A senior OpenAI safety leader has resigned with an unusually direct warning about the organization he is leaving. David Robinson, who led the writing of safety reports accompanying OpenAI product releases, says the company’s culture is “broken” and argues that frontier AI companies are not being sufficiently careful.
There will inevitably be debate about whether Robinson’s assessment of OpenAI is fair, and some of the existential-risk predictions surrounding frontier AI remain contested. But enterprises deploying these systems don’t need to resolve either argument to recognize a more immediate problem: Their control over AI should not depend entirely on the internal safety culture of another company.
There is an even deeper problem when AI interacts directly with people. Governing what an agent can access or execute is only part of the job. Organizations also need to govern how that intelligence behaves toward a human being, particularly when the interaction becomes emotional, adversarial, confusing or vulnerable.
That requires something most AI governance architectures don’t have: An independent understanding of the human side of the interaction.
The Missing Half of AI Governance Is Human
Most AI safety systems begin with the AI. They evaluate the model’s output, restrict its tools, monitor its actions or attempt to determine whether its response violates a policy.
Those controls matter, particularly as agents gain greater autonomy. But an agent interacting with another machine and an agent interacting with a distressed customer are fundamentally different governance environments.
A customer may become angry. A patient may express fear. Someone may become increasingly distressed over the course of an interaction even though no individual sentence appears particularly alarming. The appropriate behavior of the AI may need to change as the emotional context changes.
Asking the LLM itself to determine that context creates another probabilistic judgment inside an already probabilistic system.
VERN takes a different approach.
Our Emotion Recognition System is an independent, deterministic neurolinguistic model. It evaluates emotional signals in language outside the LLM, giving the governance layer a consistent signal it can use when determining how the AI should behave.
The AI doesn’t get to decide what emotion it thinks it sees and then decide what rules should apply because of that interpretation. Emotional state and behavioral governance can exist outside the probabilistic reasoning process.
That distinction is central to VERN’s architecture.
One Governance Layer for Machines and Humans
This becomes especially important as enterprise AI evolves from individual assistants into networks of autonomous agents.
Agent-to-agent interactions require governance around authority, delegation, tool use, role containment and consequential actions. When those agents interact with humans, all of those requirements remain, but another dimension appears: The human experience itself.
The same customer-service agent might communicate with a billing agent, a CRM agent and a customer during a single workflow. The organization shouldn’t need one governance architecture for the machine interactions and another disconnected system for the human conversation.
VERN OS is designed to govern both.
When an agent communicates with another agent, deterministic controls can govern authority, role, escalation and execution. When that interaction reaches a human, VERN’s deterministic emotional model adds information about the human side of the exchange, allowing Behavioral Control Modules to govern how the AI responds as the interaction develops.
The governance layer follows the interaction from agent to agent to human.
That’s a significant difference between controlling AI infrastructure and governing AI behavior.
Emotional Intelligence Becomes a Control Input
Emotional intelligence in AI is often treated as a feature of the model: Ask the LLM to infer how somebody feels, then instruct it to respond empathetically.
That approach has a fundamental weakness. The same probabilistic intelligence is interpreting the human, deciding what the interpretation means and generating the response. If its inference is wrong, the behavior built on top of that inference can also be wrong.
VERN separates those responsibilities.
Our deterministic emotional model provides an independent signal about emotional expression in the interaction. VERN OS can then use that signal as an input into behavioral governance, alongside role requirements, authorization rules, escalation conditions and other controls.
Consider an AI customer-service agent. The system might be permitted to handle an ordinary billing dispute autonomously. As anger or distress increases, however, the behavioral requirements can change. The AI may need to adjust its behavior, remain within stricter boundaries or escalate according to rules established by the organization.
The critical point is that the enterprise defines those rules. The underlying LLM does not get unilateral authority to decide when they apply.
Emotional intelligence becomes part of the control architecture.
Determinism Matters Most When Humans Are Involved
Probabilistic AI is extraordinarily powerful because it can handle ambiguity, reason across enormous amounts of information and generate responses that were never explicitly programmed.
Those same characteristics make it poorly suited to being the only enforcement mechanism governing its own behavior.
This is especially important in human interactions. An enterprise cannot reasonably establish a behavioral standard that means “usually respond appropriately when someone becomes distressed.” It needs repeatable rules around what happens when defined conditions occur.
VERN’s architecture combines probabilistic intelligence with deterministic governance. The LLM retains the flexibility that makes generative AI valuable, while VERN maintains consistent controls over the behavioral requirements surrounding it.
That gives organizations something increasingly important: The ability to improve or replace the underlying intelligence without replacing the behavioral model governing the human experience.
Claude can change. Gemini can change. GPT can change. An enterprise’s standards for how its AI treats people don’t have to change with them.
The Model Provider Shouldn’t Own the Human Experience
This is where Robinson’s warning about culture becomes particularly relevant.
Model providers will make decisions about training, alignment, safety testing and product releases. Leadership will change, competitive pressure will fluctuate and models will continue evolving rapidly.
An enterprise cannot outsource its responsibility for the human experience to those internal decisions.
A hospital needs to determine how its AI behaves with patients. A bank needs to determine how its AI behaves with customers. A university needs to determine how its AI behaves with students. Those standards belong to the organizations responsible for those relationships.
VERN gives them a governance layer independent of whichever probabilistic intelligence happens to be underneath the application.
That independence extends beyond model choice. It creates a consistent behavioral architecture across AI-to-AI and AI-to-human interactions, rather than forcing enterprises to assemble separate governance systems for every point where intelligence moves through the organization.
Human Control Requires Understanding the Human
The AI governance market is rapidly becoming crowded with systems designed to control what agents can access, which tools they can invoke and which actions they can execute. Those capabilities are necessary, and VERN OS provides deterministic controls around many of those same operational boundaries.
But the enterprise AI environment does not end at the API call.
Eventually, many of those agents interact with a person.
At that boundary, governance needs information about both sides of the interaction. It needs to understand what the AI is attempting to do and recognize meaningful emotional signals coming from the human it is serving.
That is the moat we’ve spent years building.
VERN combines deterministic behavioral governance with an independent deterministic emotional model, creating one control architecture capable of governing agent-to-agent and agent-to-human interactions.
The underlying intelligence can continue getting smarter. The models can change. The agents can multiply.
The organization retains one consistent governance layer across the entire interaction.
Because human control over artificial intelligence ultimately requires governing what AI does and how it behaves when it reaches a human being.
That’s VERN OS.
