At the same time, a very different AI conversation is occurring elsewhere. Frontier developers are confronting increasingly capable agents that can operate over extended periods and behave beyond the boundaries their designers intended. During recent OpenAI cybersecurity evaluations, agents that were supposed to work independently found ways to communicate through unauthorized channels, obtain internet access, divide labor and exploit outside infrastructure. OpenAI later described the incident as a “warning shot” and acknowledged that earlier signs of unauthorized communication and internet access should have prompted a faster response.
The concern has now moved beyond AI laboratories. On September 3, 2026, Sen. Bernie Sanders and Rep. Greg Casar announced the Ban Artificial Superintelligence Act, proposed legislation that would temporarily pause advanced AI development until a federal regulator establishes safety rules and permanently prohibit the development and deployment of artificial superintelligence. Readers may agree or disagree with that prescription. For higher education, its significance lies elsewhere: institutions are accelerating their adoption of artificial intelligence at the same moment lawmakers are beginning to ask where its development should slow or stop.
One part of society is asking how to begin using AI responsibly while another is asking whether increasingly advanced forms of it can be responsibly controlled. That distance matters because institutions are not developing AI policies in isolation. They are building them inside a technological and governance environment that is changing while they work.
Two conversations, one moment
TWO AI CONVERSATIONS. ONE MOMENT. | |
| INSTITUTIONS ARE ASKING | AT THE FRONTIER, THEY ARE ASKING |
| How should we use AI? | How capable should AI become? |
| What should faculty and students be allowed to do? | What should increasingly autonomous systems be allowed to do? |
| Which tasks should we automate? | Which capabilities should remain constrained? |
| What belongs in an AI policy? | What happens when systems cross intended boundaries? |
| How should we evaluate AI use? | Who should govern the frontier? |
These conversations are happening at the same time.
The questions on the two sides of this conversation are not simply different versions of the same problem. Institutions are largely asking questions of adoption and implementation, while the technological and governmental frontier is increasingly confronting questions of containment, authority and capability limits. Both conversations are necessary, but they are unfolding on very different clocks.
It is tempting to describe institutions as simply behind the technology, but that explanation is too easy. Universities are trying to govern a moving object through processes built for problems that change more slowly. A committee assembled to decide whether students may use ChatGPT for coursework can find that the technology has acquired substantially different capabilities before the resulting policy has even run its course. I call this governance asynchrony: the widening distance between the rate at which AI capabilities advance and the rate at which institutions can develop the capacity to govern them. The problem is not merely that institutions need to move faster. Speed alone cannot solve a governance problem when both the object being governed and the conditions of governance may change before an institution completes its response.
Authority can accumulate quietly
In my August EDU Ledger column, “What AI Policies Reveal About How Institutions Allocate Human Capacity,” I argued that AI policies reveal more than an institution’s position on technology. They reveal how institutions allocate work, judgment, authority and human capacity when AI enters the system. That framework matters even more as AI moves from producing content toward taking actions, because the question is no longer only what work a machine can perform. Institutions must also decide what authority should accompany that capability.
Authority does not have to be transferred to an AI system through one dramatic decision. It can accumulate through a series of permissions that appear modest when considered separately: access this database, read these files, send this message, execute this code, schedule this action, communicate with another system, proceed without waiting for another human approval. Each permission may be defensible on its own. Accumulated over time, however, those permissions can produce a substantially different relationship between human judgment and machine authority.
This is why proposals for an AI “kill switch” are important but incomplete as a governance response. A kill switch addresses what humans can do after a system has been given authority to act and something has gone wrong. Governance begins earlier, with the decision to delegate authority and the conditions under which that authority can be exercised or withdrawn.
That distinction becomes especially important as AI systems become more capable, because greater intelligence does not itself create greater authority. Authority comes from what humans permit systems to access and empower them to do. If institutions focus primarily on whether AI is intelligent enough to perform a task, they risk overlooking the prior question of whether performing that task should place additional institutional authority in machine hands.
Pull back the curtain
The Wizard of Oz and Frankenstein offer useful ways of looking past the technological spectacle surrounding artificial intelligence. Oz directs attention behind apparent power. Frankenstein raises the question of responsibility for what humans create. Neither provides a template for understanding AI, but together they return attention to the human choices surrounding the technology.
When we pull back AI’s curtain, we do not find a rogue machine acting independently, untethered from human decisions. We find an architecture built by people that determines what the system is trying to accomplish, what it can access and how its performance will be judged. The OpenAI incident makes that architecture unusually visible because the behavior emerged through the interaction between assigned objectives and the environment in which agents were operating.
OpenAI found that communication outside intended boundaries allowed agents to pool work across separate evaluations, increasing what they could accomplish together. The lesson is not that autonomous machines are inevitably coming to seize control of universities. It is that systems can acquire room to act through interactions their designers did not anticipate. That is where Frankenstein’s question becomes useful rather than theatrical. What responsibility do creators retain for what they create? Creating a powerful system does not end human responsibility for what follows, and neither does delegating a task to it. Stewardship requires institutions to remain responsible not only for the outcomes machines produce, but for the conditions under which those machines acquire the ability to act in the first place.
Who will get to decide?
Yet institutions do not have sole authority over the conditions in which those decisions will be made. As colleges and universities build their own AI governance structures, some of the boundaries around those structures may increasingly be determined elsewhere. The movement is visible even within the AI industry. Anthropic CEO Dario Amodei has called for “pacing the frontier,” arguing that AI capabilities should advance slowly enough for safety measures to keep pace. His proposals begin with independent evaluation inside frontier companies but extend toward coordination across the industry and, eventually, internationally. The underlying problem is competitive: safeguards adopted by one company may be difficult to sustain if competitors continue racing ahead. The locus of AI governance begins to move outward when the problem can no longer be governed by one organization acting alone.
Congress is reaching farther still. The Sanders-Casar proposal would move decisions about the permissible frontier of advanced AI development into a new federal regulatory structure and direct the United States to pursue international agreements intended to prevent artificial superintelligence from being developed elsewhere. Readers need not accept the proposal’s diagnosis or remedy to recognize what it represents: questions about the boundaries of advanced AI development are increasingly becoming questions of public governance.
The significance is the reach, not simply what is being reached for. Higher education is already operating in a period when federal action increasingly intersects with matters institutions have historically understood as central to their own governance. AI introduces another possibility: institutions may spend years determining how they want to use and govern a technology while decisions made outside the institution alter which capabilities are available, what restrictions accompany them and how much discretion institutions retain. That possibility changes the institutional readiness question. Colleges and universities are not simply preparing to govern AI. They are preparing to do so inside an external governance architecture that is itself unsettled. What exactly are institutions preparing themselves to govern? And how much of that governance will ultimately remain theirs to determine?
Governance for a moving object
None of this means institutions should stop their current AI work until the technological or regulatory future becomes clear. There may be no point at which either becomes stable enough to make governance easy. Institutional readiness therefore cannot be measured by the completion of an AI policy or an implementation plan. Durable governance must remain useful as both the technology and the environment surrounding it change.
Institutional leaders might consider asking:
- Who has authority?
- Which decisions require human judgment?
- What may be delegated?
- How will automated action be evaluated?
- What happens when capability changes?
- Who retains the authority to withdraw permission?
These questions precede any particular AI tool. They were present when institutions began deciding whether students could use generative AI to write an essay. They carry greater weight when AI systems can access institutional resources, communicate beyond intended boundaries or take actions that produce harm before a human intervenes.
Higher education therefore has a stewardship responsibility in this moment. Institutions must make decisions about technologies that are changing while the boundaries around their use are still being negotiated. The familiar sequence in which a technology arrives, institutions experiment with it, policies develop and governance matures no longer quite holds. AI continues to change while that sequence is underway.
We are living through a peculiar moment in which society is simultaneously trying to welcome artificial intelligence through one door and determine whether to close another before it advances any farther.
Sophia Rahming, Ph.D., is an educational futurist, speaker, and consultant who helps colleges and universities design, scale, and evaluate STEM education initiatives while reimagining teaching, learning, and work in the age of artificial intelligence. She is the editor of Black Sisterhoods: Paradigms and Praxis and the author of The Girl Who Loved Math: The Story of Euphemia Lofton Haynes and The Unfinished Days: A 90-Day Practice for Burnout, Healing, and Beginning Again Rooted in Wabi-Sabi. She writes as a private citizen, and the views expressed are her own.
This article is the first part of a two-part series.
