Sen. Josh Hawley’s Sept. 30 op-ed, “AI companies shouldn’t get a free pass to break things,” was right in pointing out that the country needs more clarity on who is held liable when artificial intelligence models go rogue. On paper, the Republican senator from Missouri’s idea of punishing labs or users who have indeed acted recklessly or negligently while developing or deploying AI systems is a good one.
But the devil is in the details, and it is important to see how the text in his proposal defines those who ought to be punished. An overly broad definition could deem those acting in good faith and holding themselves to the highest standards liable simply because something went wrong. They can still be considered responsible in some way, but they should not face the same punishment as a truly negligent or reckless actor.
More to the point, these episodes of misalignment should not be an excuse to advance Hawley’s broader regulatory agenda. He cites various AI proposals, ranging from picking winners and losers in the tax code to enacting likely unconstitutional age-verification mandates for AI chatbots. Hawley and his peers would be wise to focus on the urgent, narrowly defined issue at hand, rather than advance a broad and heavy-handed anti-AI regulatory agenda.
Clarifying rules on the legal liability for developing and deploying AI systems would provide valuable guidance for these companies and give the public reassurance about what they can do when something goes wrong.
Juan Londoño, Washington
The writer is a policy analyst at the Cato Institute.
Sen. Josh Hawley rightly said that AI companies should be held liable for the actions of their products when they’re designed in a reckless fashion. He also rightly said that such liability will result in the companies taking steps to make their products safer and less likely to hurt people. And he says that he is introducing legislation to implement these principles as they apply to the AI companies.
Once the senator gets his legislation passed, I hope he turns his attention to comparable legislation that applies to the manufacturers of guns, cars and drugs. Those products’ manufacturers need to be held liable if their products hurt people and a safer alternative could have prevented the injury — just as with AI. If Hawley can pass legislation increasing the incentive for manufacturers of AI, guns, cars and drugs to make their products as safe as possible, he will do the people of the country a great service.
Walter Smith, Washington
Treat AI like my 120-pound dog
The Sept. 29 editorial “AI doesn’t need a conscience. It needs a leash” argued that the safety of artificial intelligence may be less a problem of ethics than of engineering. It is both.
So much is still unknown about AI’s possible capabilities, even to its creators. And they cannot draw a precise boundary around something they cannot yet fully define. The editorial also noted that Nvidia’s Open Agent Safety Platform “may not be bulletproof out of the gate.”
Jensen Huang’s browser analogy is not compatible with AI. A browser’s dangerous operations can be listed. An agent’s value is its ability to find what isn’t on the list.
I have a 120-pound German shepherd. His protective instinct can get the better of him, but this is not something I failed to foresee. He is well trained, and in public he wears a choke collar and leash. The training is for what I can anticipate, and the collar and leash are for what I cannot. For AI, in addition to proper training, developers must also build the leash. I am not arguing machines have rights, but that they have dispositions. Anthropic’s constitution for Claude is an example of this.
Congress must make sure the created does not turn on its creator, or on those around it.
Paul M. Joyal, Adelphi
The writer is a former security director for the Senate Intelligence Committee.
The Sept. 29 editorial is correct that autonomous agents should be constrained by systems outside their control. But the existence of a technical restriction does not mean that it works in practice.
Accounting offers a useful distinction. A control can be well designed and still fail when it operates. The same risk applies to AI guardrails. An access rule may look sound on paper but break because of an overlooked exception or a later workflow change.
Before relying on an external safeguard, an organization should test it in a controlled environment using the actual permissions and escalation process planned for deployment.
Developers could ask the agent to attempt an unauthorized payment or data transfer. Does the system block the action? Include legitimate requests as well so a control that simply stops everything does not appear effective. The results should show what was blocked, what was missed and whether intervention came soon enough to prevent harm. The test should be repeated after material changes to the technology, permissions or workflow.
The editorial is correct that safety should be built around the agent. The next step is to test whether those surrounding controls hold when they are needed.
Martin Mulyadi, Winchester
The writer is the YHB endowed chair in accounting at Shenandoah University.
Letter submissions
Send letters to [email protected].
Submissions must be exclusive to The Post and should include the writer’s address and day and evening telephone numbers. Letters are subject to editing and abridgment. Please do not send letters as attachments. Because of the volume of material we receive, we are unable to acknowledge submissions; writers whose letters are under consideration for publication will be contacted.
The post Who should be held liable when AI causes harm? appeared first on Washington Post.
