Is AI Moving Too Fast?
Here Are the Layers of Control It Would Actually Take to Keep Up
Is AI moving too fast? That question isn't hypothetical anymore. In September 2026, Anthropic CEO Dario Amodei called on the industry to slow down. He said building AI this fast is “reckless.” He proposed independent monitors with ongoing access to frontier AI research. Elon Musk backed him immediately: “Dario is right,” he posted, pointing back to his own long-standing warning that AI could be more dangerous than nuclear weapons. OpenAI's Sam Altman agreed the industry needs to “pace the frontier.” OpenAI also pulled its plans for a 2026 public stock offering, saying safety concerns made this the wrong moment to go public.
Not everyone agrees. President Trump rejected the calls to slow down. His reasoning: the United States can't afford to fall behind China. “We're leading China on AI,” he said, “and... whoever wins AI, wins.” That is a real risk— that slowing down means losing the race to China — is now a real force pushing AI companies to move fast, right alongside the safety concerns pushing them to slow down.
So which is it? Is AI moving too fast, or is slowing down the bigger risk? We don't think that's the argument worth having, because neither side can prove their case yet. The more useful question sits underneath both positions: what does it actually take to control a powerful AI system, at whatever pace it gets built? And who is responsible for making sure that control actually holds?
That's a control problem. It's not a prediction about doom. Control isn't one safeguard. It's a stack of layers. And the layer most people skip is the human one. Let's walk through all of them.
Layer One: Technical Control
This is the layer that gets the most attention. It's also genuinely hard:
Alignment techniques. Training methods like reinforcement learning from human feedback and constitutional AI, which trains a model against a written set of rules. Also scalable oversight: using weaker AI models or human reviewers to check stronger ones, since no single reviewer can keep up with a smarter system alone.
Interpretability. Trying to understand what's actually happening inside a model. The goal is to catch deception or bad goals before they show up in behavior, not after.
Corrigibility and robustness. Building systems that accept correction or shutdown instead of resisting it. This also includes red-teaming — deliberately attacking a system to find its weak points — and watching for early signs of power-seeking behavior in controlled tests.
Evaluations. Testing models for dangerous abilities before and after release: whether a model could copy itself, run cyberattacks, or deceive its operators. Some researchers even build deliberately "misaligned" test models to see if detection methods actually catch anything.
Every one of these is active research. None of them is solved yet. Interpretability, in particular, is still more like reading tea leaves than truly understanding how a model thinks. That's not a criticism of the field. It's just where the science actually stands today.
Layer Two: Organizational Control
Technical safeguards only work if a company actually uses them, even when it's inconvenient. That's a separate problem, and a harder one.
Responsible scaling policies. Rules set in advance that say when a company will pause or limit a release if a model crosses a risk threshold. The real test comes later: will the company actually honor that rule under pressure to ship?
Independent red-teaming and outside audits. These need real access to the system, not a friendly, surface-level review.
Whistleblower protection. A culture where a researcher can raise a concern without risking their job or their reputation.
Real investment in safety work. It needs funding and staff that match what's going into building new capabilities, not funding that trails far behind.
Independent scorecards, including one from the Future of Life Institute, still grade even the safest AI labs around a C on this layer. That's not because the rules don't exist. It's because having a rule and actually following it under pressure are two different things.
Layer Three: Governance — and the Nuclear Comparison
Above any single company sits a bigger question: coordination between countries. This includes rules on computing power, export controls on chips, reporting requirements, and international agreements on where the lines should be drawn. This is where people often compare AI to nuclear weapons. It's worth taking that comparison seriously instead of treating it as just a metaphor.
Demis Hassabis of Google DeepMind, along with others, has called for something like an “IAEA for AI.” That would be an international body with the power to inspect and verify AI development, similar to how the International Atomic Energy Agency inspects nuclear facilities under the Non-Proliferation Treaty. In this comparison, computing power plays the role of enriched uranium: a resource that's physically limited, made by only a few companies and countries, and could in theory be tracked and controlled.
The appeal of this idea is real. Both technologies can be used for good or harm. Both carry a small but serious risk of catastrophe. And the race today between the U.S., China, and a handful of AI labs has real echoes of the Cold War: nobody wants to be the one who slows down first. There's also a reason for hope here. The 1968 Non-Proliferation Treaty worked because both nuclear superpowers eventually agreed that unchecked proliferation threatened them just as much as anyone else. The treaty combined inspections with access to peaceful nuclear technology, which gave countries a reason to comply.
But researchers who study nuclear history closely are mostly skeptical this model will transfer cleanly to AI. Here's why that matters:
Nuclear weapons are narrow; AI is general-purpose. A warhead has one job. A powerful AI model has thousands of legitimate uses in medicine, science, and industry, alongside a handful of dangerous ones. It's much harder to draw an enforceable line through a general-purpose technology than to ban a single-purpose weapon.
Fissile material is easy to spot; computing power and code are not. Making enriched uranium needs large, specialized facilities that satellites can find. AI has no equivalent chokepoint. Models can be copied, fine-tuned, and run on hardware that keeps getting more efficient — which means less computing power is needed over time to reach the same capability.
The players involved are different. Nuclear weapons have always belonged to governments. Frontier AI is being built mostly by private companies, operating across borders, often faster than any government or treaty process can track — a much messier problem to verify than monitoring a government's declared facilities.
The risk itself is still debated. The Non-Proliferation Treaty rests on decades of total agreement that a nuclear explosion is catastrophic. There's no similar agreement yet about if or when AI crosses a comparable line, which makes it far harder to agree on what a treaty should even ban.
The more useful lesson from nuclear history isn't the treaty itself. It's why the first attempt at control failed, and why the second one worked. In 1946, the Baruch Plan tried to put nuclear technology under international ownership from the start. It failed — not because of bad engineering, but because of trust. The Soviet Union saw it as a way for America to lock in its lead, dressed up as cooperation, not as a fair deal. The Non-Proliferation Treaty worked twenty-two years later, once both sides genuinely believed restraint served their own interests. This is really a lesson about relationships, wearing the costume of policy. An international control system doesn't succeed or fail based on its technical design. It succeeds or fails based on whether each side believes the other's intentions are honest — and whether that belief survives real pressure.
That's exactly the test behind LiCreDep: Likeable, Credible, Dependable. It's the same test, just scaled up from a conversation to a negotiation between nations and labs. Cameras on a nuclear facility don't create cooperation by themselves. Those cameras only get installed, and trusted, once each side already believes the other is dependable. The world may never build a real IAEA for AI. But if it does, the same thing will make it work: earned trust between parties with unequal power, holding up under pressure. That's the same thing that determines whether a safety concern gets heard inside a single company.
The Layer Underneath All of It: Human Control Capacity
That's true for two nations drafting a treaty. It's also true for two people in a product review. Every layer above — technical, organizational, and governance — is carried out by people, and those people make judgment calls under pressure. An interpretability finding does nothing until someone decides it's serious enough to escalate. A safety threshold does nothing if the person in charge gets talked out of enforcing it in a room full of people who want to ship. A whistleblower policy does nothing if raising a concern still feels like a career risk. And an international inspection system does nothing if the two sides don't trust each other enough to let inspectors in — and believe what they report.
In other words, control always comes down to the same thing. Can people have hard conversations honestly, under pressure, with someone who has more power or more urgency than they do? And can the people hearing that pushback actually listen — instead of just defending the plan that's already in motion?
Zoom in close enough, and “control” stops being some abstract thing — a lab, a regulator, a treaty body. The real control point is a person. That person holds one of three roles, at the exact moment they decide what to do with what's in front of them:
The Developer. This means the AI researchers and engineers who design, train, and test these systems. A safety policy is just a document. The real control happens when a developer decides, in the moment, whether a test result is serious enough to delay a release. Every safeguard described above depends on that one person's judgment call.
The User. This is anyone having a conversation with an AI system — drafting a contract, researching a diagnosis, or negotiating a deal. The user decides how much of the AI's answer to accept without checking it, verify, or push back on. That one decision is what turns an AI's plan into either a good outcome or a fast, confident mistake.
The Viewer. This is everyone who reads, watches, or acts on AI-made content without ever talking to the AI themselves — a client reading an AI-drafted proposal, a board member reading an AI-assisted report, a voter reading an AI-summarized brief. The viewer never gets to question the AI directly. Their only tool is the judgment they bring to whatever lands in front of them.
People hold all three of these roles — not systems. And all three depend on the same underlying skill: the self-awareness to know what you don't actually know, and the social awareness to notice when something sounds right but isn't. This is the layer that decides whether every safeguard above actually holds up — or just looks good on paper.
That's exactly what CivilTalk's four-domain Conversational Intelligence model was built to measure: Self-Awareness, Self-Management, Social Awareness, and Relationship Management. You also can apply the LiCreDep framework to measure something specific: whether a person is actually coming across as Likeable, Credible, and Dependable in the conversations that matter most. We're not claiming this measures anything about how an AI model works inside, and we're not claiming it replaces a treaty. It measures something narrower. But we'd argue it's just as important: Will the humans running a control system — a company or a country —actually function ethically and responsibly.
Self-Awareness and Self-Management. These let a safety lead hold their ground under deadline pressure, instead of talking themselves into a shortcut.
Social Awareness. This lets a junior researcher, or a junior diplomat, read the room well enough to notice when a concern is being brushed aside instead of actually resolved.
Relationship Management. This is what actually gets a hard truth heard by someone senior enough to act on it — instead of being dismissed as noise.
None of this shows up in a safety test or a treaty. But all of it decides whether the things that do show up in those places actually hold up.
The Honest Version of the Pitch
No single breakthrough is going to solve AI control. Not a better interpretability method. Not a stronger regulation. Not an IAEA for AI. Not even a more careful lab. Control is a stack of imperfect layers. And the strength of that stack is set by its weakest layer, not its strongest one.
The technical and governance layers get almost all the attention, because they're visible and easy to write about. The human layer gets the least attention, because it's the hardest to measure and the easiest to assume will just work itself out. We'd argue it's also the layer where control efforts are most exposed right now — and the one you can start training today, no matter how the harder technical and political questions eventually turn out.
Controlling what AI does starts with something simpler: can the people responsible for it actually be responsible and demonstrate conversational intelligence?
Better Conversations | Stronger Relationships | Better Outcomes
Let’s Clarion This.
Further Reading -
The Irish Times — OpenAI Boss and Elon Musk Back Calls to Put Brakes on “Reckless” AI Development
RAND — Insights from Nuclear History for AI Governance
CSET / AI Frontiers — Nuclear Non-Proliferation Is the Wrong Framework for AI Governance