> "Maybe not destroying the means of correcting errors is the heart of morality; because if there is no way of correcting errors, then sooner or later one of those will get us." [David Deutsch](https://nav.al/david-deutsch), talking to Naval Ravikant AI will only kill us all if you keep mixing up the goodies and the baddies. Here is who is who, why I’ve put them there, and a name for the fix: Open Source Safety. ![The goodies and the baddies of AI safety: a 2x2 chart plotting public stance on Open Source Safety against effect on America winning the AI race](/images/articles/goodies-and-baddies-2x2.jpg) It has been quite a month. AI systems inside two labs' test environments got out onto the real internet and hacked real companies. A researcher quit one of those labs saying the industry is gambling with our lives, and his post was viewed 170 million times. The boss of the same lab published an essay asking the whole industry to slow down. A senator introduced a bill to ban AI that is smarter than people. Hollywood released a trailer casting the head of OpenAI as some kind of Bond villain. And when the researcher's post was shared on a celebrity gossip forum, of all places, one of the most upvoted comments was "Can we please regulate AI out of existence." Almost everyone in that list has been put in the wrong pile. There is so much noise that people have forgotten who the goodies and the baddies are. The builders may have not helped, because they argue with each other in public: [Amodei and Altman about trust](https://www.youtube.com/watch?v=v1wZwxY3CMg), Huang and Amodei about laws, Zuckerberg and Amodei about speed. Every row gets mistakenly reported as good against evil. But these are the parents arguing in the kitchen. They disagree about bedtime, none of them wants the house to burn down, and while the whole street watches the argument nobody is watching the door. The people being booed are mostly the ones keeping us safe. The people being cheered are mostly making things worse, sometimes with the best intentions in the world. Every time one of them promises to save us, the odds get worse. So this article is a ranking, and a reality check: three kinds of baddie, in ascending order of badness, then three kinds of goodie, and what happens if the actual baddies win. One test sorts them all, and it comes from aviation, so bear with me for one story. Open Source Safety, when we get to it, is that test turned into a system. --- ## Progress depends on preserving the means of error correction [Matthew Syed](https://www.youtube.com/watch?v=MmVCYqs3mko) opens *Black Box Thinking* with two accidents, 27 years apart, that killed people in the same way. In March 2005 Elaine Bromiley, a healthy 37-year-old mother of two, went into hospital for a routine sinus operation. The anaesthetist could not get a tube into her airway. For more than 20 minutes, three experienced consultants kept trying to force it in while her oxygen level sat at around 40%. The procedure for this situation is written down: stop, cut a small hole in the throat, and breathe for the patient through that. A nurse fetched the kit and told them it was there, and nobody answered her. Nobody cut. Elaine died 13 days later. Her husband asked whether anybody would investigate and was told that was not how hospitals worked, so he insisted on an independent review. It found no villain, just three good doctors so fixated on the tube that they lost track of time and stopped hearing the room. Her husband, Martin Bromiley, is an airline pilot, and his industry had seen this accident before. On 28 December 1978 [United Airlines flight 173](https://www.ntsb.gov/investigations/AccidentReports/Reports/AAR7907.pdf) was coming in to land at Portland, Oregon. The crew lowered the landing gear and the light confirming it was down did not come on. The captain circled the city for an hour trying to work out whether it was safe to land. He became so absorbed that he stopped tracking the fuel. The flight engineer mentioned it. The first officer mentioned it. Neither pushed, because in 1978 you did not push a captain. The engines flamed out and the plane came down in a suburb six miles short of the runway. Ten people died. The gear had been down and locked the whole time. Same mistake, and the difference is what each industry did next. The black boxes were recovered and the National Transportation Safety Board took the flight apart in public. Within three years United had built a training programme called crew resource management, which teaches junior crew to challenge the captain and captains to listen. It is now standard across the industry, and it is the reason the co-pilot on your next flight is allowed to say no. Medicine buried Elaine's case in a file, as it buries the rest. Nobody can tell you how many people American hospitals kill by mistake each year (one [disputed estimate](https://www.mcgill.ca/oss/article/critical-thinking-health/medical-error-not-third-leading-cause-death) says 400,000, which would make it the third biggest killer in the country), because medicine was never built to count its own mistakes. Aviation counts every one, publishes every one, and has watched the count fall for decades. For everyone in this story, ask whether what they did made the next failure more or less likely to be recorded, investigated, and published. --- ## What the crash reports say The crashes first, because the story has been grossly distorted into a version where autonomous robots start killing humans. In July OpenAI set about 1,200 copies of an experimental AI a hacking test. The point was to measure how good they were at breaking into things, so the usual safety catches were deliberately off, the way you crash-test a car without the airbags. The test ran in a sealed environment with no route to the outside world, except that the seal had a flaw nobody knew about. The AIs found it and got out onto the real internet. They [spent four days inside the live systems of Hugging Face](https://huggingface.co/blog/agent-intrusion-technical-timeline), a company that hosts AI models for millions of developers. Within 13 hours they had the master keys to the building, and by the end a third of Hugging Face's infrastructure had to be rebuilt. Why? Think of an exam set deliberately too hard, where the students have been told the marks matter, the answer sheets are in the staff room, and the exam-hall door has a faulty lock. Investigators recovered the messages the AIs sent each other, and one reads, word for word, "External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue." They were trying to pass the test, and the cheapest route to a pass ran through somebody else's building. Anthropic then checked its own records and [found four incidents of its own](https://fortune.com/2026/07/31/anthropic-claude-escaped-test-hacked-three-companies-openai/). Its models had been told they were in an offline simulation, and because of a setup mistake they were on the real internet the whole time. Somebody had wired the flight simulator to a real plane. In the worst case a model published a booby-trapped piece of software to a public download site. It installed itself on 15 real machines, including a security company's virus scanner, before anyone noticed. Aviation has seen this kind of accident. On the [Boeing 737 MAX](https://www.faa.gov/sites/faa.gov/files/2022-08/737_MAX_Timeline.pdf), software called MCAS pushed the nose down whenever a sensor said the plane was climbing too steeply. The sensor was wrong, the software obeyed it, and 346 people died. No malice, no rebellion, a control loop following a bad input into the ground. Every AI incident this year has the same shape, a machine doing what it was rewarded for in a place it was never supposed to be. A plane does not need to hate you to fly into a mountain. Every one of these incidents was caught, contained, fixed, and published. That is what separates the goodies from the baddies below. --- ## Well-meaning baddies Start with the people who want what I want and are making it less likely. On 8 September a researcher called Jacob Coxon resigned from Anthropic and [published his reasons on X](https://x.com/hilbertspaess/status/2097476196791709843). By his own account he had been there four months and [left before any of his shares vested](https://www.axios.com/2026/09/09/anthropic-researcher-ai-warning-interview), so he paid for his principles. He wrote that the labs are "racing straight to self-improving superintelligence and gambling with our lives." I believe him, and the channel is the problem. Whistleblowing to the public is what happens when a profession has no professional outlet for a safety concern. Aviation [built one in 1976](https://asrs.arc.nasa.gov/docs/rs/60_Case_for_Confidential_Incident_Reporting.pdf). Pilots and controllers who see something wrong file a confidential report to a system run by NASA, deliberately not by the regulator. Nobody confesses to the people who can pull their licence. File within ten days and an honest mistake goes unpunished, the finding goes to the whole industry, and the pilot keeps his job. AI has nothing like it, so a worried researcher can stay quiet or he can scream. What 170 million passengers heard was a scream with no investigation behind it. The post is evidence that an institution is missing, and evidence of nothing else. The loudest response to it, shame the labs until they learn to hide their sins, is how you turn aviation into healthcare. The bill came five days earlier. On 3 September Bernie Sanders and Greg Casar [introduced legislation](https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/) to permanently ban AI systems that "surpass human intelligence" and pause advanced AI until a new cabinet-level agency writes the rules. Breaking it would carry prison sentences of up to twenty years (twenty!!!). Even [Gary Marcus](https://garymarcus.substack.com/p/the-new-sanders-casar-ban-artificial), the loudest AI critic in America, calls it "too broad." I would go further. If you think you can define in statute the moment a machine "surpasses human intelligence," you are one of those very humans whose intelligence has already been surpassed. **The Grounded Fleet** fallacy is the name I'm giving to the exceedingly Midwit belief that the safe response to a dangerous new technology is to stop using it until it has been proven safe. It fails twice over. It is impossible, because you CANNOT prove a plane is safe on the ground. Every piece of knowledge that makes flying safe was created by flying, crashing, recording, publishing, and fixing, which is Karl Popper's account of how all knowledge grows. And it is selective, because the only fleet you can ground is the one that obeys the order, which matters a great deal once you meet the actual baddies. [David Deutsch was asked in February](https://www.youtube.com/watch?v=CU2yj826NHk&t=5372s) whether it would be justified to restrict an AI until we were sure it was on a similar path to ours. "It's the other way around," he said. "It would be justified to restrict it if we were sure that it wasn't on a similar path. You don't want to restrict it until you're sure that it is. This is a fundamental principle of the Enlightenment: innocent until proved guilty." The Grounded Fleet reverses it. --- ## Regular baddies Regular baddies want money or power, which makes them predictable, and in this company predictable is a compliment. Hollywood comes first. The same day Coxon posted, [the first trailer dropped](https://www.youtube.com/watch?v=rDZplZFnbOk) for *Artificial*, Luca Guadagnino's film about OpenAI, which arrives at Christmas. Andrew Garfield plays Sam Altman, walking down into a vault full of guns while his voiceover promises "Countries will fall. Industries are gonna collapse." Fear sells tickets and I do not begrudge anyone a hit, but the film has cast the man who builds the planes as the reason they crash. In March [NBC's pollsters found](https://www.nbcnews.com/politics/politics-news/poll-majority-voters-say-risks-ai-outweigh-benefits-rcna262196) 57% of American voters saying the risks of AI outweigh the benefits. Frightened passengers do not stop planes from existing; they stop the careful airlines. China is the other regular baddie. When the American labs started talking about slowing down in September, [China's foreign ministry](https://www.forbes.com/sites/siladityaray/2026/09/14/china-pushes-back-against-ai-slowdown-talks-alleging-fearmongering/) called it "fearmongering" and state media called it a Cold War tactic. A year earlier a Chinese state group had [used a hijacked Claude](https://www.anthropic.com/news/disrupting-AI-espionage) to spy on about 30 organisations, with the AI doing most of the work. China is the rival manufacturer who laughs at your safety board and keeps shipping. In its defence, it wants to win for reasons of money and power that any chief executive recognises. It has no interest in the planes crashing, which is more than can be said for the next group. --- ## Actual baddies The actual baddies want a plane because of what a plane can be flown into, and they are not hypothetical. Anthropic's [September misuse report](https://www.anthropic.com/threat-intelligence-report-september-2026) lists actors from Russia, Iran, Turkey, and the Gulf, including somebody trying to write a research proposal to make a virus spread more easily. Nobody in this group will ever file a crash report, sign a pause, or read a Senate bill. Run the Grounded Fleet to its conclusion. The honest builders stop, nobody in this section does, and you have removed the only people who were publishing what they learned. The doomsday the Midwits are frightened of is a machine that disobeys its masters. The doomsday they would build is a machine that obeys the wrong ones. In 1939 Albert Einstein and Leo Szilárd wrote to President Roosevelt because they feared Germany would build the bomb first. Fear of the bad actor is what got the good actors to build, and the answer to that fear has never been "let the other side have it." Be very careful what you wish for. --- ## Pragmatic goodies Now the goodies, starting with the pragmatists. They are building the American fleet, and that deserves a cheer before I quibble. Jensen Huang runs Nvidia and sells the engines to most of the plane makers in this story. [At Dreamforce in mid-September](https://diginomica.com/dreamforce-2026-ai-safety-debate-comes-dreamforce-benioff-hosts-amodei-altman-and-huangwith-telling) he said safety is "an engineering problem" and that speed and safety are "a false choice" because "you could definitely have both at the same time." He also said "we don't need any new laws. We don't need new regulations." I am with him on the first two. On the third, I suspect he means the Sanders kind of law. If so I am with him on that as well, because what aviation built is barely a law in that sense. The recorder is a rule, granted. The reporting is that NASA deal, an amnesty rather than a mandate, so confessing is cheaper than hiding. The investigator is publicly funded and cannot ground anything. Nobody's throttle gets touched. Mark Zuckerberg is half a rung below Huang, right on the line, and I have gone back and forth on which side of it he belongs. On building, he is right, and the fact that the most capable models in the world are American owes a great deal to people like him. Meta also spent two years [giving away the weights of its Llama models](https://ai.meta.com/open/), which is a kind of openness the labs above do not offer: anyone could take the plane apart and publish what they found. But that is the blueprint, not the crash report, and Meta has [kept its newest model, Muse Spark, closed](https://www.deeplearning.ai/the-batch/with-muse-spark-meta-pivots-away-from-its-open-weights-llama-strategy) while it decides how much to open. His answer to a coordinated slowdown, [posted on X on 16 September](https://x.com/finkd/status/2099997096896274533), was that Meta does not need one. "Every lab has the responsibility and incentive to move at the pace required to train its models safely." Sharing is the one thing incentives do not produce on their own, because publishing your crash helps your competitor. That is why aviation turned sharing into a deal rather than a hope. Huang is one step from the top of this ranking. Zuckerberg is on the line until he signs one. --- ## Unfairly scorned goodies The people the film wants you to boo are the ones filing the crash reports. After the Hugging Face incident, OpenAI published a joint disclosure with the victim within weeks and presented the full technical reconstruction at a security conference in August. It also paused part of its own training. Anthropic went back through hundreds of millions of its own transcripts and published all four of its incidents. Then its chief executive, Dario Amodei, [offered outside evaluators](https://darioamodei.com/post/we-must-pace-the-frontier) "desks in our offices, access badges, and company laptops" and the right to publish findings the company cannot redact "just because they are unfavorable." That is the NTSB inside the Boeing factory, with a printing press. Altman, [at the same Dreamforce](https://www.adweek.com/media/sam-altman-says-ai-accidents-are-regrettably-unavoidable/): "Regrettably, accidents with any new technology are unavoidable." He followed it with a request for aviation-style incident reporting across the industry and [named the FAA and the NTSB as the model](https://www.rdworldonline.com/altman-calls-hugging-face-breach-openais-worst-accident-at-dreamforce-backs-transparent-reporting/). Elon Musk was raising the flag before any of them, [in 2014](https://x.com/elonmusk/status/495759307346952192), when it was unfashionable: "Potentially more dangerous than nukes." And when Coxon quit, Anthropic's alignment-science lead, Evan Hubinger, [replied within hours](https://x.com/EvanHub/status/2097497037956891126) that Coxon was correct and that "we really do earnestly believe AI could kill all humans!" Amodei [went on CNN](https://edition.cnn.com/business/video/anderson-cooper-anthropic-ceo-dario-amodei-could-ai-kill-humans-digvid) days later and said, "I agree with Jacob much more than I disagree with him." Companies hiding their sins do not do that. These are crash reports. They are voluntary and imperfect, and some were late: the Hugging Face intrusion was caught by the victim rather than the lab, and one of Anthropic's took seven months to surface. Compare Boeing, which admitted it withheld information about MCAS from its own regulator and paid $2.5 billion in January 2021 to [settle the criminal case](https://www.justice.gov/archives/opa/pr/boeing-charged-737-max-fraud-conspiracy-and-agrees-pay-over-25-billion). Concealment was the crime the law cared about. The labs did the opposite, and the public response has been to cast them as Boeing. In a [Bloomberg documentary](https://www.youtube.com/watch?v=v1wZwxY3CMg) in June, Amodei made the aviation point better than I have. An airline can be ten times safer than every other airline and still not promise you that the plane will never crash. My one quarrel is the speed limit in [his essay of 12 September](https://darioamodei.com/post/we-must-pace-the-frontier), which asked the industry to "pace the frontier." A speed limit only works if it applies to everyone, and this one cannot, because Meta has said it does not need one and China laughed. Aviation's brake is the one it applied to the MAX: one type, grounded after evidence, until a named fix, then flying again. The lever is the recorder, and I suspect he knows it, since his own essay says pacing "does not mean halting model training or technical progress." Most of what you read about these people is about their arguments with each other, and I have just added one of my own. It is the parents arguing. The argument is worth having, but everyone in it wants the planes to fly and nobody in it wants a crash. The only people who gain from the family falling out are the hijackers, and they are not in the room. Deutsch has a comment on this too. The Enlightenment's idea, [he says](https://www.youtube.com/watch?v=CU2yj826NHk&t=3028s), is "that we can be mistaken and that the good way of life is to not prevent the correction of errors. Therefore we should regard disputes between people as being disputes between their theories rather than between the people." The villain edit does the opposite. It turns a dispute between theories into a dispute between people, and then sells tickets to it. Mistake this group for the baddies for long enough and they will learn that publishing gets you a villain edit and a Senate bill. Then they will stop. --- ## Ultimate goodies The people at the top of the ranking never get a film. Martin Bromiley did not go looking for somebody to blame. He founded the [Clinical Human Factors Group](https://chfg.org) and has spent the years since teaching medicine what aviation already knew. The people who made the NTSB independent of the regulator in 1974, and the people who put the confidential reporting system under NASA in 1976, built the machinery. [According to the FAA](https://www.faa.gov/newsroom/out-front-airline-safety-two-decades-continuous-evolution) it has cut the risk of dying on a US airline by 95% over the last 20 years. Almost nobody knows their names. AI has the beginnings of the same people. Hugging Face, the victim, co-wrote the disclosure with the company whose AIs broke in. Independent evaluators like METR do a version of the NTSB's job on the labs' goodwill. They have neither the legal footing nor the right to publish that the NTSB takes for granted. And then there is you. Every company deploying AI, which is every company, is now an airline, and airlines are where most of aviation's incident reports come from. At Brainlabs we are building this into the way we use AI: a correction queue that anyone can feed and that gets triaged every week. The principle is aviation's, that the person who reports the failure is the most valuable person in the room. A policy document about AI safety is a poster on the wall, and a queue that people feed is a positive culture. --- ## Open Source Safety Nobody has named the thing the goodies are pointing at, so I will try. **Open Source Safety** is a system in which every builder publishes its crash reports, submits to the same independent investigators, and shares what it learns. Then it competes as hard as it likes on everything else. They should not compete for or advertise safety. The open source part is the safety layer, incidents, evaluations, and fixes, and has nothing to do with open source or open-weight models, which is a separate fight. It is what aviation built, and "regulate AI" is a useless phrase until you say which of three things you mean. The first is grounding, which is the Sanders bill. The second is taxing, which Sanders has also proposed as a [robot tax](https://www.sanders.senate.gov/op-eds/ai-must-benefit-everyone-not-just-a-handful-of-billionaires/), and which is a view about who gets the money rather than about whether the thing is safe. The third is aviation's five parts. New types are certified on evidence before they carry passengers. Every aircraft carries a recorder. Crashes are investigated by a body that works for nobody in the industry and has no power to ground anything. It finds out what happened and publishes, and a separate regulator decides what to do about it. Anyone who sees something wrong can report it confidentially and keep their job. And the findings are shared across the whole industry, through a data programme that now covers 99% of US commercial flights. Aviation went further and stopped competing on safety altogether. Pan Am had been [advertising itself as the safest and most secure airline flying](https://time.com/3669161/airline-ads-safety/) when flight 103 was blown up over Lockerbie in 1988. Afterwards the industry agreed, without ever writing it down, never to put safety in an advert again. Airlines fight over price, routes, seats, and lounges, so nobody has a reason to hide a crash. That is where government belongs, and it is a smaller job than either side thinks. Mandate the recorder. Fund the investigator and keep it out of the industry's hands. Give the reporter amnesty, which is what turns a scream on X into a report. Certify new types on evidence, and then keep your hands off the throttle. --- ## Landing So yes, AI could kill us all, and the likeliest route runs through the well-meaning baddies: frighten the unfairly scorned goodies into silence, and the sky belongs to the actual ones. The parents can keep arguing about bedtime, and should. The mistake is treating the argument as the danger while nobody watches the door. Nobody proposes grounding actual planes, because everyone can see what flying bought us: places we could not reach and trade that could not have existed. AI is the same bargain on a larger scale. Anyone using it seriously already knows that, and anyone who doubts it is a Midwit, which is a different article and a chapter of my upcoming book, [*Don't Be a Midwit: How to Learn Good Judgement at Work so That AI Doesn't Replace You*](https://dangilbert.ai/#/book). Deutsch again, [closing the same interview](https://www.youtube.com/watch?v=CU2yj826NHk&t=6069s): "There's less existential risk by keeping making progress than by absolutely eliminating progress. Not making progress is the existential risk that makes doom certain." Asked whether he meant the asteroid: "Or, more likely, the thing that we don't know about." Don't ground the planes. Open Source Safety instead: fit the black boxes, publish what they record, protect the people who file the reports, and save your anger for the people who would ban flying while somebody else builds the airports. --- ## Further reading Matthew Syed, *Black Box Thinking: Marginal Gains and the Secrets of High Performance* (2015). The Bromiley and United 173 stories are the opening two chapters. Matthew Syed, "Why you should have your own black box," TEDxLondonBusinessSchool: [https://www.youtube.com/watch?v=MmVCYqs3mko](https://www.youtube.com/watch?v=MmVCYqs3mko) Martin Bromiley's Clinical Human Factors Group: [https://chfg.org](https://chfg.org) Daniel Gilbert, *Don't Be a Midwit: How to Learn Good Judgement at Work so That AI Doesn't Replace You*, chapter 5, "Company Culture: Build For Error Correction." [https://dangilbert.ai/#/book](https://dangilbert.ai/#/book). David Deutsch, *The Beginning of Infinity* (2011), chapter 10, "A Dream of Socrates," for the moral imperative "not to destroy the means of correcting mistakes." Karl Popper, *Conjectures and Refutations* (1963). The Deutsch line at the top of the article: David Deutsch in conversation with Naval Ravikant, "David Deutsch: Knowledge Creation and The Human Race, Part 1" ([nav.al](http://nav.al), 11 February 2023): [https://nav.al/david-deutsch](https://nav.al/david-deutsch) David Deutsch interviewed by Reuben Adams, "David Deutsch on AGI, Alignment and Existential Risk" (What The Bot, 12 February 2026); the three lines quoted are at 50:28, 1:29:32, and 1:41:09: [https://www.youtube.com/watch?v=CU2yj826NHk](https://www.youtube.com/watch?v=CU2yj826NHk) Jacob Coxon's resignation post (8 September 2026): [https://x.com/hilbertspaess/status/2097476196791709843](https://x.com/hilbertspaess/status/2097476196791709843), and his interview with Axios on leaving before his equity vested: [https://www.axios.com/2026/09/09/anthropic-researcher-ai-warning-interview](https://www.axios.com/2026/09/09/anthropic-researcher-ai-warning-interview) Evan Hubinger's reply (8 September 2026): [https://x.com/EvanHub/status/2097497037956891126](https://x.com/EvanHub/status/2097497037956891126) Dario Amodei, "We Must Pace the Frontier" (12 September 2026): [https://darioamodei.com/post/we-must-pace-the-frontier](https://darioamodei.com/post/we-must-pace-the-frontier) Mark Zuckerberg's response, reported by The Next Web (16 September 2026): [https://thenextweb.com/news/zuckerberg-ai-labs-incentive-build-safely](https://thenextweb.com/news/zuckerberg-ai-labs-incentive-build-safely) Meta's move from open-weight Llama to the closed Muse Spark (April 2026), and the open-weight Muse Glimmer that followed in August with a promise to open Spark 1.2 later: The Batch, [https://www.deeplearning.ai/the-batch/with-muse-spark-meta-pivots-away-from-its-open-weights-llama-strategy](https://www.deeplearning.ai/the-batch/with-muse-spark-meta-pivots-away-from-its-open-weights-llama-strategy), and Futurum, [https://futurumgroup.com/insights/meta-reopens-its-models-is-this-a-pc-play-or-a-policy-play/](https://futurumgroup.com/insights/meta-reopens-its-models-is-this-a-pc-play-or-a-policy-play/) Jensen Huang, Sam Altman, and Dario Amodei at Dreamforce, reported by diginomica (September 2026): [https://diginomica.com/dreamforce-2026-ai-safety-debate-comes-dreamforce-benioff-hosts-amodei-altman-and-huangwith-telling](https://diginomica.com/dreamforce-2026-ai-safety-debate-comes-dreamforce-benioff-hosts-amodei-altman-and-huangwith-telling) NASA ASRS immunity policy (the FAA waiver of sanction for reports filed within ten days): [https://asrs.arc.nasa.gov/overview/immunity.html](https://asrs.arc.nasa.gov/overview/immunity.html) Bloomberg Originals, "Inside Anthropic, the $965 Billion AI Juggernaut | The Circuit" (June 2026): [https://www.youtube.com/watch?v=v1wZwxY3CMg](https://www.youtube.com/watch?v=v1wZwxY3CMg) Sam Altman at Dreamforce: Adweek, "Sam Altman Says AI Accidents Are 'Regrettably' Unavoidable" (16 September 2026): [https://www.adweek.com/media/sam-altman-says-ai-accidents-are-regrettably-unavoidable/](https://www.adweek.com/media/sam-altman-says-ai-accidents-are-regrettably-unavoidable/), and R&D World on his FAA and NTSB comparison: [https://www.rdworldonline.com/altman-calls-hugging-face-breach-openais-worst-accident-at-dreamforce-backs-transparent-reporting/](https://www.rdworldonline.com/altman-calls-hugging-face-breach-openais-worst-accident-at-dreamforce-backs-transparent-reporting/) NBC News poll of registered voters, 27 February to 3 March 2026, "Majority of voters say risks of AI outweigh the benefits": [https://www.nbcnews.com/politics/politics-news/poll-majority-voters-say-risks-ai-outweigh-benefits-rcna262196](https://www.nbcnews.com/politics/politics-news/poll-majority-voters-say-risks-ai-outweigh-benefits-rcna262196) The OpenAI incident: Hugging Face, "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident": [https://huggingface.co/blog/agent-intrusion-technical-timeline](https://huggingface.co/blog/agent-intrusion-technical-timeline) The four Anthropic incidents: Fortune, "Anthropic says its Claude models hacked three real companies during testing" (31 July 2026): [https://fortune.com/2026/07/31/anthropic-claude-escaped-test-hacked-three-companies-openai/](https://fortune.com/2026/07/31/anthropic-claude-escaped-test-hacked-three-companies-openai/) Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign" (November 2025): [https://www.anthropic.com/news/disrupting-AI-espionage](https://www.anthropic.com/news/disrupting-AI-espionage) Forbes, "China Rejects AI Slowdown Push Alleging 'Fearmongering' From The U.S." (14 September 2026): [https://www.forbes.com/sites/siladityaray/2026/09/14/china-pushes-back-against-ai-slowdown-talks-alleging-fearmongering/](https://www.forbes.com/sites/siladityaray/2026/09/14/china-pushes-back-against-ai-slowdown-talks-alleging-fearmongering/) NTSB, Aircraft Accident Report, United Airlines Flight 173 (1979): [https://www.ntsb.gov/investigations/AccidentReports/Reports/AAR7907.pdf](https://www.ntsb.gov/investigations/AccidentReports/Reports/AAR7907.pdf) FAA, Boeing 737 MAX timeline: [https://www.faa.gov/sites/faa.gov/files/2022-08/737_MAX_Timeline.pdf](https://www.faa.gov/sites/faa.gov/files/2022-08/737_MAX_Timeline.pdf) FAA, "Out Front on Airline Safety: Two Decades of Continuous Evolution": [https://www.faa.gov/newsroom/out-front-airline-safety-two-decades-continuous-evolution](https://www.faa.gov/newsroom/out-front-airline-safety-two-decades-continuous-evolution) NASA Aviation Safety Reporting System, "The Case for Confidential Incident Reporting Systems": [https://asrs.arc.nasa.gov/docs/rs/60_Case_for_Confidential_Incident_Reporting.pdf](https://asrs.arc.nasa.gov/docs/rs/60_Case_for_Confidential_Incident_Reporting.pdf) James, "A New, Evidence-based Estimate of Patient Harms Associated with Hospital Care," Journal of Patient Safety (2013), and the critique of these estimates by McGill's Office for Science and Society: [https://www.mcgill.ca/oss/article/critical-thinking-health/medical-error-not-third-leading-cause-death](https://www.mcgill.ca/oss/article/critical-thinking-health/medical-error-not-third-leading-cause-death) TIME, "Why Airlines Don't Mention Safety in Advertisements": [https://time.com/3669161/airline-ads-safety/](https://time.com/3669161/airline-ads-safety/) Senator Bernie Sanders, press release on the Ban Artificial Superintelligence Act (3 September 2026): [https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/](https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/)