news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

According to foreign media reports, on August 3, after a whistleblower accused the U.S. Securities and Exchange Commission (SEC) that the confidentiality agreement of artificial intelligence company OpenAI illegally prevented employees from disclosing major safety issues to lawmakers, OpenAI is facing increasing pressure to prove that it has not concealed the risks of artificial intelligence.

In a letter to OpenAI yesterday, Sen. Chuck Grassley (R-Iowa) asked for evidence that OpenAI is no longer requiring agreements that could hinder its employees from making protected disclosures to government regulators.

Specifically, Grassley asked OpenAI to provide current employment, severance, nondisparagement, and nondisclosure agreements to assure Congress that the contracts would not hinder disclosure.

This is critical, Grassley said, so that we can rely on whistleblowers who expose emerging threats to help develop effective AI policy to guard against existing AI risks posed by technological advances.

Grassley's letter said he had apparently requested the records from OpenAI twice but had not received a response.

Clare Slattery, a spokeswoman for Grassley, told The Washington Post that OpenAI has not yet responded to the most recent request to send the documents.

"It is not enough to simply claim that updates were made," Grassley said in a statement provided to Ars. "The facts speak louder than words. Altman needs to provide records and respond to my oversight requests so that Congress can accurately assess whether OpenAI is adequately protecting its employees and users."

In addition to requiring OpenAI to update its employee agreement recently, Grassley urged OpenAI to be more transparent about the total number of requests from employees seeking federal disclosures since 2023.

The senators wanted to know what information the employees wanted to disclose to officials and whether OpenAI actually approved their requests.

Similarly, Grassley asked OpenAI to confirm how many investigations the SEC has launched into OpenAI since 2023.

Taken together, the documents will reveal whether OpenAI employees are still prohibited from making federal disclosures, what types of disclosures OpenAI denies, and the extent to which the SEC is monitoring OpenAI for hidden security risks.

OpenAI must ensure that its employees can provide protected disclosures without unlawful restrictions, Grassley wrote in the letter.

He has given OpenAI until August 15 to respond so that Congress can conduct objective and independent oversight of OpenAI's safety and confidentiality protocols.

OpenAI did not immediately respond to Ars' request for comment.

On X, Altman wrote that OpenAI has taken steps to increase transparency, including an agreement with the U.S. AI Safety Institute where we will provide early access to our next foundational model so that together we can advance the science of AI evaluation.

He also confirmed that OpenAI wants current and former employees to feel comfortable raising concerns.

"This is critical for any company, but especially for us, and is an important part of our safety program," Altman wrote. "In May, we removed the non-disparagement clause for current and former employees, as well as the clause that gave OpenAI the right to cancel vested equity (although it was never used). We have been working to correct this."

In July, the whistleblower told the SEC that OpenAI should be required to provide not only current employee contracts, but also all contracts containing nondisclosure agreements to ensure that OpenAI is not covering up historical or current practices of covering up AI safety risks. They want all current and former employees to receive notice of any contract containing illegal nondisclosure agreements and ask OpenAI to be fined for each illegal contract.

Sam Altman accused of dismissing AI safety concerns

However, calls for more transparency from OpenAI are not limited to lawmakers. The day before Grassley sent his open letter, OpenAI CEO Sam Altman released a statement on X claiming that OpenAI has invested sufficient resources in safety work since July last year, but X fact-checkers commented in a comment that the statement was misleading, sparking a public outcry.

Altman wrote: As we said last July, we are committed to allocating at least 20% of our computing resources to security efforts across the company, which sparked community attention and many biased X readers felt provided more context.

The community note states that Altman was referring to a July blog post that explicitly confirmed that OpenAI was allocating 20% ​​of its computing resources to its super alignment team. Since that team has now disbanded and the chief scientist in charge of super alignment has left the company, X commentators claim that Altman's statement misled the public about what computing resources were actually allocated to OpenAI's now vaguely defined safety work.

Some X commenters asked for more details about what OpenAI’s current safety work is about. In a July 31 letter responding to five other senators about AI safety, OpenAI provided a more detailed explanation of its safety work. The company also clarified Altman’s statement that X had flagged it as needing “more context” and wrote that the computing resources mentioned in the blog should never refer only to the super-aligned team.

OpenAI said in a letter to the senators: To further advance our safety research agenda, last July we committed to allocate at least 20% of our computing resources to AI safety over a multi-year period. This commitment always applies to safety work across the entire company, not just one specific team.

This is confusing to some, because OpenAI’s blog explicitly states that OpenAI will dedicate 20% of the computing resources we have received so far to solving the superintelligence coordination problem over the next four years. At the time, OpenAI said that its main basic research bet was its former superintelligence coordination team, “which needs the computing resources because doing this well is critical to achieving our mission.”

One X user @AISafetyMemes asked OpenAI why it was shifting computing resources away from super alignment and across the company if super alignment is by far the most important problem to address in order to protect against AI risks.

OpenAI has not yet responded to the meme account, but has previously said super-aligned teams are critical to AI safety because society needs scientific and technological breakthroughs to guide and control AI systems that are smarter than us.

OpenAI details safety initiatives

In OpenAI's letter to senators, Chief Strategy Officer Jason Kwon told lawmakers that OpenAI "prevents risks by working with external experts to evaluate and red-team our models. This included consulting with over 100 external experts who helped assess the risks associated with our latest model, GPT-4o."

Kwon wrote: The findings and security mitigations from red team testing are publicly available in the system cards that come with our models, and OpenAI has published research on “measuring the chemical, biological, radiological, and nuclear (CBRN) risks associated with AI systems,” estimating the extent to which different occupations and industries may be affected by language models, assessing the impact of language models on influencing operations, and explaining the explainability of AI system decisions.

To further ensure the safety of AI, Kwon said OpenAI will do iterative deployments, releasing new features such as the Sora video model or speech engine to a limited group before releasing them broadly to the public.

“This strategy enables us to get feedback from people outside of OpenAI, update our safety guardrails as needed, and give the public a peek into upcoming AI features before fully opening up our models,” Kwon wrote.

For employees working on new features or models who still worry about retaliation for raising safety concerns, Kwon said OpenAI launched an integrity hotline in March, a channel where employees can anonymously report concerns when they are not comfortable raising them through other existing channels.

Although OpenAI says employees are free to speak about any concerns, it's clear that if OpenAI believes certain information presents a security risk, it still can't be shared.

"OpenAI continues to distinguish between raising concerns and revealing company trade secrets," Kwon wrote. "The latter (protected disclosure rights) remains prohibited under nondisclosure agreements with current and former employees. We believe this prohibition is particularly important given the impact of our technology on U.S. national security."

Given that this exception allows OpenAI to deny some federal disclosures, Grassley will likely want to examine whether OpenAI might have improperly denied disclosures on the grounds of protecting trade secrets.

Grassley's letter made clear that the senators are concerned about the risks of companies policing their own safety.