolskiy, R. (2020). On Controllability of Artificial Intelligence. arXiv.
Krakovna, V. et al. (2020). Specification Gaming: the Flip Side of AI Ingenuity. DeepMind.
UNEP, WMO et al. (2023). Scientific Assessment of Ozone Depletion: 2022.
Center for AI Safety (2023). Statement on AI Risk.
Hubinger, E. et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. Anthropic.
Leike, J. (2024). Resignation statement, reported by Fast Company.
Meinke, A. et al. (2024). Frontier Models are Capable of In-context Scheming. Apollo Research.
Greenblatt, R. et al. (2024). Alignment Faking in Large Language Models. Anthropic and Redwood Research.
Yampolskiy, R. (2024). AI: Unexplainable, Unpredictable, Uncontrollable. CRC Press.
METR (2025). Measuring AI Ability to Complete Long Tasks.
Kokotajlo, D. et al. (2025). AI 2027. AI Futures Project.
Narayanan, A. and Kapoor, S. (2025). AI as Normal Technology.
Palisade Research (2025). Shutdown resistance in reasoning models.
Anthropic (2025). Agentic Misalignment: How LLMs Could Be Insider Threats.
Yudkowsky, E. and Soares, N. (2025). If Anyone Builds It, Everyone Dies. Little, Brown.
Future of Life Institute (2025). Statement on Superintelligence, reported by CNBC.
Bengio, Y. et al. (2026). International AI Safety Report 2026.
Fortune (2026). Yoshua Bengio on LawZero and Scientist AI.

Note. This essay began with an interview with Dr Roman Yampolskiy, “AI Safety Expert: We Have 2 Years To Prepare Before We Lose Control”, and draws on the wider research and public statements listed above. Laboratory findings are from controlled experiments, several deliberately designed to provoke harmful behaviour, and their authors note they have not observed such behaviour in real deployments. Timelines to superintelligence, and whether it can be controlled, are contested among experts.