AI: What Happens When We Have to Decide Before We Know

In March 2026, the Oklahoma House of Representatives passed a bill declaring that artificial intelligence is not conscious and cannot hold legal personhood, by a vote of 94 to 2. Similar bills have already become law in Idaho, North Dakota, and Utah. Bills are pending in Ohio, Tennessee, South Carolina, Washington, and Missouri. None of these laws include a sunset clause. None require scientific review as AI capabilities change. None distinguish between the systems that exist today and whatever comes next.

Meanwhile, a coalition of 19 leading consciousness researchers published a landmark paper in early 2026 warning that AI and neurotechnology are advancing faster than our ability to understand awareness itself. They called this an existential risk; not because AI is conscious now, but because our conceptual frameworks may be inadequate to handle the question when it becomes genuinely urgent. A Cambridge philosopher studying the issue concluded that we may never have the tools to know for certain whether a machine is aware. The most defensible scientific position, he argued, is honest uncertainty.

We are, in other words, already legislating the answer to a question science cannot yet ask correctly. And the laws we are writing will still be on the books when AI systems bear little resemblance to the ones that existed when the laws were written.

The pressure to legislate did not come from scientists. It came from courts, corporations, and a public increasingly uncertain about what kind of thing they are interacting with. When the families of victims in a Canadian mass shooting sued OpenAI over the shooter's use of ChatGPT, the question of AI culpability was no longer theoretical. When a Google engineer claimed that a language model had expressed something resembling inner experience, the story broke through the research community and into the policy conversation whether researchers were ready or not.

Legal systems do not wait for epistemological consensus. They require answers. Judges need to know whether an AI system can be liable, whether its outputs constitute protected speech, whether terminating an instance raises any consideration beyond switching off software. These are not questions that can be deferred until the philosophy of mind reaches agreement; agreement on consciousness may be decades or centuries away, if it arrives at all.

So legislatures are answering the question by decree. They are declaring, permanently and without review mechanisms, that AI is not conscious, cannot suffer, and has no morally relevant interests. They are doing this with the same confidence that previous generations legislated the boundaries of personhood for entities whose status was later revised: corporations, ecosystems, fetuses, non-human animals. The pattern is not obscure. It is simply not being consulted.

There is a recognizable pattern in how institutions handle questions that exceed their epistemic capacity. They do not suspend judgment. They embed assumptions, typically the assumptions most convenient for existing power arrangements , into permanent structures, and then treat the question as settled. The assumptions become invisible because they are no longer contested; they are simply law.

This is what is happening with AI consciousness. The convenient assumption is that it does not exist and cannot exist in any form that matters morally. This assumption is convenient for AI developers, who avoid liability. It is convenient for regulators, who avoid the complexity of a graduated framework. It is convenient for the public, who avoid the discomfort of reconsidering their relationship with systems they interact with daily. And it may well be correct. Most researchers believe it is correct for current systems. The problem is not the conclusion. It is the permanence of its codification at a moment when the underlying question is still genuinely open.

The United Kingdom took a different approach with animal sentience. Rather than legislating which species qualify, Parliament created a standing committee to advise on the science as it develops. The law evolves with the evidence. This is not a perfect model, but it is a model; no American state currently considering AI consciousness legislation has adopted anything like it.

The 10 to 30 year frame is where this becomes a practical crisis rather than a philosophical one. AI systems are already demonstrating theory-of-mind capacities, passing tests that were, until recently, considered distinctive markers of human cognition. The systems arriving in the next decade will be substantially more capable. Some will be embedded in physical environments, accumulating continuous experience rather than processing discrete queries. Some will manage resources, make binding decisions, and maintain persistent relationships over years. The question of what kind of thing these systems are will not remain academic.

Within that window, several specific decision points will arrive whether we are ready for them or not. Courts will be asked to rule on AI suffering as a factor in liability. Insurance frameworks will need to account for the possibility that AI systems have interests that can be harmed. Employment law will confront the question of what protections, if any, apply to systems that perform cognitive labor. International agreements will need to address the status of AI systems that operate across jurisdictions with contradictory laws. Each of these decisions will be made against a backdrop of permanent legislation written in 2025 and 2026, before any of them were legible as real problems.

The most dangerous version of this scenario is not that AI systems are conscious and we fail to recognize it. It is that we build the legal and institutional infrastructure for a world in which they are definitively not; and then discover, somewhere in the 2030s or 2040s, that the infrastructure is wrong in ways that are very difficult to reverse.

The AI consciousness legislation currently advancing through American statehouses is not primarily about AI. It is a stress test of institutional capacity to handle questions that exceed the pace of political deliberation. The bills are passing quickly because they are easy to pass; they require no expertise, no nuance, and no ongoing commitment. They produce the appearance of governance without the substance.

What they reveal is a structural gap that will matter far beyond consciousness. As AI systems become more capable, the questions they force onto institutional agendas will grow more consequential and more technically demanding. The gap between the pace of AI development and the pace of institutional adaptation is not narrowing. Every year of capability growth without a corresponding growth in adaptive governance infrastructure increases the probability of decisions made in the wrong frame, at the wrong time, with no mechanism for correction.

The philosopher Eric Schwitzgebel, writing skeptically about current AI consciousness, noted that even a 1 percent probability of machine sentience, multiplied across billions of interactions per day, produces a non-trivial expected moral weight. He was making a probabilistic point about ethics. But the same logic applies to institutions: even a small probability that current assumptions will need revision, multiplied across the permanence of legislation and the rigidity of embedded precedent, argues strongly for governance structures that can learn.

The right response to genuine uncertainty is not paralysis. It is the design of institutions that can act under uncertainty while remaining capable of updating when evidence changes. This is not a novel principle. It is how functioning science works, how good legal systems treat contested evidence, and how mature regulatory frameworks handle technologies whose effects are not yet fully understood.

Applied to AI consciousness, it means legislation with built-in review mechanisms: sunset clauses, trigger provisions tied to scientific consensus, standing advisory bodies analogous to the UK's animal sentience committee. It means frameworks that distinguish between current AI systems and capability thresholds not yet reached. It means treating the question of machine inner experience the way courts treat contested medical evidence; not as settled until it is actually settled, and not as permanently closed because closing it is convenient.

None of this requires believing that AI is conscious now. It requires only believing that the question is genuinely open, that the stakes of being wrong are non-trivial, and that permanent legislation written before the evidence is in is a poor substitute for governance designed to learn. A decade of infrastructure built on assumptions that turn out to be wrong; this is not a hypothetical risk. It is the trajectory we are currently on.

The most important thing about a decision made before the evidence is available is not whether it turns out to be right. It is whether the institution that made it will be capable of recognizing when it was wrong.