The Wall Street Journal has put its finger on perhaps the most consequential conflict developing inside artificial intelligence.
OpenAI and Anthropic are racing toward increasingly capable systems while competing against each other and China. Researchers are reporting startling advances, including progress toward systems capable of contributing to their own improvement. Meanwhile, people inside these same organizations are increasingly warning that AI capability may be advancing faster than our ability to control it.
Then add money.
These aren’t research projects operating in isolation. They’re companies pursuing enormous commercial opportunities while spending extraordinary amounts of capital. Competition for customers, talent, investment and technological leadership creates tremendous pressure to keep advancing. Potential public offerings add another dimension to those incentives.
That’s what makes the current moment so important.
The companies developing frontier AI are simultaneously being asked to build the most capable intelligence possible and determine how that intelligence should be controlled.
Safety Is Competing With Capability
There is an uncomfortable economic reality underneath the safety debate.
Imagine one AI company decides that a new capability is too risky to release. Its competitor releases something comparable anyway. Customers move. Developers follow. Capital follows. The cautious company may have made the responsible decision while creating a competitive disadvantage for itself.
Add international competition and the problem becomes even harder. The Journal describes the race with China as another powerful reason American labs believe they cannot simply stop advancing frontier models.
That doesn’t mean the people running these companies don’t care about safety. Anthropic, in particular, has built much of its identity around responsible AI development. Dario Amodei’s recent call to slow frontier development demonstrates how seriously he views the problem.
But good intentions don’t eliminate incentives.
And incentives matter when we’re designing safety architecture.
If our primary answer to AI safety is asking the companies racing to build increasingly powerful probabilistic systems to make those same systems reliably control themselves, we’ve created a structural weakness.
Control Shouldn’t Depend on Which Model Wins
This is where I think the industry needs to separate two questions.
How do we build safer artificial intelligence?
And how do humans maintain control over artificial intelligence?
We need to work on both.
Model companies should improve alignment, evaluation, interpretability and safeguards. Governments should establish reasonable standards. Researchers should continue investigating the risks associated with increasingly capable systems.
But the organization deploying AI should also have an independent mechanism for governing what that intelligence is allowed to do.
VERN is human control over artificial intelligence.
VERN OS sits outside the underlying model as an independent governance layer. The model provides intelligence. VERN governs the AI-human interaction according to deterministic behavioral requirements established by the organization deploying it.
That distinction becomes more important as models become more powerful.
An enterprise shouldn’t lose its behavioral safeguards because a provider releases a new model. It shouldn’t have to trust that a model update interprets its safety instructions exactly the same way. It shouldn’t surrender its definition of acceptable behavior because another model suddenly becomes faster, cheaper or smarter.
The intelligence can change.
Human control should remain.
The Human Has to Be the Center of the Architecture
There’s another problem with leaving governance entirely to model providers.
Their objectives and the human user’s interests aren’t necessarily identical.
A model company might optimize for usefulness. An application developer might optimize for engagement. A business might optimize for conversion, completion or cost reduction.
None of those automatically means the interaction is improving the condition of the person using the AI.
A human-centric AI system needs to consider what is happening to that person during the interaction.
Is frustration escalating? Is the AI reinforcing harmful thinking? Is it creating emotional dependency? Is it manipulating someone toward an outcome? Has the interaction moved outside the AI’s appropriate role? Is continuing the conversation helping the person or potentially harming them?
These aren’t hypothetical questions anymore. AI systems are already interacting with people experiencing loneliness, distress, financial problems, healthcare concerns and other emotionally consequential situations.
The goal of governance therefore can’t simply be making sure the AI follows a policy document.
The AI-human interaction should operate for the betterment, rather than the detriment, of the human being.
That is a much higher standard.
It’s also why emotional recognition belongs alongside behavioral control. If the system can recognize changes in the human side of the interaction, governance can respond to what is actually happening rather than simply checking whether certain prohibited words appeared.
External Governance Changes the Incentive Problem
This is where the architecture gets interesting.
If behavioral governance is independent of the underlying model, enterprises no longer have to choose between capability and control in quite the same way.
A better model comes along? Use it.
A cheaper model becomes available? Switch.
Different models perform better for different tasks? Route between them.
The behavioral requirements governing the human experience remain external to those choices.
That creates an important separation between the competitive race to build intelligence and the responsibility to govern how that intelligence interacts with humans.
The model companies can compete.
The enterprise can maintain control.
And the human doesn’t have to depend entirely on either one making the right decision every time.
We Need More Than Promises
The Journal’s story captures an industry confronting a difficult contradiction. The potential economic value of increasingly powerful AI is enormous, while some of the people closest to its development are warning that the risks are increasing with it.
We shouldn’t expect that contradiction to disappear.
The financial incentives will remain. The geopolitical competition will remain. Model capabilities will continue advancing. Different companies will make different judgments about acceptable risk.
That makes independent governance more important, not less.
Human control over artificial intelligence shouldn’t depend on which company wins the model race, which safety philosophy its leadership adopts, or whether commercial pressure eventually outweighs caution.
It should be built into the architecture surrounding the intelligence.
And when that intelligence interacts with a person, the human should remain the reason the control exists
