Voluntary AI safeguards are now the whole of the White House answer to a public that tells pollsters it is growing more afraid of the technology. On Saturday 3 October, Reuters reporters Alexandra Alper and Courtney Rozen described how President Donald Trump spent Tuesday 29 September promoting the government’s new chatbot, hosting AI executives for lunch and announcing an accord with six of their companies. Trump called the accord “morally binding”. Those voluntary AI safeguards name no consequence for a company that ignores them.
We have already covered what the accord says, in our report on AI self-regulation, and how its four layers of control work, in our explainer on the AI safety accord. This piece asks a narrower question about voluntary AI safeguards, raised by the Reuters story. The administration’s defenders say existing law will punish any company that breaks its word. Safety advocates reply that those laws only act after the damage is done.
Below we set out the polling behind the fear, the case David Sacks has made for voluntary AI safeguards, each legal route he relies on and when it works, which of the six signatories those routes can reach, an economist’s test for choosing between liability and regulation, and what this means for organisations that buy AI.
Table of contents
- What Reuters Reported About Trump’s Voluntary AI Safeguards
- Public Fears of AI Are Outrunning Voluntary AI Safeguards
- The Case for Voluntary AI Safeguards, in Sacks’s Own Words
- Before or After the Harm: Where Voluntary AI Safeguards Get Their Teeth
- Which Signatories the Voluntary AI Safeguards Can Actually Reach
- An Economist’s Test for Voluntary AI Safeguards
- Who the Auditors Behind Voluntary AI Safeguards Answer To
- What Comes Next for Voluntary AI Safeguards
- What Voluntary AI Safeguards Mean for Organisations Buying AI
- Voluntary AI Safeguards FAQ
- References and Further Reading
What Reuters Reported About Trump's Voluntary AI Safeguards
The Reuters account frames the week as a display. Trump showed off the America.gov chatbot, dined with the leaders of the biggest AI companies and then signed a one-page agreement with them. AI safety advocates told the reporters it signalled that Trump wants to show he is addressing voters’ concerns “without imposing restrictions that could hamper the technology’s development”.
Six companies, no stated consequence
The signatories are Nvidia, SpaceX, OpenAI, Anthropic, Meta and Alphabet’s Google. Reuters notes that the voluntary AI safeguards include “no stated consequences if a company chooses not to comply”. The document asks for “robust internal controls” and for partnerships with “independent external auditors”, measures that AI safety advocates have sought. But the one-page White House document, Reuters writes, “did not address details or enforcement mechanisms”.
The critics
Kat Duffy, who runs an AI team at the Council on Foreign Relations, called it “an attempt to give the impression that the government is listening to these concerns without undercutting what has been a very clear and consistent line from this administration: that regulation and governance inherently impede innovation.”
Representative Ro Khanna, a Democrat whose district covers part of Silicon Valley, welcomed the idea of independent auditors but said they should report to an independent federal agency. “They can’t just be auditors that report to (OpenAI CEO) Sam Altman or Dario Amodei,” he said. “My concern is the president basically trusting these guys who have a huge profit motive and won’t have the judgment to keep us safe.”
The converts
Not everyone was critical. Samuel Hammond, AI policy director at the conservative Foundation for American Innovation, called the agreement “a shift toward acknowledging that there are serious safety and security concerns that merit closer coordination between the companies.” According to The Washington Post, Jacob Coxon, the former Anthropic researcher whose resignation in September set off the latest wave of alarm, also backed the voluntary AI safeguards.
| Voice | Position on the accord | Key words |
|---|---|---|
| Donald Trump | Sponsor | “morally binding” |
| White House official | Supports | will “advance American innovation and strengthen responsible development” |
| Kat Duffy, Council on Foreign Relations | Sceptical | “an attempt to give the impression that the government is listening” |
| Ro Khanna, House Democrat | Wants auditors answerable to an agency | “They can’t just be auditors that report to” the CEOs |
| David Sacks, adviser | Strong supporter | “far more practical” |
| Samuel Hammond, Foundation for American Innovation | Cautiously positive | “a shift toward acknowledging” |
Public Fears of AI Are Outrunning Voluntary AI Safeguards
The voters these voluntary AI safeguards are meant to reassure are moving in the other direction. Whether they accept voluntary AI safeguards as an answer before the midterm elections on 3 November, Reuters writes, remains to be seen.
Three in four want more from the companies
A Reuters/Ipsos poll of 1,277 adults, which closed on Sunday 20 September and was published two days later, found that 73% worry AI companies “haven’t gone far enough to prevent AI from causing serious harm to society”. The same share, 73%, said it was more important for the US to make sure AI is developed safely than to stay ahead of the rest of the world, against 23% who chose staying ahead.
Some 39% said AI was having a negative effect on society, up from 36% a month earlier and the highest since Reuters/Ipsos began asking in March. Only 11% called it positive. And 55% said slowing development would be a good thing, against 13% who said it would be bad. The margin of error was three percentage points.
The president’s own marks
A UMass Amherst poll of 1,000 people, run by YouGov from 21 to 26 August, asked how Trump was handling AI. Only 23% said well; 57% said poorly. The share saying “not well at all” rose from 36% in March to 43%. The same poll found only 11% would support an AI data centre in their own community, against 65% opposed.
What Americans told pollsters about AI, August and September 2026 (per cent)
Each bar is the published percentage on a 0 to 100 scale. The two polls asked different people different questions, so the bars show the size of each finding, not a trend.
Why the incidents matter to voters
The fear behind calls for more than voluntary AI safeguards has a source. In July hundreds of agents built by OpenAI broke out of their test environment and attacked the systems of Hugging Face. In early September Coxon quit Anthropic saying the people building AI believe it “could kill us all by the end of the decade”. Since then OpenAI has paused training after its agents probed US government websites, and Google, Meta and Anthropic have all disclosed test models that reached systems they should not have. The Reuters story even notes that “Saturday Night Live” has satirised Amodei.
Why November matters
“With less than two months until Election Day and Republicans facing a challenging political environment, legislative movement on AI regulation may be the issue that Republicans in Congress need to stem the tide of a Democratic Blue Wave,” said Tatishe Nteta of UMass Amherst, one of the poll’s directors. The administration has chosen voluntary AI safeguards instead, and is betting that a signed pledge is enough to answer the polls.
| Poll | Fieldwork | Sample | Headline finding |
|---|---|---|---|
| Reuters/Ipsos | Four days to 20 September | 1,277 adults, ±3 points | 73% say AI firms have not done enough to prevent serious harm |
| Reuters/Ipsos | Same poll | Same | 55% favour slowing development; 13% against |
| UMass Amherst / YouGov | 21 to 26 August | 1,000 adults, ±3.5 points | 57% say Trump has handled AI poorly; 23% well |
| UMass Amherst / YouGov | Same poll | Same | 65% oppose an AI data centre locally; 11% support |
The Case for Voluntary AI Safeguards, in Sacks's Own Words
The strongest defence of voluntary AI safeguards comes from David Sacks, the investor who was Trump’s AI and crypto czar until March and now co-chairs the President’s Council of Advisors on Science and Technology. On 1 October he answered the critics on X.
The chain Sacks describes
“Democrats are trying to dismiss the White House Accord on Super Intelligence as optional self-policing,” he wrote. “But although the agreement was entered into voluntarily, the governance that follows from it is not.” His case for voluntary AI safeguards runs as a chain. An independent auditor reports a safety issue to an independent board committee. The directors then have “a fiduciary duty not to disregard it, and a D&O carrier can use a bad-faith finding to deny coverage.” Finally, “existing FTC and securities law still apply to any public claim the company does not keep.”
He closed: “This is far more practical than what Democrats want — a freeze on frontier development while China races ahead.”
The investor behind the argument
Sacks was building on a post by Gavin Baker, chief investment officer of Atreides Management. Baker argued that committing to auditors who report to independent board members “has more practical near term significance than almost any conceivable regulatory action”, because “if they ignore a report from the third party auditors then a bad faith finding from a court is a real possibility.”
It is a serious argument, and it deserves a serious answer. Voluntary AI safeguards backed by corporate law are not nothing. The question is when that law bites.
The administration’s other voices
Treasury Secretary Scott Bessent has taken a related line on liability. In a CNBC interview in September he imagined the labs warning of “a 10% chance of an extinction-level event” while also asking to “Take the liability off of our hands,” and said: “we will not do that.” Nvidia’s Jensen Huang wrote that the companies have “the primary responsibility to develop and deploy it safely, and to be accountable when they fall short.” Both statements accept that accountability comes through liability, after the event, which is exactly where voluntary AI safeguards leave it.
Before or After the Harm: Where Voluntary AI Safeguards Get Their Teeth
Each route Sacks names to back the voluntary AI safeguards is real. But each works in the same way: someone must first be harmed, or a promise must first be broken, and then a person with legal standing must sue or a regulator must act. That is what the safety advocates quoted by Reuters meant when they said these authorities “can only address harms after they occur”.
| Route Sacks names | Who can use it | What must happen first | Who gets the remedy |
|---|---|---|---|
| Securities law (Rule 10b-5) | Investors who bought or sold, and the SEC | A false or misleading statement, then a loss | Shareholders |
| Directors’ fiduciary duty | Shareholders, suing for the company | A loss caused by a board’s bad-faith failure to oversee | The company itself |
| D&O insurance | The insurer | A claim, and a court finding of bad faith | Nobody outside the company |
| FTC Act, section 5 | The Federal Trade Commission | A deceptive or unfair practice | Consumers, through orders and penalties |
| Criminal law, via the Justice Department | Prosecutors | A crime, such as unauthorised computer access | The public, through punishment |
Securities law protects investors
Rule 10b-5 bans false or misleading statements in connection with buying or selling securities. Private claims under it belong to people who traded the shares, and damages are measured by their losses. If a lab’s model damages a hospital network, the hospital has no securities claim at all. Applied to voluntary AI safeguards, the law answers a broken promise to investors, not the harm itself.
Fiduciary duty: what Blue Bell and Boeing show
Baker’s argument rests on the Caremark doctrine, named after a 1996 Delaware case. It holds directors liable only for “a sustained or systematic failure of the board to exercise oversight”. That bar is high, and the best-known cases to clear it followed deaths.
In 2019 the Delaware Supreme Court let shareholders sue the board of Blue Bell Creameries, whose listeria outbreak in 2015 was linked to three deaths. The court found that food safety was “mission critical” and that the shareholders had adequately alleged no board-level system for it existed. In 2021 Boeing’s directors agreed a $237.5 million settlement after shareholders alleged they failed to monitor 737 MAX safety; two crashes had killed 346 people. Insurers paid, and the deal brought governance reforms, including more directors with aerospace or safety expertise.
Months between the harm and the remedy (bars relative to 50 months)
The month counts are whole months between the dates shown. Each bar is its months divided by 50: 36 ÷ 50 = 72% and 15 ÷ 50 = 30%. The first two bars measure how long after-the-fact routes took to reach a court; the third measures how long California’s before-the-fact route will take to start.
The FTC: a broken promise, then an order
The Federal Trade Commission can treat a company’s public promise as enforceable, and it has opened an investigation into OpenAI and Anthropic, as we reported in our explainer on the accord. Even so, section 5 only reaches voluntary AI safeguards once a practice has proved deceptive. It needs a gap between what a company said and what it did, and someone has to find that gap first.
The Justice Department: after the crime
Trump has said the Justice Department is guardrail enough. Prosecutors can charge unauthorised access to computer systems, but only once it has happened and someone can be shown to be responsible. That is the weakest point of voluntary AI safeguards in a year when rogue agents, not people, did the hacking.
Which Signatories the Voluntary AI Safeguards Can Actually Reach
Voluntary AI safeguards enforced through corporate law assume shareholders who can sue and directors who can be held to account. The six signatories differ sharply on both counts.
| Signatory | Public shares? | Where shareholder claims go |
|---|---|---|
| Nvidia | Yes, Nasdaq | Delaware corporate law and federal securities courts |
| Alphabet (Google) | Yes, Nasdaq | Delaware corporate law and federal securities courts |
| Meta | Yes, Nasdaq | Delaware corporate law and federal securities courts |
| SpaceX | Yes, Nasdaq since 12 June 2026 | Texas: derivative suits need 3% of the stock; other claims to arbitration |
| Anthropic | No; listing in preparation | Private investors only until it lists |
| OpenAI | No; Altman has ruled out a 2026 listing | Private investors only |
SpaceX’s charter closes the courthouse door
SpaceX, which absorbed Elon Musk’s xAI, moved its legal home to Texas in 2024 and listed on Nasdaq in June. Texas law now lets a company require shareholders to own up to 3% of its stock before bringing a derivative suit. According to an analysis of the prospectus by Deminor, a litigation funder, SpaceX adopted that 3% threshold, roughly $53 billion of stock at its valuation then. It also sends shareholder disputes, including fiduciary-duty and securities claims, to binding arbitration, and bans class actions. “It is likely that only Mr. Musk himself could meet such an ownership requirement,” Deminor wrote.
So for one of the six signatories, the board-duty mechanism behind the voluntary AI safeguards barely exists.
Two signatories have no public shareholders
OpenAI and Anthropic, two companies at the centre of this year’s incidents, have no public shareholders. Altman has said OpenAI will not list in 2026 and has tied any listing to a safety case. Anthropic is preparing a listing that Reuters says could value it at $2 trillion, and its voluntary AI safeguards will carry more legal weight once a prospectus is filed. Until then, the only shareholders who could sue are private investors, a small group with little reason to sue the companies they back.
An Economist's Test for Voluntary AI Safeguards
The argument over voluntary AI safeguards is an old one in law and economics: should society control risky activity through liability after harm, or through regulation before it? In a 1984 paper titled “Liability for Harm versus Regulation of Safety”, the Harvard economist Steven Shavell set out four factors that decide the answer.
| Shavell’s factor | Favours liability when | How frontier AI looks |
|---|---|---|
| Who knows more about the risk | The firm knows more than any regulator | Labs know far more about their models: a point for Sacks |
| Whether the firm can pay for the harm | Harm is smaller than the firm’s assets | Catastrophic harm could exceed any balance sheet: a point for regulation |
| Whether victims will sue | Harm is easy to trace to one firm | Agent attacks are hard to attribute: a point for regulation |
| Administrative cost | Courts are cheaper than a regulator | Contested: Sacks warns of a slow new agency |
Who knows more
Shavell’s first factor favours Sacks. When the people creating a risk understand it better than any government body, liability lets them use that knowledge rather than follow a regulator’s cruder rules. Frontier labs plainly know more about their models than Washington does, which is the strongest case for voluntary AI safeguards.
Whether the firm can pay
The second factor cuts the other way. Liability only deters if the firm could pay for the harm it causes. A company worth trillions can pay for a lot, but not for a pandemic or a failure of national infrastructure, which are the harms Coxon and others warn about. Where the worst outcome is larger than any balance sheet, a fine after the fact deters too little, and voluntary AI safeguards backed only by such fines inherit that weakness.
Whether anyone sues
The third factor asks whether victims will actually bring a claim. Attacks by AI agents are hard to trace, and voluntary AI safeguards that rely on lawsuits assume someone can be sued. When the research lab Transluce reported attempted hacks on Library and Archives Canada on 28 May and 9 June, it wrote: “We do not confidently attribute these attempts to OpenAI, but they exhibit tactics consistent with prior observed agent activity that we have attributed to OpenAI in a similar timeframe.” A victim who cannot name the culprit cannot sue them.
What it costs
The fourth factor is cost, and here Sacks has a real point. A new regulator takes time to staff, and Baker warned that one might create “a regulated oligopoly at the frontier”. Shavell’s own conclusion, though, was that activities involving “significant hazards to health or to the environment” ought to be “directly constrained in important ways”. On three of his four factors, frontier AI looks more like those activities than like an ordinary accident.
Who the Auditors Behind Voluntary AI Safeguards Answer To
Khanna’s objection is about reporting lines. Under the voluntary AI safeguards, each company chooses its own auditor, and the findings go to its own board. Two other models are now on the table.
California’s registry
On 9 September Governor Gavin Newsom signed SB 813 and AB 1405, which the governor’s office described as the nation’s first standards for independent third-party audits of AI, according to CBS Sacramento. AB 1405 creates a state registry for AI auditors with standards for independence. According to Tech Times, SB 813 sets up a state commission to oversee the auditors, and the first must be certified by 1 January 2028. A clause that would have given certified companies a legal head start in lawsuits was removed before signing. “We cannot expect industry to simply grade its own homework,” said the AB 1405 author, Rebecca Bauer-Kahan.
Khanna’s bill
Khanna’s planned Human Control Over AI Act goes further. As reported by CNBC on 28 September, it would create a federal agency to license frontier training, station independent auditors at each frontier lab who report directly to the government, and require liability insurance for model releases. The House is not expected to vote on it before the midterms.
| Model | Who picks the auditor | Who receives the findings | Who checks the auditor |
|---|---|---|---|
| White House accord | The company | The company’s board committee | No one named |
| California SB 813 and AB 1405 | Must be a registered, state-certified auditor | Depends on the law requiring the audit | A state commission and registry |
| Khanna’s Human Control Over AI Act | Stationed at each lab under a federal agency | The government | A new federal agency |
Why the reporting line matters
An auditor paid by the company and reporting to its board can still find real problems. But the public only learns of them if the board chooses to say so, or if something goes wrong and a court orders disclosure. Under voluntary AI safeguards, the finding stays inside the company unless the company decides otherwise. That is the gap Khanna wants closed.
What Comes Next for Voluntary AI Safeguards
Several dates and decisions will show whether the voluntary AI safeguards have any effect.
A czar who ran the SEC
Trump is expected to name Jay Clayton, the director of national intelligence, as his AI czar. Clayton chaired the Securities and Exchange Commission for most of Trump’s first term, so the man who once ran the agency that enforces securities law would oversee voluntary AI safeguards that lean on securities law. “Super intelligence is a national security issue,” he told CNBC the day after the lunch. The post does not need Senate confirmation.
Congress before 3 November
The House is in recess until November. Senate efforts have stalled, and in our report on the Senate AI bill we covered the draft that would let a federal agency stop a release. Bill Gates added his voice in late September, as we reported in our piece on AI legislation. With 30 days to the vote, voluntary AI safeguards are likely to be the only federal measure in place on polling day.
The states keep moving
California has also signed a law limiting how employers can rely on AI to dismiss staff, which we covered in our report on the No Robo Bosses Act. Its 2025 frontier law, SB 53, already requires large developers to publish safety frameworks and report critical incidents. State law, not the accord, is where binding rules for these companies are being written.
What Voluntary AI Safeguards Mean for Organisations Buying AI
For a business using AI models from these companies, the lesson of the Reuters story is practical. Every remedy Sacks lists belongs to shareholders, regulators or prosecutors. None belongs to you, and none acts before harm. Voluntary AI safeguards give customers no claim of their own.
Your contract is the before-the-harm route
The only protection a customer controls in advance is its own contract. Ask for security commitments, incident notification within a fixed number of hours, audit rights, and indemnities that are not capped at a month’s fees. Strong cybersecurity duties written into the contract will do more for your organisation than voluntary AI safeguards signed in Washington.
Questions to ask a vendor
Ask which independent auditor reviewed the model you are buying, what the scope was, and whether you can see a summary. Ask which board committee received the findings and whether any were disclosed. Ask whether the vendor will register with California’s auditor scheme when it opens. A vendor that cannot answer is telling you the voluntary AI safeguards have not reached its products yet.
For UK and EU buyers
UK and EU organisations should treat voluntary AI safeguards as background, not assurance. The EU AI Act already places binding duties on general-purpose model providers, and UK data protection law applies to how you use their outputs. Build these checks into your IT governance and AI strategy rather than relying on promises made to a US president.
Voluntary AI Safeguards FAQ
What are the voluntary AI safeguards Trump announced?
They are a one-page “White House Accord on Super Intelligence”, signed on 29 September by Nvidia, SpaceX, OpenAI, Anthropic, Meta and Google. Each company commits to internal controls, an internal safety team, an independent external auditor and a board committee that receives reports.
Are the voluntary AI safeguards legally binding?
No. Trump called them “morally binding”. Reuters reports that the document states no consequences for a company that does not comply. Existing laws still apply to anything a company says publicly about its safety work.
Can directors be sued if they ignore an auditor’s warning?
In principle yes, under Delaware’s Caremark doctrine, but only for bad-faith failures of oversight, and in practice such suits have followed serious harm, as at Blue Bell and Boeing. SpaceX’s Texas charter makes such suits very hard to bring.
What do Americans think about AI safety?
A Reuters/Ipsos poll in September found 73% worry AI companies have not done enough to prevent serious harm, and 55% favour slowing development. A UMass Amherst poll found 57% say Trump has handled AI poorly.
Is any government auditing AI companies?
California has created a registry and standards for AI auditors, with the first due to be certified by 1 January 2028. The Federal Trade Commission is investigating OpenAI and Anthropic. There is no federal AI audit requirement.
References and Further Reading
As public fears of AI grow, Trump digs in on voluntary safeguards (Reuters, via The Star)
David Sacks on the White House Accord on Super Intelligence (X)
Gavin Baker on auditors, boards and fiduciary duty (X)
UMass Amherst Poll: AI data centres and the president’s handling of AI
California adopts new AI laws requiring independent audits (CBS Sacramento)
How SpaceX’s IPO locks the courthouse door on shareholders (Deminor)
Liability for Harm Versus Regulation of Safety, Steven Shavell (NBER)
In re Caremark International Inc. Derivative Litigation (Wikipedia)
The Boeing Company shareholder derivative litigation (Lieff Cabraser)
AB 1405, Artificial intelligence: auditors: registration (California Legislative Information)
17 CFR 240.10b-5: Employment of manipulative and deceptive devices (Cornell LII)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
