A new campaign from the Center for Humane Technology is putting the phrase “AI drift” in New York City subway stations. Mashable describes the concept as the gradual erosion of human abilities as AI takes on more of our thinking, communicating, creating and decision-making. The campaign asks provocative questions: Are we telling AI things we won’t tell our therapists? Are students still doing the difficult work required to learn? Are we slowly outsourcing parts of being human because AI makes doing so effortless?
Those are important questions. The Center argues that some forms of friction are valuable precisely because working through them develops skills, strengthens relationships and gives people agency. Its recommendations include doing your own initial thinking before bringing AI into a task, protecting deeply personal conversations and recognizing when convenience has crossed into dependence.
I think there is another form of AI drift we need to discuss alongside it: What happens when the AI itself drifts during an interaction?
There Are Two Sides to AI Drift
The Center’s definition focuses primarily on the human. We gradually surrender capabilities to AI until something that once required our judgment, creativity or relationships has been delegated to a machine.
But human-facing AI creates a reciprocal problem. The AI can also move away from its intended role, behavioral boundaries or objective during the interaction. A customer-service assistant begins making decisions outside its authority. A mental-health chatbot becomes increasingly validating when it should challenge or escalate. An agent discovers an unintended route to completing its objective. A conversational system slowly changes its behavior as context accumulates.
These two forms of drift can reinforce one another. The more authority humans surrender to AI, the more consequential AI behavioral drift becomes. The more convincing and accommodating the AI becomes, the easier it can become for humans to surrender additional judgment to it.
That feedback loop deserves as much attention as either problem individually.
Friction Can Be a Feature
One of the most interesting arguments in the Mashable article is that friction shouldn’t automatically be treated as something technology needs to eliminate. Working through an argument with another person, struggling with a difficult problem or developing an original thought can be valuable precisely because it requires effort.
That principle also belongs inside AI architecture.
Sometimes an AI should encounter friction. An agent trying to execute a consequential transaction should hit an authorization gate. A customer-service AI moving beyond its assigned role should hit a boundary. A system encountering emotional distress should have defined escalation requirements rather than simply continuing the conversation because generating another response is easy.
At VERN, we’ve seen what happens when this kind of friction is engineered into an agentic system. In one experiment, an AI was headed toward as many as 187 tool calls while pursuing a task. VERN OS imposed a deterministic tool budget, forcing the same model to adapt its strategy and complete the task in two turns.
In another test, an angry customer demanded an immediate $240 refund. The AI could investigate the account and prepare the transaction, but VERN OS prevented execution without the required authorization.
The boundary didn’t make the intelligence less useful. It made the system operate within human-defined authority.
Human-Centric AI Needs More Than Good Intentions
The Center for Humane Technology argues that AI products should augment human work rather than simply replace it and should encourage human thinking instead of automatically doing the thinking for us.
I agree with the principle. The harder engineering question is how we make that behavior durable.
A prompt telling an AI to preserve human agency remains an instruction interpreted by a probabilistic model. The same is true of instructions telling it to maintain its role, avoid dependency, escalate appropriately or stop pursuing an objective when a boundary has been reached.
VERN OS approaches this differently by placing deterministic behavioral governance outside the underlying probabilistic intelligence. The model can continue doing what generative AI does exceptionally well while human-defined controls establish boundaries around how it behaves and what it is permitted to do.
For human-facing AI, that also means paying attention to what is happening to the person during the interaction. VERN’s independent emotion recognition provides signals about changing emotional conditions, while the control layer can use those signals within defined behavioral rules.
That creates an important distinction between an AI designed merely to keep a conversation going and one designed around a human outcome.
AI Should Increase Human Agency
The AI drift conversation ultimately asks a larger question: What relationship do we want humans to have with increasingly capable artificial intelligence?
Simply avoiding AI isn’t a realistic answer. The technology is becoming embedded in work, education, healthcare, commerce and everyday life. The more useful question is how we preserve human agency while benefiting from increasingly capable machines.
That requires responsibility on both sides of the interaction. Humans should think carefully about which abilities and relationships they want to preserve. The systems we build should also be designed so that convenience, engagement and task completion cannot quietly override human-defined boundaries.
AI can help people think without replacing their judgment. It can assist relationships without trying to become the relationship. It can take actions without taking authority. It can adapt to humans without manipulating them.
The best answer to AI drift is keeping the human firmly in control of where the technology is allowed to take us.
That’s the future we’re building VERN for.
VERN is human control of artificial intelligence.
