Sam Altman Calls on Senate to License Most Powerful AI
OpenAI CEO Sam Altman has called on the U.S. Senate to create a dedicated agency and require licenses for the most advanced AI models. The proposal opens a debate over how to control real risks without shutting new competitors out of the market.
On May 16, 2023, Sam Altman testified before a US Senate subcommittee; his proposal should be read as testimony from an interested party, not as legislation.
Sam Altman, CEO of OpenAI, told the U.S. Senate on Tuesday to create a federal agency dedicated to artificial intelligence and establish licenses for the most powerful systems. The request matters because it comes from the company behind ChatGPT and GPT-4, two products that have brought generative AI to the center of public debate in just a few months.
Altman appeared before the Senate Subcommittee on Privacy, Technology, and the Law, chaired by Democrat Richard Blumenthal. Gary Marcus, a researcher and critic of the current race to build ever-larger models, and Christina Montgomery, IBM’s chief privacy and trust officer, also testified.
A license for models that cross a threshold
OpenAI’s proposal does not call for licensing every program that uses artificial intelligence. Altman has backed a system aimed at frontier models: those with capabilities advanced enough to pose significant risks.
The proposed agency would define that threshold, require safety testing before deployment, and have the power to grant or revoke licenses. It could also mandate independent audits and set rules for how models are evaluated before being made available to companies or the public.
The problem is no small matter. Language models are systems trained on enormous amounts of text to predict the next word and generate responses. That technique allows them to write, summarize, program, and hold convincing conversations, but it can also produce false information with apparent confidence, help automate fraud, or facilitate the creation of disinformation campaigns.
Altman expressed concern that the industry could cause significant harm to the world. It is an unusual statement in a sector accustomed to presenting every advance as inevitable progress, although it comes as OpenAI has become one of the companies benefiting most from the technology’s expansion.
Regulation as a matter of safety and competition
The U.S. debate starts with a practical challenge: regulating AI without treating a spam filter, a medical diagnostic system, and a model capable of generating text, code, and images at scale as if they were the same thing.
A license limited to the most capable systems would, in theory, spare small companies, researchers, and developers using existing models from excessive paperwork. But drawing that line will be difficult. A model may not appear dangerous in an isolated test and yet acquire new capabilities when connected to external tools, databases, or software that allows it to take action.
There is also the question of who can comply with the rules. Large companies have legal teams, computing infrastructure, and the resources to put their systems through costly evaluations. Poorly designed regulation could strengthen the companies that already dominate the market, leaving less room for competitors and open projects.
During the hearing, Gary Marcus argued that companies should not decide for themselves when their products are safe. His position reflects a growing concern: internal testing and voluntary commitments can be useful, but they cannot replace public oversight when the consequences of a failure fall on third parties.
A Congress catching up after ChatGPT
The hearing took place six months after ChatGPT’s public launch and two months after GPT-4. The speed with which these tools have reached schools, offices, and digital services has exposed the fact that the United States lacks a specific federal law for AI.
The Biden administration has already published a framework of principles, the AI Bill of Rights, but it has no force of law. Various federal agencies can intervene in specific areas, such as consumer protection, competition, employment, or data protection, although none oversees general-purpose models comprehensively.
The Senate did not emerge today with a bill ready for passage. It has begun treating AI as infrastructure with potentially broad economic and social effects, rather than simply another technology product. The next step will be deciding whether to create a new authority, strengthen existing regulators, or combine both approaches—and, above all, whether it can move fast enough to keep pace with the technology it wants to oversee.
Testimony, proposal and law are not synonyms
The official hearing page identifies the date, witnesses and written documents. That record makes attribution possible: Altman represented OpenAI, Christina Montgomery represented IBM and Gary Marcus appeared as an academic. A sentence spoken before the Senate may influence an agenda, but it does not create an agency or require model licences. Binding law requires text, procedure, approval and a competent authority.
A careful reading separates diagnosis from institutional design. “Some systems can cause serious harm” is a risk assessment. “A licence should apply above a threshold” proposes a mechanism. The threshold, tests, regulator, penalties, appeals and relationship with sector rules still need definition. Without those components, the word licence sounds more precise than it is.
The witness's incentives belong in the analysis. A large laboratory may sincerely believe oversight is necessary while also being better placed than a small competitor to fund audits and compliance. The alignment of safety and economic interest does not invalidate the proposal; it requires assessment of entry barriers, open innovation and concentration.
How to test threshold-based regulation
Every threshold creates incentives around its boundary. If it uses model size alone, a more efficient technique may provide similar capabilities below it. If it relies on one test, developers may optimise for that test. If it follows use, a general model may change category when connected to tools or sensitive data. A durable design therefore combines measurements, context and periodic review.
Test a proposal against three cases: a frontier laboratory, a small company adapting a model and an organisation deploying it in a high-impact decision. Ask which duty falls on each actor, who possesses the necessary information and what cost is introduced. A rule aimed at the wrong actor can be demanding without being effective.
The transferable skill is to read a hearing as a map of attributable positions. Preserve the video or testimony, attribute each proposal and then find the bill that does—or does not—turn it into text. This prevents one company's preference being presented as technical consensus or settled law.
The primary record also reveals what was not said. If a report attributes a figure, exact quotation or endorsement of a measure to a witness, it should appear in testimony or the recording. When it exists only in a third-party summary, attribute that source or remove it. This habit prevents a complex hearing from shrinking over time into a memorable quotation that nobody can locate.
Finally, safety and competition should be measured together. A rule can reduce one risk while creating another by concentrating capability. Mature analysis does not choose one concern and erase the other: state the objective, measure avoided harm, count who is excluded and revisit the threshold as models, uses and observed costs change.
Traceability matters more than memory of a headline.
This article was produced with artificial intelligence under human editorial oversight.
