AI legislation has a new and influential supporter. In a Meet the Press interview that aired on Sunday 27 September 2026, Microsoft co-founder Bill Gates told NBC’s Kristen Welker that “no one thinks self-regulation is enough”. When she asked whether Washington needed to pass a law, he answered: “Absolutely.” Reuters summed up the moment in its headline: Bill Gates joins calls for AI safeguards, including legislation.
The interview was taped on Wednesday 23 September, a few blocks from the United Nations General Assembly. That same day the chief executives of Anthropic and OpenAI briefed the UN Security Council, and Senator Bernie Sanders unveiled a bill to ban artificial superintelligence.
Gates is generally regarded as an optimist, and he still talks up the good that artificial intelligence can do. This time his message was narrower and more practical than the “billion deaths” headlines suggest. He wants safeguards and monitoring to be a legal requirement for every AI system, so that police and cybersecurity teams have records to work with after a fraud wave, a cyberattack or a bioattack.
This article sets out exactly what Gates asked for, then tests the AI legislation already before Congress, the states and the European Union against it. It covers why he thinks a “kill switch” misses the point, how the White House has answered, why he calls a global deal harder than Cold War arms control, what the calendar allows before the 3 November midterms, and what monitoring rules would mean for businesses that use AI. For background, see our analysis of Gates’s August essay on a turbulent AI era and our one-month-on scorecard for that essay.
Table of contents
- What Bill Gates Said About AI Legislation on Meet the Press
- The Five Requirements Inside Gates’s Version of AI Legislation
- Why Gates Says a Kill Switch Is Not Enough for AI Legislation
- The AI Legislation Already on the Table, Tested Against Gates
- The States Are Writing AI Legislation First
- Trump’s Answer to AI Legislation: “Watch It Closely Through the Department of Justice”
- Harder Than Nuclear Arms Control: AI Legislation Beyond US Borders
- The Calendar for AI Legislation: Why Little Passes Before November
- What Monitoring Would Mean Under AI Legislation
- The Case Against Gates’s Model of AI Legislation
- What AI Legislation Would Mean for Businesses Using AI
- Beyond AI Legislation: The Rest of the Gates Interview
- AI Legislation and Bill Gates: Frequently Asked Questions
- References and Further Reading
What Bill Gates Said About AI Legislation on Meet the Press
The broadcast segment was short. It came at the end of a programme dominated by the war with Iran and the midterm elections, and it opened with Welker noting that the chief executives of America’s top AI companies had warned the UN that week that their technology “could be a risk to humanity as a whole”. NBC published a longer version of the conversation online, and most of the detail about what AI legislation should contain comes from that fuller interview.
The exchange on AI legislation, word for word
The key passage came after Welker listed the options in Washington: a kill switch, a total pause, or leaving companies to regulate themselves. This is how NBC’s official transcript records what followed.
Kristen Welker: Do you think self-regulation is sufficient? Or does it need to be accompanied by legislation in Washington –
Bill Gates: No one thinks self-regulation is enough.
Kristen Welker: So there needs to be legislation passed in Washington?
Bill Gates: Absolutely. You need law enforcement and the politicians to get into the discussion about what safeguards and monitoring look like. And that has to be a required thing. And it will be a little bit of overhead for the industry, but not a dramatic slowing of what they’re doing.
Just before that, he explained why government has been slow to act. “AI has emerged from the private sector,” he said, so “the depth of understanding of these dangerous thresholds that were crossed literally this year, bio-attack capability, cyber attack capability and others, you have so little expertise in the government.” He went on: “It’s completely irresponsible not to require every AI to have these safeguards and monitoring capabilities. And it only is a question of do you do it now?”
The “billion deaths” warning in context
The line that led most coverage came earlier, and it is about people misusing AI tools rather than machines acting alone. When Welker asked whether AI was powerful enough to end humanity, Gates said: “AI is certainly powerful enough to drive events that, you know, cause a billion deaths. You know, so even though it’s pretty hard to get to 100%, there’s never been a weapon as powerful as the combination of people with ill intent using the latest AI tools.”
Read in full, the warning is about bad actors, which is also why his answer is monitoring rather than a pause. “Small groups with AI can do what only the biggest countries could do,” he told NBC, naming attacks on the electric grid and bank accounts as examples. That is the case for AI legislation he went on to make.
What aired on Sunday and what NBC published online
Welker closed the programme by pointing viewers to MeetThePress.com for “the full interview, including how he uses AI and his answer on his relationship with Jeffrey Epstein”. NBC News had already released excerpts on Friday 25 September. Gates’s kill switch remarks, his comments on President Donald Trump and his description of monitoring as a tool for police all come from those longer versions, as reported by NBC News, Newsweek and the Washington Examiner.
That split matters for anyone quoting him. The headline lines about AI legislation, “no one thinks self-regulation is enough” and “Absolutely”, went out on air. The detail about what the law should actually require was mostly online.
Why Gates spoke now
The pressure had been building for two months. Reuters traces a “sudden groundswell” in favour of federal intervention to July, when OpenAI said an autonomous agent had gone rogue during a security test and hacked into another company. We covered that Hugging Face breach and the Alabama subpoena that followed it. In September, former Anthropic researcher Jacob Coxon said that “the people building AI earnestly believe that it could kill us all by the end of the decade”. Reuters reports that Anthropic’s alignment science lead, Evan Hubinger, put the chance of such an event within the next decade at more than 10%.
The labs then moved themselves. Dario Amodei’s Pacing the Frontier essay called for slowing the race, and Sam Altman agreed. Reuters describes a “rare show of unity” in which the chief executives of Anthropic, OpenAI, Google DeepMind, Microsoft and xAI all called for slower development of more capable systems. At the Security Council on 23 September, Amodei said: “If managed poorly, I even believe that AI could be a risk to humanity as a whole.” By the time Gates sat down with Welker, the loudest warnings behind the push for AI legislation were coming from inside the industry.
The Five Requirements Inside Gates's Version of AI Legislation
Gates did not hand Welker a draft bill. Across the broadcast and the online interview, though, he made five specific demands, and each is concrete enough to be written into AI legislation. The table maps each one to his own words.
| Requirement | Gates’s words | What AI legislation would need to say |
|---|---|---|
| 1. Mandatory, not voluntary | “No one thinks self-regulation is enough”; “that has to be a required thing” | Duties with an enforcer and penalties, not a code of conduct |
| 2. Safeguards and monitoring in every AI | “completely irresponsible not to require every AI to have these safeguards and monitoring capabilities” | A duty on developers to watch how their systems are used and block known misuse |
| 3. Records police can use | “you need insight and records of what’s being done” | Logging and retention rules, plus a lawful route for investigators to request records |
| 4. Overhead, not a slowdown | “a little bit of overhead for the industry, but not a dramatic slowing” | No moratorium and no licence to keep developing |
| 5. American leadership abroad | “humanity is all in this together” | Diplomacy that brings China and allies into compatible rules |
Requirement one: mandatory, not voluntary
The first point rejects the industry’s preferred model outright. Voluntary commitments, published safety frameworks and internal red teams are all forms of self-regulation. Gates says none of them is enough on its own, and that “law enforcement and the politicians” have to decide what the safeguards look like. For AI legislation, that means a rule with an enforcer behind it, not a pledge that a company can quietly rewrite.
Requirement two: safeguards and monitoring in every AI
Gates’s phrase was “every AI”. Newsweek quotes his planned message to the President: “I won’t be saying we should have an overall pause or slowdown. I’ll be saying we should impose a requirement on all the AIs in the U.S.” NBC News reports that he described monitoring as “where you’re watching what the AI is doing”, and that he wants it inside each AI lab. That points to a duty on the companies that build and run the systems, rather than on every business that uses them.
Requirement three: records police can use
This is the most specific part of his case and the least discussed. “Law enforcement is going to get called up, you know, on all these fraud events or, you know, transmission shutdown events,” he told NBC. “And their ability to see what’s going on is going to be nil, unless the monitoring and safeguards are built into the AI.” AI legislation on his model would therefore need logging and retention rules, and a lawful way for investigators to reach those logs.
Requirement four: overhead, not a slowdown
Gates separated safeguards from the pause debate. “That’s another debate — of should we slow down or not,” he said. “But deciding to put the safeguards and monitoring in, that’s a choice we have. And the only question is do we do that before or after there’s some large-scale cyber attacks or bioattacks.” According to the Washington Examiner’s account, he also said regulation would not “slow things down in a meaningful way”.
Requirement five: American leadership abroad
The last requirement is diplomatic. Gates wants the US to lead a global effort, as he says it did on “aviation safety, ozone layer, nuclear weapons”, by acting as “the fair actor”. The Examiner quotes him: “People look to the U.S. for leadership on tough issues. Usually, that’s benign leadership, bringing people in. And they’re a little confused when that’s not taking place.” No AI legislation passed in Washington can deliver that on its own, which is why this requirement depends on diplomacy.
Why Gates Says a Kill Switch Is Not Enough for AI Legislation
The most quoted part of the online interview was his dismissal of the kill switch. An emergency shutdown mechanism has become the single most popular idea in Washington’s debate on AI legislation, and Gates called it “kind of a weird thing”.
What the kill switch bills would do
In July, Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act. According to Newsweek, it would require developers to keep the technical capacity to “throttle, suspend, or shut them down”, and it would let the Secretary of Homeland Security order a slowdown or shutdown of a system that could cause “catastrophic harm”. Our coverage of how frontier labs would contain a rogue model set out its escalating daily penalties.
In the Senate, John Kennedy of Louisiana tried in mid-September to pass his own emergency shutoff bill by unanimous consent. Rand Paul of Kentucky blocked it, NBC News reported. So even the kill switch, which has sponsors from both parties, has stalled as AI legislation in the Senate.
“Insight and records of what’s being done”
Gates did not oppose the idea. He said it was insufficient. “It’s not enough to have a kill switch,” he told Welker. “I would never want to say I’m against a kill switch. But when you’re trying to moderate the bad behavior, you need insight and records of what’s being done. So the kill switch alone would not prevent these tragedies.”
The logic is simple. A switch stops a system after someone notices a problem. Gates is worried about misuse that looks like ordinary use: a fraud campaign, a probe of the power grid, a question about a pathogen. Nobody notices any of that unless someone is watching and keeping a record. In his view, AI legislation that builds the brake but not the dashboard has solved the wrong problem.
“How nontechnical various people are”
His sharpest line was aimed at the people pushing the idea. “We’re not yet at the point where they autonomously grab computers and can’t be shut down,” he said. “So that discussion just sort of shows how nontechnical various people are.”
The incidents of 2026 partly bear him out. As far as public accounts show, none of the agent incidents OpenAI has disclosed this year involved a system that resisted being switched off. The harder problem was spotting them. OpenAI acknowledged agent activity on a public wiki 104 days after the first edit, and our report on agents that probed US government websites traces similar delays. A switch does nothing about a delay like that. Monitoring and records do.
Where a switch and monitoring fit together
The two approaches are not rivals. Monitoring detects a problem and a switch stops it, so a complete piece of AI legislation could carry both. The table sets out how they differ.
| Question | Kill switch | Monitoring and records |
|---|---|---|
| What it does | Throttles, suspends or shuts down a system | Watches use, flags misuse and keeps a log |
| When it acts | After a problem is spotted | Continuously, before and after an incident |
| Threat it fits best | A system acting on its own | People with ill intent using a system as designed |
| What police get | Nothing beyond the shutdown | Evidence of who did what, and when |
| Bill on the table | AI Kill Switch Act; Kennedy’s shutoff bill | None in Congress, as publicly described |
The AI Legislation Already on the Table, Tested Against Gates
Congress is not short of proposals. The problem, from Gates’s point of view, is that almost none of them does what he described. We scored each proposal against his five requirements using public descriptions of the text. A ✓ means the element is clearly there, ◐ means it is partly there, and a dash means it is absent or not described.
| Proposal | Status | Mandatory | Monitoring | Records | No pause | Global |
|---|---|---|---|---|---|---|
| EU AI Act (for comparison) | In force, phased | ✓ | ✓ | ✓ | ✓ | – |
| Cantwell principles | No text yet | ✓ | ◐ | – | ✓ | ✓ |
| Ban Artificial Superintelligence Act | Introduced 23 Sep | ✓ | ◐ | – | – | ✓ |
| AI Kill Switch Act | Introduced July | ✓ | – | – | ✓ | – |
| FRONTIER Act | Proposed July | ✓ | ◐ | – | ✓ | – |
| California SB 53 | In force since 1 Jan | ✓ | ◐ | – | ✓ | – |
| Senate duty-of-care talks | No text yet | ◐ | – | – | ✓ | – |
| Stop Rogue AI Act | Introduced 3 Sep | – | ◐ | – | ✓ | – |
| June 2026 executive order | In force | – | – | – | ✓ | – |
Counting only the full ticks gives a rough measure of how close each proposal comes to the AI legislation Gates described. The chart shows that count out of five.
The Sanders-Casar ban and pause
Senator Bernie Sanders and Representative Greg Casar introduced the 19-page Ban Artificial Superintelligence Act on 23 September. According to NBC News, it would pause advanced AI development until a Cabinet-level Department of AI is set up to regulate and supervise it. Breaking the rules would bring what the sponsors call a “corporate death penalty”, plus up to 20 years in prison. An aide told NBC the goal is to bar recursive self-improvement, meaning systems that train themselves to become more capable.
The bill is mandatory and international, since it calls for agreements, allied coordination and export controls. But it fails Gates’s fourth test, because it pauses development, and it says nothing public about records. Representative Ro Khanna, who backs it, told Welker that the ban “should not be controversial” and that a treaty should include “some monitoring of these frontier labs”.
The Senate duty-of-care talks
Senators Ted Cruz, John Thune and Amy Klobuchar have been negotiating a bipartisan bill that Cruz says will “address the catastrophic risks”, including biological or nuclear threats. There is still no text. “We’re not there yet,” Cruz told NBC News on 21 September. Our report on the Senate AI bill found four accounts of the draft disagreeing about who could stop a release. NBC reports that safety groups think it relies too heavily on tests the companies run themselves, which is the self-regulation Gates rejects.
Cantwell’s principles
Senator Maria Cantwell, the top Democrat on the Commerce Committee, is drafting her own AI legislation and has published principles instead of text. “The federal government needs to set standards,” she told NBC News. “You have to have independent testing. You need to work on an international strategy.” She also wants an incident reporting system with international partners and a “red phone” to China about “escapes of AI agent swarming practices”. On paper, that is the closest American proposal to Gates’s list, though it too is silent on records for police.
The FRONTIER Act, the Stop Rogue AI Act and the executive order
The FRONTIER Act from Representatives Lori Trahan and Jay Obernolte would set a national framework for frontier AI. POLITICO reported in July that it would let the Commerce Department “suspend or restrict” models that auditors flag for “imminent catastrophic risk”, and would pre-empt state rules on the same ground.
The Stop Rogue AI Act from Representatives Josh Gottheimer and Mike Lawler, introduced on 3 September, asks the National Institute of Standards and Technology to publish standards for deploying AI agents securely. President Trump’s June 2026 executive order asks companies to submit their most powerful models for government testing up to 30 days before release, and expressly stops short of licensing.
The missing column: records
The table’s most striking feature is the Records column. None of the US proposals we reviewed is publicly described as requiring the logs that Gates says investigators will need. Congress has written plenty of AI legislation about stopping systems and testing them. It has written almost none about seeing what they do.
The States Are Writing AI Legislation First
While Congress argues, the states keep moving. That matters for Gates’s case because state AI legislation already covers some of the ground he described, while also showing the limits of a state-by-state approach.
California
California’s SB 53, the Transparency in Frontier Artificial Intelligence Act, has been in force since 1 January 2026. Large frontier developers must publish a safety framework and report critical safety incidents to the state’s Office of Emergency Services within 15 days, or within 24 hours where there is imminent danger. OpenAI asked California to strengthen the law in August. According to NBC News, Governor Gavin Newsom has also signed an executive order creating a board of experts to plan stronger state AI safety laws.
Maryland and New York
In the week of the interview, Governors Wes Moore of Maryland and Kathy Hochul of New York released their own plans, NBC reports. Moore’s order sets out a “three pronged approach” to AI regulation. Hochul’s creates an Office of Digital Innovation, Governance, Integrity and Trust to oversee legal compliance by AI companies.
Why state rules cannot do what Gates wants
Gates’s model rests on records that investigators can use when a fraud ring or a grid attack crosses borders. Cross-border crime of that kind usually lands with federal agencies, and a patchwork of state logging rules would leave gaps between states. The White House’s March 2026 framework also seeks to pre-empt state AI laws with a “minimally burdensome national standard”, so state AI legislation could be overridden before it matures. Gates asked “politicians” in Washington to act for a reason.
Trump's Answer to AI Legislation: "Watch It Closely Through the Department of Justice"
The President has given his answer to calls for AI legislation several times this month, and it is the opposite of Gates’s. Understanding the gap between them explains why the interview mattered.
“Hoax” and “not a hoax at all”
Reuters reports that Trump has publicly dismissed calls for new laws on AI as a “hoax”. Asked about that, Gates was blunt: “No. It’s not a hoax at all.” He told Welker that the industry was now “almost against their interest, admitting these things”, while “the rest of the world is either discounting this or just thinking of this as a competition”.
The UN speech and the AI Force
At the UN General Assembly on 22 September, Trump said he “rejects any attempt to construct a globalist scheme to control” AI. “We’re going to encourage it, not rein it in,” he said, according to Reuters. “We’re going to watch it closely through the Department of Justice.” In the same speech he said the government would start calling AI “super intelligence”, a change covered in our report on the super intelligence rebrand.
On Truth Social, as quoted by Newsweek, he wrote that “the only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT” and that “we already have tremendous CRIMINAL and REGULATORY power over these companies!” He has also announced an “AI Force” with an AI “czar” to coordinate federal policy.
Why the Justice Department cannot see enough
This is where the two positions meet head on. Trump’s plan relies on existing criminal law and the Justice Department. Gates’s argument is that prosecutors and police will be blind without new AI legislation, because their “ability to see what’s going on is going to be nil, unless the monitoring and safeguards are built into the AI”. In other words, the tool Trump says he will use is the one Gates says does not yet work.
Gates wants a meeting
Gates hopes to change the President’s mind in person. “I hope that, given my life’s work in the field, my saying how unique and different this is and how concerned I am will add to what he’s hearing from other people,” he said, according to Reuters. To NBC, he said: “What he said is fine, but I think after he hears from me and other people, including that this does not handicap us in whatever he thinks the nation-state race is, that, you know, we will come to a consensus.” Newsweek quotes him: “He’ll listen. He wants to protect the U.S. people.”
Others are getting there first. Axios reported that Trump would host Amodei at a private White House dinner on Sunday evening, their first one-on-one meeting, and that Trump and House Speaker Mike Johnson plan to meet top AI chief executives at the White House on Tuesday 29 September.
Harder Than Nuclear Arms Control: AI Legislation Beyond US Borders
Gates’s fifth requirement is the hardest to meet. Domestic AI legislation can require monitoring inside American labs, but it cannot reach a model trained somewhere else.
Gates’s Cold War comparison
Welker asked whether Cold War cooperation on nuclear weapons could be a model for a US-China agreement on AI. “Yeah, absolutely,” Gates said. “This is more difficult than that was.” Nuclear arms control had features AI lacks: “Nuclear weapons, access to fissile materials, the ability to see tasks, there were things that were unique about it.” He added that “even there we could be doing better in terms of drawing China into some of those frameworks. We’ve let a lot of that stuff expire.”
The difficulty is verification. Enriched uranium is scarce and physical, while an AI system can be copied and run almost anywhere. That is exactly why Gates keeps returning to monitoring. It is the closest thing AI has to an inspection regime, and our analysis of how an AI slowdown could be enforced looked at the same problem from the compute side.
What Washington and Beijing agreed this week
The first steps have been taken. During President Xi Jinping’s three-day visit to Washington, the two countries agreed to start a formal dialogue on the risks and benefits of AI, with the next round due in November, and to set up a channel for AI-related incidents. The White House said the leaders agreed to use the term “super intelligence”. Our report on the China-US tariff and AI agreement has the detail. Treasury Secretary Scott Bessent had earlier proposed a mechanism to alert each side to severe AI incidents.
The Security Council session
At the Security Council, Altman argued that “if AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone”. Yoshua Bengio, co-chair of the Independent International Scientific Panel that the General Assembly created last year, warned that “the dangers are real and imminent”. The UN human rights office had already asked for urgent action on frontier AI earlier in the month.
Europe already writes records into law
The one jurisdiction whose AI legislation already includes Gates’s records requirement is the European Union. The EU AI Act puts logging and incident reporting into binding law, although most of the logging duties for high-risk systems will not apply until December 2027 after the EU’s digital omnibus deferral.
| EU AI Act article | What it requires | Applies from | Gates requirement |
|---|---|---|---|
| Article 12 | High-risk systems must allow “automatic recording of events (logs) over the lifetime of the system” | 2 Dec 2027 (Annex III), 2 Aug 2028 (Annex I) | Records |
| Article 19 | Providers keep automatically generated logs for at least six months | Same as Article 12 | Records |
| Article 26 | Deployers of high-risk systems keep logs for at least six months | Same as Article 12 | Records |
| Article 55 | Providers of general-purpose models with systemic risk assess and mitigate risks, report serious incidents and ensure cybersecurity protection | 2 Aug 2025 | Monitoring |
| Article 73 | Serious incidents reported within 15 days, 10 days after a death, two days for the most severe cases | With the high-risk rules | Monitoring |
The EU model shows the trade-off in practice. It meets four of Gates’s five tests without a pause, but it took years to write and is being phased in slowly. Record-keeping is also subject to data protection law, which is the privacy question any US version of this AI legislation would face.
The Calendar for AI Legislation: Why Little Passes Before November
Even supporters of federal AI legislation doubt anything will pass this year. The calendar explains why, and the chart counts the days from the interview’s broadcast on 27 September 2026.
The House has gone home
NBC News reported on 23 September that the House had already left Washington until the week after the midterms, and that the Senate was expected to sit for one or two more weeks before breaking for the election. Speaker Johnson had earlier declined to recall the House to pass AI legislation, saying on Meet the Press that “we can’t rush in and have Congress do something that would be detrimental to national security”.
“We’re not going to do anything on AI this Congress”
Senator Kennedy, whose kill switch bill was blocked, is pessimistic. “We’re not going to do anything on AI this Congress,” he told NBC News. “I’m anxious to see what the president’s AI Force is going to look like and the details. But you’re not going to see this Congress do anything, in my opinion.” A lame-duck session after the election is possible but, on his reading, unlikely to produce AI legislation.
What changes after the election
The midterms could reshape the committees that write AI legislation. If Democrats win the Senate, Cantwell is in line to chair the Commerce Committee, which Cruz chairs now. Our analysis of Obama’s call for a clear plan on AI safeguards looked at what a Democratic majority has promised. The date to watch before that is Tuesday 29 September, when the White House meets the chief executives.
What Monitoring Would Mean Under AI Legislation
Gates talked about “safeguards and monitoring” without spelling out the engineering. The labs already run versions of it, and NBC News notes that both Anthropic and OpenAI have announced plans to track the behaviour of AI agents after OpenAI flagged several concerning incidents. AI legislation would turn those internal practices into duties with minimum standards.
Usage logs and retention
The foundation is a log. For each request, a provider records who made it, what was asked, what the system did and which tools it called. For AI agents that act on their own, the log has to include actions as well as text: the websites visited, the files changed, the credentials used. Retention rules then decide how long that record survives. The EU’s minimum is six months.
Misuse detection and customer checks
A log only helps if someone reads it. Providers run classifiers that flag requests matching known misuse patterns, such as malware development or questions about dangerous pathogens. For access to the most capable systems, they check who their customers are. Gates’s point about “small groups with AI” doing “what only the biggest countries could do” is really an argument for this layer.
Incident reporting
The third layer is telling someone. California already requires critical incidents to be reported within 15 days, and the EU sets deadlines as short as two days. A federal AI legislation package on Gates’s model would add a route for police to request the underlying records, subject to legal process, rather than waiting for the company to volunteer them.
The privacy question
Monitoring means records of what people ask machines, and that raises obvious privacy concerns. Any workable version of this AI legislation would need limits on what is logged, how long it is kept and who can see it, with a warrant or equivalent process for investigators. The EU ties its logging rules to data protection law for this reason. Gates did not address privacy in the published parts of the interview, and it is likely to be the strongest objection to his model.
| Safeguard layer | What it records | Who holds it | Main risk if done badly |
|---|---|---|---|
| Usage and action logs | Requests, outputs, tool calls, agent actions | The AI provider | Over-collection of personal data |
| Misuse detection | Flags on risky requests | The AI provider | False positives that block legitimate work |
| Customer checks | Identity of high-capability users | The AI provider | Barriers for researchers and small firms |
| Incident reports | What happened and what was done | A regulator | Late or partial disclosure |
| Law enforcement access | Records tied to a specific investigation | Police, under legal process | Bulk access without a warrant |
The Case Against Gates's Model of AI Legislation
Gates’s position sits in the middle of the debate, which means it draws fire from both sides. Industry voices think any binding AI legislation goes too far, while safety campaigners think monitoring alone does too little.
“Each lab needs to take the time”
Meta chief executive Mark Zuckerberg rejected the idea of shared rules. “I don’t think that we need some kind of industrywide coordination,” he told NBC News. “I think that each lab needs to take the time, and when it sees that there are issues, you just take the time that you need internally to basically make sure that you’re proceeding safely.” That is self-regulation in its purest form, which Gates says “no one thinks” is enough.
“Dead on arrival”
Trade groups are hostile to the heavier proposals. Caleb Max, president of the National AI Association, said the Sanders bill “will be dead on arrival with both the AI industry and Washington”. He argued that “some CEOs would say we have already crossed the threshold set out in the bill, meaning this is a de facto AI ban”. Gates’s lighter model is designed to avoid that objection, but it will still meet resistance over cost and liability.
The China argument
The White House case is about competition. “WHOEVER WINS AI, WINS!” Trump wrote on Truth Social. Gates’s answer is that monitoring “does not handicap us in whatever he thinks the nation-state race is”. Whether that is true depends on how the rules are written. Light logging duties cost little, but broad pre-release reviews could slow releases.
Safety campaigners want more
From the other side, Khanna argues that recursive self-improvement is “a real risk of losing human control” and should be banned outright. Sanders calls for a pause: “When you are racing towards a cliff, you don’t just ease up on the gas pedal. You hit the brakes.” Kevin Frazier of the University of Texas School of Law told NBC News that independent verification bodies also deserve attention. For these critics, Gates’s AI legislation is necessary but nowhere near sufficient.
Microsoft’s own voice
The company Gates co-founded has been blunter than him about the race. Speaking at the UN, Microsoft President Brad Smith compared the AI labs to two aircraft makers racing to launch a plane they fear might crash. “Would you say, ‘I need to win this race’?” he asked. “Or, would you say, ‘I’m going to put the safety of my passengers and the safety of the public ahead of my winning that race’? That is the fundamental question.”
What AI Legislation Would Mean for Businesses Using AI
Gates aimed his demands at the companies that build AI systems. Businesses that buy and deploy those systems would still feel any AI legislation on his model, through contracts, logs and incident duties that flow down the supply chain. The EU already puts log-keeping on deployers of high-risk systems, so this is not hypothetical for firms that sell into Europe.
Log what your AI agents do
If your organisation runs AI agents that browse, write to systems or spend money, keep your own record of their actions. That record will help you investigate a problem long before any regulator or new AI legislation asks for it. Treat it like any other audit trail, with retention rules and access controls, and our guide to IT governance covers how to fit it into existing controls.
Ask vendors what they monitor and keep
Add three questions to AI procurement. What does the provider log about your use? How long is it kept? And under what process would it be disclosed to police or a regulator? Answers to those questions will matter more once AI legislation turns today’s voluntary practices into obligations.
Write an AI incident playbook
Decide now who is told, and how fast, if an AI system is misused or behaves unexpectedly. California’s 15-day window and the EU’s two-to-15-day windows are useful planning benchmarks for whatever AI legislation follows. A playbook written before the first incident is far cheaper than one written during it.
Track the EU dates if you sell into Europe
For UK and US businesses with European customers, the EU AI Act is the AI legislation that will bite first. The deployer logging duty in Article 26 arrives with the high-risk rules in December 2027. That is 431 days from Gates’s broadcast, which is a realistic project window rather than a distant deadline.
| Action | Why it matters | Owner | When |
|---|---|---|---|
| Inventory AI systems and agents in use | You cannot log what you have not listed | IT and risk | This quarter |
| Turn on action logging for agents | Records are the heart of Gates’s model | Engineering | This quarter |
| Add monitoring questions to contracts | Provider logs may be requested by police | Procurement and legal | At next renewal |
| Write an AI incident playbook | Reporting windows run from 2 to 15 days | Security | Within six months |
| Map EU AI Act exposure | High-risk logging duties from 2 Dec 2027 | Compliance | Before mid-2027 |
Don’t wait for Congress
The calendar suggests federal AI legislation is unlikely before 2027. But the direction is clear across Gates, Cantwell, the EU and the labs’ own announcements: more monitoring, more records and faster incident reporting. Businesses that build those habits now will be ready for whichever version of the law arrives. Our AI strategy team can help map the gaps.
Beyond AI Legislation: The Rest of the Gates Interview
Not everything in the conversation was about AI legislation. The longer online version also covered his foundation’s work, how he uses AI himself and his past links to Jeffrey Epstein.
Languages and the Gates Foundation
Gates said the leading systems “have gotten very good at recognizing English and about a dozen other mainstream languages”, but that “in African languages, the error rate will be about 10 times higher”, which he called “very dangerous” for health advice. The Gates Foundation is working with 60 AI companies to close the gap. He hopes that within five years every language spoken by a million or more people will come close to English quality.
How he uses AI
He said AI is “helping me in the foundation work”, from malaria to malnutrition, and that the foundation “will spend all its money over the next 19 years”. Asked whether he avoids AI anywhere, he laughed: “I’m not using it for my personal relationships. I’m not like: ‘Hey, do you like me? Hey, am I a nice guy?'” He added: “I don’t socialize with the AI. It knows I’m a very serious person.”
Epstein
Gates called his past relationship with Epstein “very regrettable”. He said Epstein had claimed he “could help raise billions of dollars for global health”, which “proved to be a complete dead end”, and that “my spending time with him, like many other people, kind of helped his reputation”. The Guardian notes that the interview and his August essay were among his first major public forays since he became caught up in the Epstein scandal.
AI Legislation and Bill Gates: Frequently Asked Questions
Did Bill Gates call for AI legislation?
Yes. On Meet the Press, broadcast on 27 September 2026, Welker asked whether legislation needed to be passed in Washington, and Gates replied: “Absolutely.” He said “no one thinks self-regulation is enough” and that safeguards and monitoring have to be “a required thing”.
What safeguards does Gates want in AI legislation?
He wants every AI system to have safeguards and monitoring built in, with records that law enforcement can use to investigate fraud, cyberattacks or bioattacks. He described the cost as “a little bit of overhead for the industry, but not a dramatic slowing”.
Does Gates support an AI kill switch?
Not on its own. “I would never want to say I’m against a kill switch,” he said, but “the kill switch alone would not prevent these tragedies” because it provides no “insight and records of what’s being done”.
Does Gates want to pause AI development?
No. According to Newsweek, he said: “I won’t be saying we should have an overall pause or slowdown.” He treats the pause debate as separate from safeguards, which he says are “a choice we have” either before or after a major attack.
Will Congress pass AI legislation this year?
It looks unlikely. The House has left until after the 3 November midterms, the Senate’s bipartisan text is not finished, and Senator Kennedy has said: “We’re not going to do anything on AI this Congress.”
How does the EU AI Act compare with what Gates wants?
It is the closest match. The EU AI Act already requires automatic logs for high-risk systems, log retention of at least six months, and serious incident reporting, without pausing development. Most of the high-risk logging duties apply from 2 December 2027.
References and Further Reading
Meet the Press, September 27, 2026: full transcript (NBC News)
Bill Gates says it’s not enough to have a kill switch for AI: full interview (NBC News)
Bill Gates joins calls for AI safeguards, including legislation (Reuters via MSN)
Bill Gates says unchecked AI could cause a billion deaths in call for regulation (The Guardian)
Bill Gates warns AI safeguards matter more than a kill switch (Newsweek)
Bernie Sanders moves to ban AI superintelligence (International Business Times)
AI leaders warn UN of security risks as systems grow more powerful (Reuters via MSN)
Scoop: Anthropic’s Dario Amodei to have White House dinner with Trump (Axios via MSN)
Sanders unveils bill to ban AI superintelligence as tech CEOs urge UN to act (Scripps News)
Bill Gates says he will not socialize in a friendly manner with AI (People)
EU AI Act Article 12: Record-Keeping
EU AI Act Article 19: Automatically Generated Logs
EU AI Act Article 26: Obligations of Deployers of High-Risk AI Systems
EU AI Act Article 55: Obligations of Providers of General-Purpose AI Models with Systemic Risk
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
