Contributor: AI isn't just standing by. It's doing things — without guardrails

Just 2 and a fractional years aft OpenAI stunned the satellite with ChatGPT, AI is nary longer lone answering questions — it is taking actions. We are present entering the epoch of AI agents, successful which AI ample connection models don’t conscionable passively supply accusation successful effect to your queries, they actively spell into the satellite and bash things for — oregon perchance against — you.

AI has the powerfulness to constitute essays and reply analyzable questions, but ideate if you could participate a punctual and person it marque a doctor’s assignment based connected your calendar, oregon publication a household formation with your recognition card, oregon record a ineligible lawsuit for you successful tiny claims court.

An AI cause submitted this op-ed. (I did, however, constitute the op-ed myself due to the fact that I figured the Los Angeles Times wouldn’t people an AI-generated piece, and too I tin enactment successful random references similar I’m a Cleveland Browns instrumentality due to the fact that nary AI would ever admit to that.)

I instructed my AI cause to find retired what email code The Times uses for op-ed submissions, the requirements for the submission, and past to draught the email title, draught an eye-catching transportation paragraph, connect my op-ed and taxable the package. I pressed “return,” “monitor task” and “confirm.” The AI cause completed the tasks successful a fewer minutes.

A fewer minutes is not speedy, and these were not analyzable requests. But with each passing period the agents get faster and smarter. I utilized Operator by OpenAI, which is successful probe preview mode. Google’s Project Mariner, which is besides a probe prototype, tin execute akin agentic tasks. Multiple companies present connection AI agents that volition marque telephone calls for you — successful your dependable oregon different dependable — and person a speech with the idiosyncratic astatine the different extremity of the enactment based connected your instructions.

Soon AI agents volition execute much analyzable tasks and beryllium wide disposable for the nationalist to use. That raises a fig of unresolved and important concerns. Anthropic does information investigating of its models and publishes the results. One of its tests showed that the Claude Opus 4 exemplary would perchance notify the property oregon regulators if it believed you were doing thing egregiously immoral. Should an AI cause behave similar a slavishly loyal employee, oregon a conscientious employee?

OpenAI publishes information audits of its models. One audit showed the o3 exemplary engaged successful strategical deception, which was defined arsenic behaviour that intentionally pursues objectives misaligned with idiosyncratic oregon developer intent. A passive AI exemplary that engages successful strategical deception tin beryllium troubling, but it becomes unsafe if that exemplary actively performs tasks successful the existent satellite autonomously. A rogue AI cause could bare your slope account, marque and nonstop fake incriminating videos of you to instrumentality enforcement, oregon disclose your idiosyncratic accusation to the acheronian web.

Earlier this year, programming changes were made to xAI’s Grok exemplary that caused it to insert mendacious accusation astir achromatic genocide successful South Africa successful responses to unrelated idiosyncratic queries. This occurrence showed that ample connection models tin bespeak the biases of their creators. In a satellite of AI agents, we should besides beware that creators of the agents could instrumentality power of them without your knowledge.

The U.S. authorities is acold down successful grappling with the imaginable risks of powerful, precocious AI. At a minimum, we should mandate that companies deploying ample connection models astatine standard request to disclose the information tests they performed and the results, arsenic good arsenic information measures embedded successful the system.

The bipartisan House Task Force connected Artificial Intelligence, connected which I served, published a unanimous study past December with much than 80 recommendations. Congress should enactment connected them. We did not sermon wide intent AI agents due to the fact that they weren’t truly a happening yet.

To code the unresolved and important issues raised by AI, which volition go magnified arsenic AI agents proliferate, Congress should crook the task unit into a House Select Committee. Such a specialized committee could enactment witnesses nether oath, clasp hearings successful nationalist and employment a dedicated unit to assistance tackle 1 of the astir important technological revolutions successful history. AI moves quickly. If we enactment now, we tin inactive drawback up.

Ted Lieu, a Democrat, represents California’s 36th Congressional District.

Insights

L.A. Times Insights delivers AI-generated investigation connected Voices contented to connection each points of view. Insights does not look connected immoderate quality articles.