The argument over artificial intelligence is becoming easier to hear and harder to evaluate. One side warns of extraordinary danger. Another dismisses those warnings as exaggeration or obstruction. Between them stand citizens, workers, business owners, parents, and public officials who must make decisions before every technical dispute is settled.
BBC News reports that President Trump called AI safety fears a hoax and rejected calls for greater safeguards. The same BBC News report on the AI safety dispute says Anthropic co-founder Jack Clark told the outlet that a mandatory AI kill switch may be needed.
Those positions are far apart, but the public should not have to choose between them as articles of faith. A responsible debate would turn broad claims into questions that can be answered, challenged, and revised.
Start by naming the harm
AI safety is too broad a phrase to guide policy by itself. It can refer to an unreliable answer in a consumer application, discrimination in an automated decision, exposure of private information, manipulation of critical systems, or a more speculative loss of human control. These risks differ in probability, severity, and remedy.
Any proposal for a safeguard should therefore begin with a plain description of the harm it is meant to prevent. Who could be injured? Through what chain of events? Would the damage be reversible? What evidence would show that the danger is increasing or decreasing?
The same discipline should apply to arguments against regulation. It is not enough to say that safeguards will slow innovation. Which safeguard would impose the burden? On whom? Would the cost fall mainly on a small developer, a large platform, a hospital, a bank, or a government contractor? Could a narrower requirement address the risk without blocking ordinary uses?
Separate emergency controls from everyday oversight
The image of a kill switch is powerful because it suggests a single moment when someone must stop a dangerous system. Yet public accountability usually depends on less dramatic work carried out earlier. That includes defining who may deploy a system, recording important changes, testing known failure modes, protecting access, reporting serious incidents, and assigning responsibility when something goes wrong.
An emergency shutoff may be sensible in some settings and impractical in others. A tool used by one company is not governed in the same way as software distributed across millions of devices. A system assisting with casual writing does not present the same stakes as one controlling physical equipment. The question should not be whether every AI system needs one universal switch. It should be what level of control is appropriate for a particular capability and use.
Demand an accountable decision maker
Every serious safeguard needs an answer to a basic civic question: Who decides?
If a private company alone determines when its system is safe, the public may reasonably worry about commercial incentives. If a government agency holds broad power to halt technology without clear standards, citizens may worry about arbitrary enforcement or political abuse. If no institution has authority, warnings can circulate without anyone being responsible for acting on them.
A workable framework would specify the decision maker, the evidence required, the scope of the action, and the process for review. It would also distinguish between confidential technical details and information the public needs in order to judge whether oversight is functioning. Transparency does not require publishing every line of code. It does require explaining the rules and reporting whether those rules were followed.
Prefer rules that can survive a change in power
AI policy should not depend on whether citizens trust one president, one executive, or one group of researchers. Durable safeguards must remain legitimate when political leadership changes and when the technology industry produces new winners and losers.
That favors clear definitions, limited authority, documented decisions, independent review, and regular reconsideration. It also favors different obligations for different levels of risk. A rule designed for the most powerful systems should not automatically burden a neighborhood business using a simple scheduling tool.
The country does not need to settle every question about AI before taking any action. Nor should it accept every proposed safeguard merely because the technology is unfamiliar. The proper standard is more demanding: identify the harm, test the claim, match the remedy to the risk, and make someone answerable for the result.
Labels such as hoax, catastrophe, progress, and safety may rally an audience. They do not govern a technology. Public tests, visible responsibility, and rules that can be corrected do.