Dan Hendrycks: AI Risk Warnings and the SB 1047 Fight

Dan Hendrycks is an American artificial intelligence (AI) researcher who directs the Center for AI Safety (CAIS), a nonprofit based in San Francisco. He signed a 2023 statement that mitigating AI extinction risk should be a global priority. He is also listed as president of the CAIS Action Fund, which sponsored California's SB 1047 bill, and he advises AI companies, including xAI. Critics argued during the SB 1047 fight that his stake in the auditing company Gray Swan AI was a conflict of interest.

Who Dan Hendrycks Is and What He Signed

Dan Hendrycks is the director of the Center for AI Safety, a nonprofit in San Francisco, California. He is also a signatory of the Statement on AI Risk, a single sentence from May 2023. The full text reads: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."

The sources checked do not say how many people signed the statement. The statement's page on safe.ai carries the exact text and his signature, so that is the primary place to read it.

His title depends on which document you read. CAIS's About page, as of 21 October 2026, lists him as "Executive & Research Director." The nonprofit's Form 990 for fiscal year 2024 lists "Daniel James Hendrycks, Executive Director." For fiscal years 2022 and 2023, the same filings list his title as "President."

He is also listed as president of the separate Center for AI Safety Action Fund, a 501(c)(4) group. The filings show $0 compensation for that role in both the FY2023 and FY2024 filings.

He earned a B.S. from the University of Chicago in 2018. He earned a Ph.D. in computer science from UC Berkeley in 2022.

The bill his group sponsored, SB 1047, aimed to regulate frontier AI models. Governor Gavin Newsom vetoed it in September 2024. Hendrycks also advises AI companies, which the advisory section below covers. For the organization itself, see our explainer on the Center for AI Safety.

His Technical Record: GELU, MMLU and MATH

His research record is the part of his profile you can check against published papers. Three of his contributions have public arXiv records. They are the GELU activation function, the MMLU benchmark, and the MATH dataset.

GELU, a 2016 Activation Function

He was the main author of the 2016 paper introducing GELU, which stands for Gaussian Error Linear Unit. The paper is arXiv:1606.08415. Wikipedia says GELU is used in many later transformer models.

MMLU, a 57-Task Benchmark

He is first author of "Measuring Massive Multitask Language Understanding," the paper that introduced the MMLU benchmark. It was submitted to arXiv on 7 September 2020, and its latest version was revised on 12 January 2021. The abstract says "The test covers 57 tasks," including elementary mathematics, US history, computer science and law.

The co-authors are Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song and Jacob Steinhardt. Anyone who wants to confirm the dates and authors can open the arXiv record for the MMLU paper.

MATH, a Competition Mathematics Dataset

He is first author of "Measuring Mathematical Problem Solving With the MATH Dataset." It was submitted to arXiv on 5 March 2021 and revised on 8 November 2021. The abstract describes "12,500 challenging competition mathematics problems."

The co-authors are Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song and Jacob Steinhardt. Four of those names, Burns, Basart, Song and Steinhardt, also appear on the MMLU paper. That overlap is easy to check on the two arXiv pages.

Other Published Work

Wikipedia lists earlier work including "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks" (arXiv:1610.02136, ICLR) and "Deep Anomaly Detection with Outlier Exposure" (arXiv:1812.04606, ICLR 2019).

He also co-authored "What Would Jiminy Cricket Do? Towards Agents That Behave Morally" (NeurIPS 2021, arXiv:2110.13136).

Wikipedia credits him as the creator of the large language model (LLM) benchmark "Humanity's Last Exam," developed with Scale AI. Wikipedia doesn't give a date for it, and the sources checked don't either.

His Advisory Roles at AI Companies

Hendrycks advises AI companies, and Fortune reported both roles in November 2024. Fortune says he has been a safety adviser to xAI, Elon Musk's AI company, since Musk founded it in 2023.

Fortune also reported, citing his LinkedIn profile, that he had become an adviser to Scale AI. Fortune said Scale AI is valued at $14 billion. Scale AI did not respond to Fortune's request for comment.

What He Has Said About the Pay

Hendrycks told Fortune the Scale AI role pays "12x my xAI role." He described that as $12 a year instead of $1. On equity he said: "As with other advisory roles, I don't have any equity."

Wikipedia gives a different account of the Scale AI pay. It says he takes a symbolic one-dollar salary from xAI and holds no company equity. It describes that arrangement as a way to avoid "any potential conflicts of interest." It dates the Scale AI advisory role to November 2024 and also describes it as a one-dollar salary. The two accounts of the Scale AI pay don't match, and the sources checked don't resolve the gap.

Pay From the Nonprofit

The Form 990 for CAIS shows his compensation from the nonprofit. It lists $119,465 for fiscal year 2022, $213,641 for fiscal year 2023 and $314,534 for fiscal year 2024. Over the same three years, the nonprofit reported total revenue of $6,660,060, $16,089,191 and $10,235,085.

Those figures come from CAIS's tax filings on ProPublica. The xAI and Scale AI pay figures come from Fortune and Wikipedia, not from these filings.

What Hendrycks Argues

His published work argues that advanced AI could cause catastrophic harm. The 2023 statement is a one-sentence version of that view. His longer papers sort the risks into categories. The wider debate over these claims is covered in our explainer on whether AI is dangerous.

The Statement's Core Claim

The statement places AI extinction risk alongside pandemics and nuclear war as a societal-scale risk. It asks that the risk be treated as a global priority. Our AI doomers explainer covers the debate over that framing.

Four Categories of Catastrophic AI Risk

His 2023 paper "An Overview of Catastrophic AI Risks," written with Mantas Mazeika and Thomas Woodside, sorts AI risk into four categories. They are malicious use, AI race dynamics, organizational risks and rogue AI agents. For the response side of the last category, read our guide on how to stop rogue AI.

Natural Selection and Agent Goals

In March 2023 he published a paper arguing that natural selection and competitive pressures could shape the goals of artificial agents. The paper is arXiv:2303.16200, dated 28 March 2023.

Policy Work and the SB 1047 Bill

In February 2022 he co-authored recommendations to the National Institute of Standards and Technology (NIST) on managing AI risks. The CAIS Action Fund, where he is listed as president, sponsored California's Senate Bill 1047 (SB 1047). Its full name is the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act. State Senator Scott Wiener introduced it on 7 February 2024.

The bill passed the state Senate 32 to 1 on 21 May 2024. It passed the Assembly 48 to 16 on 28 August 2024. Governor Gavin Newsom vetoed it on 29 September 2024. The criticism section below covers the reasons the bill drew opposition.

Other Papers and a Textbook

He also wrote a 2022 paper on how AI research itself affects societal risks. Wikipedia's text dates it to September 2022, but its own reference list gives 13 June 2022 for the arXiv version. The page doesn't explain the difference.

In 2024 he wrote the textbook Introduction to AI Safety, Ethics, and Society. A Coursera course called Introduction to AI Safety is attributed to him, with a 2026 date.

The Strongest Criticism of Hendrycks

One criticism in the sources checked is a conflict-of-interest argument tied to SB 1047. It rests on his ties to AI companies and to Gray Swan AI, which develops tools to audit AI model risks. Critics said the bill would have required affected companies to use such auditing tools.

The Conflict-of-Interest Argument

During the SB 1047 debate, critics said Hendrycks had a conflict of interest because he was an investor in Gray Swan AI. Fortune reported the criticism.

Pirate Wires, a tech-industry outlet, made the sharpest version of the argument in June and July 2024. It said what it called CAIS's lobbying arm co-sponsored SB 1047 while Hendrycks had co-founded Gray Swan. It also said Gray Swan's tools appeared positioned to supply the auditing data the bill would require.

His Response and Gray Swan's

Fortune reports that Hendrycks responded by posting on X that he would divest from Gray Swan and continue as an unpaid adviser. Fortune doesn't give a date for that post or for the divestment, and the sources checked don't give one either.

Gray Swan's own response says it does not hold any position on SB 1047. It says its products "are not and will not be designed to satisfy any of the auditing requirements proposed in the bill." That answers the auditing link, but it doesn't describe his stake.

The sources also differ on his role. Fortune says investor. Pirate Wires says co-founder. The size of any stake, before or after divestment, is not published in the sources checked.

His own record has answers on the other side too. He said he has no equity in his advisory roles, and he said he would divest from Gray Swan. Gray Swan says its products are not built to meet the bill's auditing rules. Those answers are on the record, though the divestment date and the size of the stake are not.

Opposition to the Bill

The bill drew named opposition. Fei-Fei Li argued in Fortune on 6 August 2024 that it would harm the US AI ecosystem. Andrew Ng called for narrower measures, and Yann LeCun of Meta argued that it would kill open-source AI models. Nancy Pelosi, Zoe Lofgren and Anna Eshoo were among the congressional opponents, and Andreessen Horowitz campaigned against it.

Newsom's veto message argued that the bill targeted models by compute size alone. It said the bill "does not take into account whether" a system is deployed in high-risk environments. It warned the bill could "give the public a false sense of security."

For where he sits in the wider debate, see our guide to who is fighting AI.

The Newsletter Disclosure Dispute

A reader of the CAIS newsletter argued that it should have disclosed his xAI adviser role when writing about Musk and SB 1047. CAIS replied that it "didn't bother since I assumed the ~all readers were aware." The exchange is in the newsletter's comments.

What Stays Unsettled

Several points on this page are not settled by the sources checked.

• The date he divested from Gray Swan AI is not published.
• Sources differ on whether his Gray Swan role was an investment, a co-founding stake or both.
• The size of any Gray Swan stake is not published.
• Fortune quotes his Scale AI pay as $12 a year, while Wikipedia describes a one-dollar salary.
• No date is published for the "Humanity's Last Exam" benchmark.
• Wikipedia dates the 2022 societal-risk paper two different ways.
• Roles change. The CAIS title is as of 21 October 2026, and the xAI and Scale AI roles are as Fortune reported them on 13 November 2024.

The filings give these legal details for the two nonprofits. CAIS files as "Center For Artificial Intelligence Safety Inc," EIN 88-1751310, a 501(c)(3) in San Francisco, with IRS exemption issued in January 2023. The CAIS Action Fund files as a 501(c)(4) in San Francisco, EIN 93-2442608, with exemption issued in October 2023. The checks found no separate registration for Hendrycks himself, and nothing further was verified.

Before you decide how much weight his advocacy deserves, read the statement text and the Fortune report on his adviser roles. Then check the CAIS Action Fund filings on ProPublica for the unpaid president role.

Sources

• Wikipedia: Dan Hendrycks
• arXiv: Measuring Massive Multitask Language Understanding (MMLU paper)
• arXiv: Measuring Mathematical Problem Solving With the MATH Dataset
• Center for AI Safety: About
• Center for AI Safety: Statement on AI Risk
• ProPublica: Center for Artificial Intelligence Safety Inc filings
• ProPublica: Center for AI Safety Action Fund filings
• Fortune: Dan Hendrycks, xAI, Scale AI and Gray Swan (13 November 2024)
• Pirate Wires: SB 1047 and Dan Hendrycks's conflict of interest
• Gray Swan AI: response on SB 1047
• Wikipedia: Safe and Secure Innovation for Frontier Artificial Intelligence Models Act
• CAIS AI Safety Newsletter #44: comments