Experts and analysts don't agree. The AI race can't be controlled, but it can be responsibly governed if we permit open discussion on the subject.

Anyone claiming that “the experts” have reached a consensus about artificial intelligence should spend a week listening to the experts.

They agree about remarkably little.

Former President Barack Obama says AI-doomsaying is not overhyped. Although he stops just short of predicting killer robots, he believes increasingly autonomous systems could pursue objectives that diverge from human intentions. 

Is it possible?

Former Anthropic researchers assign disturbingly high probabilities to catastrophe. Anthropic CEO Dario Amodei wants frontier laboratories to slow down and coordinate with government — apparently while America’s geopolitical friends and enemies pursue breakneck access to this “dangerous” new tech without leadership from the worlds most successful democracy, home of Silicon Valley, and birthplace of AI.

Should we step back?

Nvidia CEO Jensen Huang takes nearly the opposite position. He says predictions of imminent extinction are unsupported and warns that laboratories demanding new regulation may really be seeking relief from existing liability. If an AI company negligently allows its agents to invade another company’s systems, Huang asks, why invent an entirely new regulatory regime before enforcing cybersecurity and product-liability laws already on the books?

He might be right.

Palantir CEO Alex Karp follows that reasoning toward an even more provocative conclusion. If an AI company sincerely believes its creation carries a meaningful chance of destroying humanity, how can it remain an ordinary private corporation? A company cannot easily tell investors that its product may end civilization while proceeding toward a trillion-dollar public offering. Karp suspects that AI laboratories are leading government toward nationalization, whether they admit it or not.

He has a point.

Writer Robert Malone raises the spectre of another dire possibility: regulatory capture. The largest AI companies can afford federal licenses, compliance departments, mandatory evaluations and armies of lawyers. Smaller competitors cannot. Rules ostensibly designed to restrain Big Tech could protect its dominant companies by making competition prohibitively expensive.

He’s right, of course.

Then there is antiquity scholar Zachary Porcu, who sees AI through humanity’s oldest stories about forbidden knowledge. From Eden and Prometheus to Faust, the golem and the genie, civilizations have warned about acquiring immense power without understanding its price. Porcu wonders whether AI is, metaphorically or even literally, a deal with the devil.

Unconvincing, if interesting. A language model invoking pagan gods reflects its training material and its user’s prompting, not proof of spiritual possession. His metaphorical concern is much stronger: AI offers convenience in exchange for intellectual independence, instantaneous answers in exchange for patient learning and simulated companionship in exchange for human relationships. The devil need not inhabit the machine. He may reside in the temptation to believe power comes without consequences.

In this way, Porcu sounds just like Socrates, who believed that books were a threat to human creativity and independent thought. How could anyone learn to think for themselves when they could just learn the secrets of philosphy from others?

He was right, of course. As all tech naysayers.

And he was wrong. 

If books cut off one area of indpendent human thought and creativity, they opened other areas of which Socrates had no inkling.

But these new disagreements over AI expose the central problem with demands that everyone “trust the experts.” Which experts?

The researchers building frontier models do not agree with one another. Executives have enormous and conflicting financial interests. Safety advocates may sincerely fear catastrophe while supporting regulations that strengthen the companies funding them. Accelerationists may correctly identify the danger from China while profiting handsomely from continuous expansion.

Even the terminology remains unsettled. 

Obama says machines are beginning to teach themselves, but “self-learning” can mean several different things. Models can revise their answers, create synthetic training data, play against themselves and assist researchers developing better systems. None is quite the same as a machine autonomously designing a superior successor without meaningful human direction.

The uncertainty is not an argument for doing nothing. It is an argument for intellectual humility and open debate.

Government should begin with demonstrated dangers. High-risk agents should be isolated from unauthorized networks. External actions with serious consequences should require human approval. Companies should report major incidents, preserve records for independent investigators and remain liable for negligent containment. Existing cybersecurity, privacy, contract and consumer-protection laws should be enforced before Congress grants anyone immunity from them.

Any additional regulation should rise with demonstrated capability and available resources. A university researcher adapting an open model should not face the same compliance burden as a trillion-dollar laboratory training frontier systems on enormous computing clusters. Otherwise, Washington will consolidate American AI while Chinese open-weight models become the world’s default alternative.

America also cannot pretend it controls the global race. China and Russia will not stop because American experts sign a letter. Previously released model weights cannot be recalled from the internet. A unilateral American pause could sacrifice democratic leadership without eliminating the technology or its risks.

There are no perfect solutions here, only tradeoffs. Moving too quickly carries dangers. Moving too slowly carries different dangers, including dependence on systems developed by governments that do not share American values.

Responsible governance is possible, but only if disagreement remains permissible. We should distrust anyone who portrays a contested theory as settled science, whether predicting certain extinction or dismissing every danger as fantasy.

AI is too consequential to leave entirely to its creators. It is also too consequential to regulate through panic. The task is to preserve innovation, enforce responsibility and keep democratic societies in the lead — not by manufacturing consensus, but by allowing honest disagreement to guide better judgment.

Today and in the future.

(Contributing writer, Brooke Bell)