I recently watched a talk by Chloe Lubinski of Anthropic about how artificial intelligence works and the questions increasingly capable systems are forcing us to confront.
One idea stayed with me.
We are building systems that are becoming remarkably capable, yet even their creators cannot always fully explain why they behave as they do.
As a software engineer, I find that fascinating.
Traditional software is built from explicit instructions. We write code, define rules, and specify what should happen under particular conditions. Modern AI is different. We design the architecture, training process, objectives, and safeguards, but we do not explicitly program every behavior that emerges.
How do you regulate something whose capabilities are advancing faster than our ability to fully understand them?
But I think there is an even harder question.
What if everyone agrees we should be careful, but nobody can afford to slow down?
Suppose an AI company decides its next model needs significantly more safety testing before release. That sounds responsible. But what if its competitors keep moving?
The cautious company could lose customers, investment, talent, and technological leadership.
Countries face the same dilemma. A government may believe stronger safeguards are necessary while worrying that imposing them alone could surrender an economic or strategic advantage to another nation.
AI safety is not simply a regulatory problem. It is a coordination problem.
The 2026 International AI Safety Report describes this tension directly. Companies that invest more heavily in reducing risk may be placed at a competitive disadvantage against companies that prioritize development speed. The same dynamic can occur between countries when governments fear that rivals will not adopt comparable safeguards.
A safety system that works only when participants willingly accept a competitive disadvantage is unlikely to remain stable.
Uncertainty Does Not Mean We Ignore the Risk
There is real disagreement about how dangerous advanced AI could eventually become. Some researchers believe highly capable systems could create catastrophic risks. Others consider those scenarios too speculative and argue that more attention should be paid to harms AI is already causing.
That disagreement matters.
Worst case scenarios should not be presented as predictions.
But recent warnings have come from people with direct experience inside the organizations building advanced AI systems.
Former Anthropic researcher Jacob Coxon resigned while publicly raising concerns about advanced AI and loss of human control. Former Google DeepMind researcher Bilal Chughtai has expressed similar concerns.
Paul Christiano, a member of OpenAI's nonprofit board and an experienced AI safety researcher, recently warned of the possibility of catastrophic and irreversible loss of control if capabilities advance without sufficiently strong alignment and safeguards.
They could be wrong.
But risk management does not require certainty.
Engineers routinely make decisions under uncertainty. We look at both the probability of failure and the consequences if that failure occurs.
A low probability event with minor consequences may justify limited attention. A low probability event with extraordinary consequences deserves a different level of scrutiny.
We do not have to believe catastrophe is inevitable to ask whether our safeguards are adequate.
We Have Faced Coordination Problems Before
Artificial intelligence and nuclear technology are very different, and the comparison should not be pushed too far.
But the history of nuclear nonproliferation offers an important lesson.
The two superpowers did not stop competing before they began cooperating. They remained strategic adversaries.
What gradually emerged was recognition that certain outcomes would be disastrous regardless of who caused them.
The Nuclear Non-Proliferation Treaty created a framework intended to limit the spread of nuclear weapons while allowing peaceful uses of nuclear technology.
Because trust alone was not enough, safeguards and verification became central parts of the international system, including inspections conducted through the International Atomic Energy Agency.
International cooperation does not require competitors to trust one another. It requires rules that can function even when they do not.
AI Safety Cannot Depend on Someone Volunteering to Lose the Race
A company can sincerely care about AI safety while believing that slowing development would hand an advantage to a competitor.
A country can support international safeguards while refusing restrictions that its rivals are not required to follow.
Everyone can believe safety matters and still keep accelerating.
That is the core of the problem.
Effective AI governance has to make responsible behavior compatible with continued competition.
One approach may be to focus regulation on capability rather than treating every AI system the same.
A small AI tool that summarizes documents does not present the same risks as a system capable of autonomously discovering cybersecurity vulnerabilities, conducting scientific research, or operating across critical infrastructure.
As systems cross agreed thresholds of consequential capability, requirements could increase accordingly.
Independent evaluation
Security testing
Incident reporting
Controlled access to particularly dangerous capabilities
Stronger safeguards as capabilities increase
The exact thresholds will be debated, and they should be.
But the larger principle is important.
Those requirements must apply broadly enough that responsible developers are not punished for following them.
We should not design a safety system whose success depends on every participant voluntarily making the same responsible choice. That is not a safeguard. It is an assumption.
The Deeper Problem Is Human
We often talk about aligning AI with human values.
But humanity itself is not aligned.
We disagree about politics, religion, economics, morality, freedom, privacy, and countless other things.
Perhaps global AI governance does not need to begin by resolving those disagreements.
Perhaps it begins with a narrower question.
What outcomes can we agree are unacceptable, regardless of who creates them?
That may be a more realistic starting point.
Humanity has confronted versions of this problem before, imperfectly and often later than we should have.
History shows that competition and cooperation can exist at the same time.
Responsible Behavior Has to Be Rational Behavior
The hardest part of AI regulation may therefore not be deciding what rules to write.
It may be creating a system in which responsible behavior is also the rational competitive choice.
Human beings have often learned to cooperate after experiencing the consequences of failing to do so.
With AI, the question is whether we can learn to cooperate before we have to.
The challenge is not simply convincing individual companies or countries to behave responsibly. It is designing a system where cooperation, verification, competition, and safety can exist together.
References and Further Reading
- Alliance for Responsible Citizenship: Anthropic's Chloe Lubinski Explains How AI Works
- International AI Safety Report 2026
- Axios: Anthropic Researcher Jacob Coxon Resigns Over AI Safety Concerns
- Reuters: Former Google DeepMind Researcher Bilal Chughtai Raises AI Safety Concerns
- Paul Christiano: Personal Statement on Joining the OpenAI Board
- United Nations: Background on the Treaty on the Non-Proliferation of Nuclear Weapons
- International Atomic Energy Agency: Nuclear Safeguards and Verification