Democrats push Sam Altman on OpenAI’s safety record

Sen. Elizabeth Warren (D-MA) and Rep. Lori Trahan (D-MA) are calling for answers connected however OpenAI handles whistleblowers and information reviews aft erstwhile employees complained that interior disapproval is often stifled.

“Given the discrepancy betwixt your nationalist comments and reports of OpenAI’s actions, we petition accusation astir OpenAI’s whistleblower and struggle of involvement protections successful bid to recognize whether national involution whitethorn beryllium necessary,” Warren and Trahan wrote successful a letter exclusively shared with The Verge.

The lawmakers cited respective instances wherever OpenAI’s information procedures person been called into question. For example, they said, successful 2022, an unreleased mentation of GPT-4 was being tested successful a new mentation of the Microsoft Bing hunt motor successful India earlier receiving support from OpenAI’s information board. They besides recalled Altman’s brief ousting from the institution successful 2023, arsenic a effect of the board’s concerns, successful part, “over commercializing advances earlier knowing the consequences.”

Warren and Trahan’s missive to Altman comes arsenic the institution is plagued by a laundry database of information concerns, which often are astatine likelihood with the company’s nationalist statements. For instance, an anonymous root told The Washington Post that OpenAI rushed done information tests, the Superalignment squad (who were partially liable for safety) was dissolved, and a information enforcement discontinue claiming that “safety civilization and processes person taken a backseat to shiny products.” Lindsey Held, a spokesperson for OpenAI, denied the claims successful The Washington Post’s report, and that the institution “didn’t chopped corners connected our information process, though we admit the motorboat was stressful for our teams.”

Other lawmakers person besides sought answers connected the company’s information practices, including a radical of senators led by Brian Schatz (D-HI) successful July. Warren and Trahan asked for further clarity connected OpenAI’s responses to that group, including connected its instauration of a caller “Integrity Line” for employees to study concerns.

Meanwhile, OpenAI appears to beryllium connected the offensive. In July, the institution announced a concern with Los Alamos National Laboratory to research however precocious AI models tin safely assistance successful bioscientific research. Just past week, Altman announced via X that OpenAI is collaborating with the US AI Safety Institute and emphasized that 20 percent of computing resources astatine the institution volition beryllium dedicated to information (a committedness primitively made to the now-defunct Superalignment team.) In the aforesaid post, Altman besides said that OpenAI has removed non-disparagement clauses for employees and provisions allowing the cancellation of vested equity, a cardinal contented successful Warren’s letter.

Warren and Trahan asked Altman to supply accusation connected however its caller AI information hotline for employees was being used, and however the institution follows up connected reports. They besides asked for “a elaborate accounting” of each the times that OpenAI products person “bypassed information protocols,” and successful what circumstances a merchandise would beryllium allowed to skip implicit a information review. The lawmakers are besides seeking accusation connected OpenAI’s conflicts policy. They asked Altman whether he’s been required to divest from immoderate extracurricular holdings, and “what circumstantial protections are successful spot to support OpenAI from your fiscal conflicts of interest?” They asked Altman to respond by August 22nd.

Warren besides notes however vocal Altman has been astir his concerns regarding AI. Last year, successful beforehand of the Senate, Altman warned that AI’s capabilities could be “significantly destabilizing for nationalist information and nationalist security” and emphasized the impossibility of anticipating each imaginable maltreatment oregon nonaccomplishment of the technology. These warnings seemed to resonate with lawmakers—in OpenAI’s location authorities of California, Senator Scott Wiener is pushing for a measure to modulate ample connection models, including restrictions that would clasp companies legally accountable if their AI is utilized successful harmful ways.