Sam Altman testifies before the US Senate on AI oversight

Background

The session was held by the Senate Judiciary subcommittee on Privacy, Technology and the Law, under the title "Oversight of AI: Rules for Artificial Intelligence"^[1]. It was chaired by Senator Richard Blumenthal, who described it as the first in a series of hearings^[1]. That framing shaped how the session was read: not as a single inquiry but as the opening of a longer examination of what rules, if any, Congress should write for AI.

What happened

On 16 May 2023, Sam Altman, chief executive of OpenAI, testified before the subcommittee^[1]. It was his first appearance before Congress^[1][2]. Two other witnesses appeared alongside him: Christina Montgomery, IBM vice-president and chief privacy and trust officer, and Gary Marcus, a former New York University professor^[1].

Blumenthal opened the hearing by playing a speech written by ChatGPT and read aloud in a voice clone resembling his own^[2]. The device put the subject on the table before any witness had spoken: the technology under discussion could draft remarks and imitate a voice.

Key details

Altman's central argument was that governments should step in. "We think that regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models," he told the subcommittee^[1]. Asked about the dangers, he said his worst fear was that the field and the technology industry would cause significant harm to the world^[1].

He offered a mechanism rather than a slogan: government could require licences and testing for the development and release of AI models once their capabilities passed a threshold^[1]. Drawing the line by capability rather than by product means the rules would attach to how strong a model is, and licensing and testing would be the conditions attached to crossing that line.

He also flagged a political risk. The potential for AI to manipulate voters and to target disinformation was among his areas of greatest concern, especially with an election due the following year^[1].

At the same time he resisted the framing of AI as an autonomous actor. GPT-4, he said, is a tool, not a creature, and people have great control over it^[1]. The distinction matters for regulation: a tool implies an operator who can be licensed, audited or held responsible, whereas a creature suggests something that acts on its own and is harder to pin down.

Impact

By its chair's account the hearing was the first in a series^[1], so it functioned as a starting point rather than a verdict. It put two things on the record: a request for regulation coming from the head of a leading AI company, and a specific design for it — licences and testing keyed to capability thresholds^[1]. Because a corporate privacy officer and an academic testified in the same session^[1], the discussion was not left to a single vantage point. The opening itself, a ChatGPT-written speech in a clone of the chair's voice^[2], became part of the argument: it showed both what generative models can do and why some people are uneasy about them.