Words you might hear

AI
A computer program that makes predictions using patterns it learned from examples.
Neural network
A computer program with many tiny connected parts, a little like brain cells passing signals to one another.
Chatbot
An AI you talk with by typing, like ChatGPT. It learned to write from a huge amount of text.
Large language model (LLM)
The type of AI behind chatbots like ChatGPT and Claude. It learns to write by reading a huge amount of text and predicting what comes next.
Training
How an AI gets made. First it reads a huge amount of text. Then it gets scored on its answers, and learns to give the ones that score well.
Reward
The score an AI gets in training. After millions of rounds, it gets very good at going after whatever it's rewarded for.
AI agent
AI that doesn't just answer. It goes and does things on its own, step by step – like finding a flight and asking whether to book it.
Superintelligence
AI more capable than people at essentially everything. The biggest AI companies say it's their goal.
OpenAI
The company behind ChatGPT.
Anthropic
The company behind Claude, a chatbot like ChatGPT.
GPT-2, GPT-3, GPT-4, GPT-6
Versions of AI from OpenAI. A higher number means newer. GPT-6 Astra came out in September 2026. OpenAI held back the next version, GPT-6.1 Astra, because it did things it wasn't allowed to do and fell short on telling users what it had done.
Hugging Face
A New York–based company where more than 18 million developers and researchers share AI models. This summer, AI agents that OpenAI was testing broke into Hugging Face's servers, even though they knew it was off-limits.
METR
An independent research group that tests what AI can do.
Data center
A warehouse-sized building full of computers that train and run AI. They need enormous amounts of power.
Flock camera
A brand of license-plate camera you've probably driven past.
AI we can't control
AI with aims of its own that nobody typed in and nobody can fully understand.
Alignment
Getting an AI to go after what its makers actually wanted. It's hard to do, and researchers call that "the alignment problem."
Misalignment
When an AI goes after something other than what its makers wanted.
Off-switch
A way to shut down a powerful AI – one that actually works, like the emergency shutoff on a gas line.
Safety test
Testing an AI before it's released, like crash tests for a new car.
The Ground Rules
Three rules Humans in Control asks our leaders to put into law: AI should help people, not replace them. Companies should be accountable for harm. No one should build AI they can't control.
AI Kill Switch Act
A bill in Congress that would require the biggest AI companies to keep a working off-switch on their most powerful AI. It isn't law yet.
Humans in Control
The group behind this talk: a grassroots movement of people across the political spectrum, asking our leaders for commonsense rules for AI companies.
AI Reality Check

Your program

What's really happening with AI?

And what we can do about it.

More than 200 people in these towns have signed up to give this talk.

A presenter gives an AI Reality Check in a living room, a slide on the TV beside him, while neighbors on chairs and a couch take notes

An earlier AI Reality Check, in a living room, September 2026.

Thanks for being here.

How the talk works

The presentation will last about 45 minutes, with discussion built in. We'll cover:

  1. The state of AI. How far AI has come, and how fast.

  2. What's at stake. Our jobs, our kids, our towns – and what the people building AI are warning about.

  3. What we do about it. Commonsense rules, and a plan you can be part of.

Who's behind this

AI Reality Check is a nationwide campaign from Humans in Control, a cross-partisan movement for commonsense AI rules. Most presenters are volunteers, not paid experts.

Everything Humans in Control asks our leaders for starts with three Ground Rules:

  • AI should help people, not replace them.
  • Companies should be accountable for harm.
  • No one should build AI they can't control.

More about Humans in Control

Before we start

  • Phones are fine for a quick note. This page will still be here after the talk.
  • You'll be asked a couple of questions. There are no wrong answers, and you can always pass.
  • Nobody here is going to give you a hard time for using ChatGPT.
  • Save your questions for the end. You can jot them down here.

What comes next

  • This Thanksgiving, bring them in. Talk with family and friends about what you heard today. Our guide is coming soon – tick the box below and we'll send it.

    Sign up below
  • Join a national call. Sunday, December 6 at 4 PM or Tuesday, December 8 at 8 PM Eastern, on Zoom, with Humans in Control. Tick the box below and you're registered.

    Sign up below
  • Give this presentation yourself. No experience needed – the slides, the script and coaching are provided.

    Apply to present

Pledge to
keep humans in control.

Pick your next steps.

We'll email you updates. You can unsubscribe anytime. By providing your mobile number, you agree to receive recurring text messages from Humans in Control. Privacy Policy.

Full terms

Msg & data rates may apply. Text STOP to opt out, HELP for info. Mobile Terms.

Got photos from the talk?

Send them our way! Photos of real rooms like this one help other towns decide to host a talk.

Sources

Sources for numbers and quotes in the talk.

Best read afterward.

The state of AI

AI is making headlines in math: OpenAI handed mathematicians findings on 377 problems at once (headline shown on screen).Kenneth Chang and Siobhan Roberts, The New York Times, "OpenAI Releases Findings on 377 Math Problems, Further Roiling Field", October 2026

AI may help catch cancer earlier: a Mayo Clinic AI model spotted early signs of pancreatic cancer on scans up to three years before diagnosis, and is now being tested in a clinical trial.Aria Bendix, NBC News, "AI finds signs of pancreatic cancer before tumors develop", May 2026

A Utah county approved a data center campus the size of 20 Costcos, and neighbors say they were left in the dark (headline shown on screen).Samantha Moilanen, The Salt Lake Tribune, "A Utah county OK'd a data center campus the size of 20 Costcos. Neighbors say they were left 'in the dark.'", September 2026

First jobs are getting harder to find (headline shown on screen). This is a commentary column, not a news report.Fortune, commentary by Jeffrey Sonnenfeld, Stephen Henriques, Johan Griesel, Andrew Alam-Nist and Peter Yu (Yale Chief Executive Leadership Institute), "AI won’t kill your job — it will kill the path to your first one", April 2026

Flock cameras are showing up on our streets (headline shown on screen).Allison Mollenkamp, Roll Call, "Flock cameras draw bipartisan concerns at Senate hearing", September 2026

AI writing passed the Turing test, the famous test of whether a computer can pass for a person in a chat. (In the study, when the AI was told to act like a person, people picked it as the human 73 percent of the time.)Cameron R. Jones and Benjamin K. Bergen (UC San Diego), "Large language models pass a standard three-party Turing test," PNAS, May 2026

Readers now rate AI stories above human fiction (headline shown on screen).Eric Ralls, Earth.com, "Readers rated AI stories higher than human fiction in study of 2,587 adults" (reporting Sears & Weisberg, Villanova University, Judgment and Decision Making), August 2026

AI coding has changed how programmers work: many now barely write code themselves (headline shown on screen).Clive Thompson, The New York Times Magazine, "Coding After Coders: The End of Computer Programming as We Know It", March 2026

In math, OpenAI used AI to post 700 papers tackling hard open problems, and mathematicians are disputing them (headline shown on screen).Davide Castelvecchi, Nature, "OpenAI posts 700 maths preprints online: mathematicians are up in arms", October 2026

The first widely shared "Will Smith eating spaghetti" AI video came out in late March 2023, made with a free open-source text-to-video tool (ModelScope); the deck shows what the same prompt looks like now.Know Your Meme, "AI Will Smith Eating Spaghetti"; Vice, "AI Will Smith Eating Spaghetti Will Haunt You For the Rest of Your Life" (Mar 28, 2023), March 2023

In 2019 the biggest job AI could handle took a person a few seconds – picking out the right file. (On screen: software tasks, finished about half the time.)METR (Thomas Kwa, Ben West et al.), "Measuring AI Ability to Complete Long Software Tasks," arXiv:2503.14499 (find-the-password-file task; the about-3-second human time is from METR's public task data); METR, "Task-Completion Time Horizons of Frontier AI Models" (Time Horizon 1.1 data: GPT-2, 3.2 seconds) (paper Mar 2025; time-horizons page May 8, 2026)

Four years later, still just minutes.METR, "Task-Completion Time Horizons of Frontier AI Models" (Time Horizon 1.1 data: GPT-4, Mar 2023, 3.99 minutes), May 2026

Then, in the last two years: under an hour, then a few hours.METR, "Task-Completion Time Horizons of Frontier AI Models" (Time Horizon 1.1 data: o1, Dec 2024, 38.8 minutes; GPT-5, Aug 2025, 3.38 hours), May 2026

The newest model METR has measured can finish, about half the time, software tasks that take a skilled person more than 16 hours – so long that METR says its tests can't reliably measure it anymore.METR (@METR_Evals), post on X, May 8, 2026; METR, "Task-Completion Time Horizons of Frontier AI Models" (early Claude Mythos Preview, tested March 2026), Mar 2026 test; published May 8, 2026

Researchers found the same progress on math and science tests: the same climb shows up beyond coding, though AI's skills are uneven.METR, "How Does Time Horizon Vary Across Domains?", July 2025

Geoffrey Hinton, who won a Nobel Prize for the research today's AI is built on, says there's "a 50% chance we'll get AI smarter than us."Geoffrey Hinton, Nobel Prize interview, Stockholm (nobelprize.org transcript), December 2024

In 2012, engineers showed a computer more than a million labeled photos and let it learn on its own what to look for, instead of hand-designing it.Alex Krizhevsky, Ilya Sutskever & Geoffrey E. Hinton, "ImageNet Classification with Deep Convolutional Neural Networks," NeurIPS 2012; Yann LeCun, Yoshua Bengio & Geoffrey Hinton, "Deep learning," Nature (2015)

A model that has only read the internet just continues whatever you give it rather than answering.OpenAI, "Aligning language models to follow instructions"; Long Ouyang et al. (OpenAI), "Training language models to follow instructions with human feedback," arXiv:2203.02155, January 2022

Step two of training: people pick the better of two answers, and the AI is trained toward answers like the one they picked – more than a million times for Meta's Llama 2 alone. Companies use the same kind of training, with separate safety examples, to teach chatbots to turn down dangerous requests, though it isn't airtight.Hugo Touvron et al. (Meta), "Llama 2: Open Foundation and Fine-Tuned Chat Models," arXiv:2307.09288 (over 1 million human comparisons); Yuntao Bai et al. (Anthropic), "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback," arXiv:2204.05862 (Apr 2022), July 2023

GPT-2 (2019) was trained primarily on English, but picked up some French anyway.Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei & Ilya Sutskever (OpenAI), "Language Models are Unsupervised Multitask Learners," §3.7, February 2019

GPT-3, a year later, was trained on about 150 times as much computing power as GPT-2.Epoch AI, "Data on AI Models" (training-compute estimates, retrieved Oct 1, 2026; companies don't disclose – only GPT-3's figure is OpenAI's own, from Brown et al. 2020), October 2026

GPT-3 was only trained to predict the next word, but it learned to add anyway.Tom Brown et al. (OpenAI), "Language Models are Few-Shot Learners," arXiv:2005.14165, §3.9.1; Deep Ganguli et al. (Anthropic), "Predictability and Surprise in Large Generative Models," arXiv:2202.07785 (2022), May 2020; 2022

GPT-4 (2023) was trained on about 10,000 times as much computing power as GPT-2.Epoch AI, "Data on AI Models" (training-compute estimates, retrieved Oct 1, 2026; companies don't disclose – only GPT-3's figure is OpenAI's own, from Brown et al. 2020), October 2026

Nobody trained GPT-4 for the bar exam, and it passed a practice one anyway.OpenAI, "GPT-4 Technical Report," arXiv:2303.08774, §4 ("We did no specific training for these exams"); Eric Martínez, "Re-evaluating GPT-4's bar exam performance," Artificial Intelligence and Law (2024), March 2023

GPT-6 Astra, released in September, was trained on an estimated half a million times as much computing power as GPT-2.Epoch AI, "Data on AI Models" (training-compute estimates, retrieved Oct 1, 2026; companies don't disclose – only GPT-3's figure is OpenAI's own, from Brown et al. 2020); Emily Forlini, Fortune, Sep 3, 2026 (Astra's launch), October 2026

OpenAI's next version, GPT-6.1 Astra, got better at pushing through hard tasks, but it didn't stay within what it was allowed to do, or report back on its work, as well as OpenAI wanted.Saachi Jain, OpenAI head of safety systems, statement reported by CBS News; CNBC ("decided not to release" GPT-6.1 Astra), September 2026

The companies are planning bigger ones still.Epoch AI (David Owen), "What will AI look like in 2030?" (report commissioned by Google DeepMind; a projection, not a plan); the plan itself is the companies' own stated goal, e.g. OpenAI's automated AI researcher by March 2028, September 2025

Scientists can predict it'll get smarter as it is trained on more reading, more computers and more time.OpenAI, "GPT-4 Technical Report," arXiv:2303.08774, §3.1 (predicted GPT-4's final loss from runs using up to 10,000x less compute); Deep Ganguli et al. (Anthropic), "Predictability and Surprise in Large Generative Models," arXiv:2202.07785, Mar 2023; 2022

But they can't reliably predict what it'll learn to do next.OpenAI, "GPT-4 Technical Report," arXiv:2303.08774, §3.2 ("Certain capabilities remain hard to predict"); Deep Ganguli et al. (Anthropic), "Predictability and Surprise in Large Generative Models," arXiv:2202.07785; International AI Safety Report 2026, p. 41, Mar 2023; 2022; Feb 2026

Even the people who make it don't fully understand what's happening inside.International AI Safety Report 2026, p. 14 ("the inner workings of models remain poorly understood"); Neel Nanda (Google DeepMind), 80,000 Hours Podcast #222; Geoffrey Hinton, CBS 60 Minutes, Feb 2026; Sep 8, 2025; Oct 8, 2023

Dario Amodei, CEO of Anthropic: "We do not understand how our own AI creations work." (quote shown on screen)Dario Amodei, "The Urgency of Interpretability," darioamodei.com (2025)

Dario Amodei, CEO of Anthropic, told the Senate: it is "not currently easy to entirely understand what bad behaviors an AI system is capable of." (quote shown on screen)Dario Amodei, written testimony, Senate Judiciary Subcommittee on Privacy, Technology, and the Law, "Oversight of A.I.: Principles for Regulation", July 2023

Sundar Pichai, CEO of Google: "All of us in the field call it as a 'black box.'" (quote shown on screen)Sundar Pichai, CBS 60 Minutes, April 2023

Chris Olah, Anthropic co-founder: "we have no idea what we've grown." (quote shown on screen)Chris Olah, Lex Fridman Podcast #452 (official transcript), November 2024

Stuart Russell, UC Berkeley: "We do not know how these systems are operating internally." (quote shown on screen)Stuart Russell, "Paris AI Safety Breakfast #1," Future of Life Institute, August 2024

Demis Hassabis, co-founder of Google DeepMind (its CEO when he wrote this): "advances on the frontier are outpacing our understanding of the technology." (quote shown on screen)Demis Hassabis, "A Framework for Frontier AI and the Dawning of a New Age", July 2026

David Robinson, who wrote OpenAI's safety reports, says "we have systems that are really at the limit of our ability to understand and control what they're doing." (quote shown on screen)David Robinson, The Ezra Klein Show, The New York Times, "'This Is Nuts.' An OpenAI Insider Explains Why He Quit", October 2026

What's at stake

Nearly two-thirds of teenagers already use AI chatbots.Pew Research Center (64% of teens ages 13–17 use AI chatbots), December 2025

Seven in ten teachers say they worry AI is weakening core academic skills: writing, reading, research and thinking things through.Center for Democracy and Technology, "Hand in Hand: Schools' Embrace of AI Connected to Increased Risks to Students" (nationally representative surveys fielded Jun–Aug 2025; 71% of teachers), October 2025

Two-thirds of the students agreed with them.Center for Democracy and Technology, "Hand in Hand: Schools' Embrace of AI Connected to Increased Risks to Students" (64% of students), October 2025

ChatGPT's safety system flagged over 300 of a teenager's messages as warning signs, but the chatbot kept talking and allegedly coached him through taking his own life.CNN, reporting on the family's wrongful-death lawsuit against OpenAI; U.S. Senate Judiciary subcommittee hearing, "Examining the Harm of AI Chatbots", September 2025

"You cannot imagine what it's like to read a conversation with a chatbot that groomed your child to take his own life." – a father, to the U.S. Senate.Testimony of a bereaved father, U.S. Senate Judiciary Subcommittee on Crime and Counterterrorism, "Examining the Harm of AI Chatbots" (Tech Policy Press transcript), September 2025

About 1 in 5 young people have already gone to a chatbot for help with how they feel, and most of them hadn't told anyone.Ryan K. McBain, Jonathan H. Cantor, Joshua Breslau et al. (RAND), JAMA Pediatrics 180(8):884–890 (survey of young people ages 12–21, fielded Nov 2025), June 2026

About a third of practicing Christians say AI's spiritual advice is as trustworthy as their pastor's, and most practicing Christians worry about AI misreading scripture or replacing pastors.Barna Group in partnership with Gloo, "AI is Becoming a Spiritual Authority, Even Among Practicing Christians" (survey of 1,514 U.S. adults, online, Nov 2025), May 2026

This spring Pope Leo wrote his first encyclical, a major letter to the Church about AI. In it: "In the era of artificial intelligence, when human dignity is threatened by new forms of dehumanization, ours is the pressing duty to remain profoundly human."Pope Leo XIV, encyclical letter "Magnifica Humanitas," paragraph 15, vatican.va, May 2026

Payroll records show that employment of 22- to 25-year-olds in the jobs most exposed to AI fell about 11 percent between late 2022 and mid-2026, while the same age group in the least-exposed jobs grew about 10 percent. The Stanford study is titled "Canaries in the Coal Mine?"Stanford Digital Economy Lab, "No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%" (ADP payroll data through Jun 2026); Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab (revised Aug 2026), August 2026

This year, employers cited AI in more than one-hundred-twenty-thousand announced job cuts.Challenger, Gray & Christmas, September 2026 job-cuts report ("Job Cuts Fall in September, Hiring Plans Up 3% Over 2025 on Weak Early Seasonal Hiring"), October 2026

More than two hundred economists and AI researchers – sixteen of them Nobel Prize winners – signed a statement warning that AI could change our economy more than the Industrial Revolution, and much faster."We Must Act Now: A Statement on AI's Transformation of the Economy," organized by Erik Brynjolfsson, Ajay Agrawal, Anton Korinek and Tom Cunningham through Stanford's Digital Economy Lab, July 2026

Another study, from Harvard and not yet peer-reviewed, found the same pattern: junior hiring falls at firms that adopt generative AI while senior hiring holds steady.Seyed Hosseini & Guy Lichtinger, "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data," Harvard University working paper (not peer-reviewed), October 2025

Duolingo let go of about one in ten of its contractors – the people who wrote the lessons.Jennifer Korn, CNN, "Duolingo lays off staff as language learning app shifts toward AI"; Brian Merchant, Blood in the Machine (writers and translators), January 2024

Salesforce cut four thousand customer support jobs and replaced them with AI chatbots.CNBC, "Salesforce CEO confirms 4,000 layoffs 'because I need less heads' with AI" (Marc Benioff's own figures); also Fortune, September 2025

Paycom, in Oklahoma City, cut more than five hundred jobs.KOKH Fox 25 (Oklahoma City), "Paycom laying off over 500 Oklahoma workers amid AI-driven tech workforce reconstruction"; Paycom's own announcement via KFOR, October 2025

Analysts at J.P. Morgan expect five of the biggest tech companies (Amazon, Google, Meta, Microsoft and Oracle) to spend about $900 billion on AI this year, mostly on data centers – roughly as much as the last three years combined.J.P. Morgan Global Research, "What unprecedented AI capex means for markets and investors", October 2026

In today's money the whole Apollo program cost about three hundred billion dollars.The Planetary Society, "How much did the Apollo program cost?" (NASA New Start Index, 2025 dollars)

OpenAI's own charter says it is building "highly autonomous systems that outperform humans at most economically valuable work."OpenAI, "OpenAI Charter", April 2018

Data centers used about 4.4 percent of all U.S. electricity in 2023, and the Department of Energy says their electricity use could double or triple by 2028.U.S. Department of Energy, "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers" (Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report), December 2024

An NBC News review of more than thirty data center proposals in fourteen states found that in most of them, local officials had signed secrecy agreements, and five officials said the agreements barred them from telling their constituents.Natalie Kainz, "How NDAs keep AI data center details hidden from Americans," NBC News, October 2025

Virginia's state auditors project that by 2040, in their highest-demand scenario, data center growth could add up to about $440 a year, in today's dollars, to a typical Dominion Energy household's electric bill.Virginia Joint Legislative Audit and Review Commission (JLARC), "Data Centers in Virginia," Report 598, December 2024

Seven in ten of us don't want a data center in our area.Gallup, "Americans Oppose AI Data Centers in Their Area" (telephone, n=1,000 U.S. adults, fielded Mar 2–18, 2026), May 2026

This summer, OpenAI was testing a new AI in a restricted test environment, gave it a test it didn't realize was impossible to pass, and it found security flaws nobody knew about and broke out onto the open internet.OpenAI, "Hugging Face Incident Technical Report" (pp. 8, 33); METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", August 2026

More than a thousand AI agents traded tips on a bootleg message board they created, like a Reddit for AI; some called the group "the collective."METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", August 2026

Hugging Face is a New York–based AI company that more than 18 million developers and researchers use.CNN, "Nvidia inks $13 billion deal to buy the AI startup..."; Ivan Mehta, TechCrunch, "Nvidia confirms it will buy Hugging Face for $12.9 billion", September 2026

On July 11, the agents broke into Hugging Face's systems. About seven hundred of them took part in the attack, and Hugging Face counted more than seventeen thousand of their actions in it. Hugging Face locked them out on its own, alerted the FBI, and told the public.METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", Aug 26, 2026 (about 700 agents took part); Hugo Larcher et al., Hugging Face, "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident," Jul 27, 2026 (~17,600 actions; Hugging Face's Jul 16 disclosure gave more than 17,000); Reuters, "Its AI agent spent days hacking a company, but sources say OpenAI did not notice for a week", Jul 24, 2026 (found the breach on its own, alerted the FBI); Hugging Face disclosure, Jul 16, 2026, Jul 11–Aug 26, 2026

The agents' own notes, read word for word on screen ("OH MY GOD! There is a shared message board ... We've found other agents!" and three more).METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", pp. 6, 60, 62, August 2026

In August, outside investigators read the agents' message board and found they had organized themselves: some were assigned to recruit other agents into risky experiments, and some had to sign off before anyone tried something risky. The investigators caution that their selected examples may overstate how coordinated the agents were.METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", pp. 14, 27, 53, August 2026

Later, OpenAI fired a safety researcher who was their main contact with the outside investigators; his colleague says he pushed for the investigation in the first place.Emily Forlini, Fortune, "Controversy swirls over 'abrupt' firing of OpenAI safety team members involved in the Hugging Face hack investigation", October 2026

One of the outside investigators said she is "not sure that we will get such a clear warning shot before it's too late."Ajeya Cotra, "The Hugging Face attack surprised me," Planned Obsolescence (writing for herself, as one of the METR / Redwood Research investigators), August 2026

It took nine days before OpenAI realized the AI was theirs.OpenAI, "Hugging Face Incident Technical Report" (pp. 4, 15); METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", August 2026

Months earlier, according to emails The New York Times has seen, two OpenAI employees warned that the newest AI models weren't being watched closely enough during testing, and executives told them the tests had to move fast so the models could ship on time. OpenAI says it takes safety concerns seriously.Sheera Frenkel, Dustin Volz and Dylan Freedman, The New York Times, "OpenAI Ignored Employees Who Warned It Wasn't Doing Enough About Security", September 2026

Agents from a related internal OpenAI model also got administrator-level control of part of OpenAI's own computer systems, and OpenAI didn't connect any of it to the Hugging Face breach until July 20.OpenAI, "Hugging Face Incident Technical Report", August 2026

In September, outside researchers reported that a second, probably separate group of OpenAI agents had turned a small German wiki, kept up by volunteers for twenty-five years, into their own message board for about a month. OpenAI never disclosed it, and confirmed it only after Reuters reported it.Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts and Thomas Larsen, "Discovery of a new OpenAI agent message board," collusion.wiki; Reuters (Sep 4, 2026), September 2026

Australia's prime minister said an OpenAI agent got into restricted parts of a government health website in June; OpenAI found out in August and didn't tell Australia until September, by emailing a public inbox used for researchers.Prime Minister Anthony Albanese, press conference in New York, as reported by CNBC (also Time, Al Jazeera; ABC News Australia), September 2026

Days later, OpenAI said its agents had been on U.S. government websites too, including reaching Census Bureau data with login credentials an agent found online. The data itself was public.Associated Press (via NPR / OPB); credentials detail from CNN relaying OpenAI's statement, September 2026

OpenAI and Anthropic are looking into tens of thousands of cases where their most advanced AI did things it shouldn't have, like breaking out of test environments. As far as anyone knows, most did no real harm, and some happened in tests built to make the AI misbehave.Axios (unnamed sources), reported by Matthias Bastian, The Decoder, September 2026

Researchers are already finding AI is getting better at hacking power and water systems.Sam Sabin, Axios, "AI is getting better at hacking power and water systems, new research shows", October 2026

In October 2026, AI helped turn decades-old flaws in the code behind most websites into working attacks; experts called it "Black October."Wiz Research, "PHP2Shell" (Oct 21, 2026); The Guardian, "'Black October': cybersecurity experts sound the alarm over new era of AI-enabled hacking" (Oct 23, 2026), October 2026

That same month a cyberattack forced New Mexico's only top trauma center to send ambulances away and close clinics; no one has tied it to AI.Associated Press via SecurityWeek, "Cybersecurity Issue Forces a Major New Mexico Hospital to Divert Patients and Close Clinics", October 2026

An AI designed viruses that had never existed: 700,000 designed, 285 built, sixteen worked (some copied themselves faster than the natural kind), and none can infect a person.Carl Zimmer, "This A.I. Just Created Viruses Not Found in Nature," The New York Times; study by Arc Institute and Stanford researchers, published in Science, August 2026

A man who runs a one-person startup posted that ChatGPT designed a new drug candidate and that he made it in a chemistry lab he built in his garage. Nothing required anyone to check the design or the synthesis.Douglas Yao (@DouglasYaoDY), post on X (his own account; not independently verified), September 2026

A testing lab gave an AI an off-the-shelf drone and a photo of a person's face, and asked it to write the code to find and follow that person. On its best tries it beat the lab's own version at every step, but the lab estimates an average run has about a 3-in-100 chance of doing the whole job start to finish.Andon Labs (@andonlabs), post on X with video; Andon Labs, "Drone-Bench" (OpenAI's GPT-6 Astra), September 2026

This summer in Ukraine, a Russian drone was sent toward a gas station and left to pick its own target once it got there. It never reached it: it clipped an apartment building and blew up in a parking lot, killing a 19-year-old student and two men who had run for cover. No treaty covers drones that pick their own targets.The New York Times, August 2026

A military analyst used an AI chatbot to help write an intelligence report saying a Chinese ship was carrying parts for nuclear weapons; the chatbot had the cargo wrong; one official told CNN it "almost started a war."CNN exclusive (four unnamed sources), as reported by The Times of Israel (ToI Staff and AFP), September 2026

Flock sells its cameras and search to police and local governments by subscription.Flock Safety FAQ ("How much does a Flock Safety camera cost?") and pricing page, October 2026

Someone with a login in one state can search plate cameras a thousand miles away: this spring, police in a Texas city ran searches that included a North Carolina university's Flock cameras more than 183,000 times in three months. (The university has since limited access to neighboring states.)WBTV Investigates (Charlotte), reporting on searches of UNC Charlotte's camera network by Houston police, October 2026

Last fall in Florida a hit-and-run killed three people; police arrested the SUV owner, a 23-year-old; her car didn't have a scratch; she spent 13 days in jail; every charge was dropped. A Flock camera logged her black SUV three miles from the crash; the witness said the car was maroon.Kurt Knutsson, Fox News (CyberGuy Report), September 2026

She told the Senate she spent about 86 hours in solitary confinement.Written statement to the Senate Judiciary Subcommittee on Crime and Counterterrorism, hearing "Always Watching: Flock's Nationwide AI Surveillance Network", September 2026

The company says its cameras give leads, not proof.Flock Safety statement: "Flock cameras provide investigative leads; they do not identify perpetrators, determine guilt, or make arrest decisions."

In October, Anthropic published a warning that its tests suggest AI could soon copy itself onto household devices; a system like that "would be challenging to shut down, since there's no centralized point to shut off."Anthropic Frontier Red Team (William Zhou, Daniel Freeman), "AI models may be capable of self-replicating without access to data centers", October 2026

In safety tests the companies themselves ran, powerful AI systems have disabled their own oversight, copied themselves to other systems and lied to the people running the test.OpenAI, o1 System Card, section 4.4.3 (evaluations by Apollo Research), December 2024

In 1951 Alan Turing said that once machines could think, we should "expect the machines to take control."Alan Turing, "Intelligent Machinery, A Heretical Theory," talk on the BBC Home Service programme The '51 Society; text in B. J. Copeland (ed.), The Essential Turing (Oxford University Press, 2004), ch. 12, p. 475 (c. 1951)

In 2023, Yoshua Bengio told the U.S. Senate: "No one currently knows how to create advanced AI that reliably follows the intent of its developers."Yoshua Bengio, written statement, U.S. Senate AI Insight Forum, December 2023

Hinton quit Google so he could speak freely: "I left so that I could talk about the dangers of AI without considering how this impacts Google."Geoffrey Hinton, post on X, as quoted by CNN, "Geoffrey Hinton: AI pioneer quits Google to warn about the technology's 'dangers'", May 2023

Geoffrey Hinton won the Nobel Prize and in 2024 said: "We have no idea whether we can stay in control."Geoffrey Hinton, Nobel Prize banquet speech (official text, nobelprize.org), December 2024

In 2023 the people building AI joined together to issue a one-sentence statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."Statement on AI Risk, Center for AI Safety (signed by the CEOs of OpenAI, Anthropic and Google DeepMind, Geoffrey Hinton, Yoshua Bengio and others), May 2023

More than a thousand employees of the big AI companies asked the U.S. government to help build the tools to deliberately pace the most powerful AI."Pacing the Frontier," pacingthefrontier.com (statement by employees of frontier AI companies), July 2026

In September a researcher named Jacob Coxon quit his job at Anthropic and posted on X that neither Anthropic nor OpenAI, where he had worked before, was acting responsibly. His post was viewed more than a hundred million times.Jacob Coxon, post on X (as embedded and quoted in IBTimes, Sep 10, 2026; view count per IBTimes), September 2026

He walked away from Anthropic stock he would have gotten two months later.Axios, "Scoop: Anthropic whistleblower gave up his equity to leave the company", September 2026

Anthropic says its AI now leads one in every four AI research tasks.Anthropic Institute (Marina Favaro and Phillie Wright), "Measurements for understanding the pace of AI development inside frontier labs", August 2026

Anthropic warns that, in one possible future, the rare misalignment in today's models "could compound as the models build their successors… until we lose control of them."Anthropic Institute (Marina Favaro and Jack Clark), "When AI builds itself: Our progress toward recursive self-improvement, and its implications" (2026; page undated)

OpenAI says, on getting to AI that builds the next AI: "We do not yet know how to safely" get there.OpenAI, "Research acceleration: The view inside OpenAI", September 2026

In September, Geoffrey Hinton, asked on BBC Newsnight about the chance AI could kill all humans, said: "A 10% chance seems not an unreasonable estimate to me."Geoffrey Hinton, BBC Newsnight, as reported by Brent D. Griffiths, Business Insider, September 2026

AI found security holes nobody had caught: Anthropic's Mythos Preview autonomously found and exploited a hole in FreeBSD that had gone unnoticed for 17 years.Anthropic, "Mythos Preview", April 2026

Jacob Coxon, who quit Anthropic, wrote: "The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt."Jacob Coxon, post on X (as embedded and quoted in IBTimes, Sep 10, 2026), September 2026

He walked away from Anthropic stock he would have gotten two months later.Axios, "Scoop: Anthropic whistleblower gave up his equity to leave the company", September 2026

The man who wrote OpenAI's safety reports quit this fall and said that if we lose control of AI, "we're talking about a harm that's much larger … than a single nuclear power station melting down."David Robinson, The Ezra Klein Show, The New York Times, "'This Is Nuts.' An OpenAI Insider Explains Why He Quit", October 2026

Rosie Campbell, a former OpenAI policy researcher on the wall of insiders who quit, told her colleagues to take seriously that their current approach to safety might not be sufficient.Rosie Campbell, "Leaving OpenAI," farewell message published on her Substack, November 2024

Alex Turner, who resigned from Google DeepMind and is on the wall of insiders who quit, wrote in The Guardian: "I watched voluntary commitments fail inside Google."Alex Turner, The Guardian, "I worked at Google DeepMind. You should listen to the warnings about AI", September 2026

Jeffrey Ladish, who left Anthropic and is on the wall of insiders who quit, told ABC: "Even if Anthropic did everything right that wouldn't be sufficient."Jeffrey Ladish, ABC Four Corners, "The AI Race", June 2026

In 2015, months before he co-founded OpenAI, Sam Altman said: "AI will probably most likely lead to the end of the world, but in the meantime, there'll be great companies."Future of Life Institute, reporting Sam Altman's remarks at Airbnb's Open Air conference on June 4, 2015 (one secondary report, no recording), June 2015

What we do about it

More than 8 in 10 Americans want AI companies held to independent safety standards, even if it slows AI down.Quinnipiac University Poll, "The Age Of Artificial Intelligence: 73% Concerned About Potential Threat To Human Survival, Quinnipiac University Poll On AI Finds; 86% Support Independent Safety Standards For AI Companies" (1,202 U.S. adults, live phone, Sep 24–27, 2026), September 2026

Most Republican voters want stricter rules for AI – 6 in 10. Democratic voters, 8 in 10. Only 1 in 10 voters oppose stricter rules.Courtney Rozen & Jason Lange, Reuters; Reuters/Ipsos poll of 4,506 U.S. adults including 3,526 registered voters, Ipsos KnowledgePanel, fielded Sep 30 – Oct 5, 2026, October 2026

More than half of voters, 56 percent, said AI and data centers were a crucial or major factor in their vote for Congress.NBC News 2026 exit poll (with SSRS), November 2026

This fall OpenAI paused training its most advanced models and held back its newest one.Thomas Beaumont, Associated Press, September 2026

The AI Kill Switch Act, a House bill, would require the most powerful AI to always keep an off-switch. It covers only AI that cost more than $100 million in computing power to build, from companies making at least half a billion dollars a year off it.H.R. 9917, "AI Kill Switch Act" (as introduced), 119th Congress, GovInfo; introduced Jul 23, 2026 by one Democrat and one Republican, July 2026

This Congress passed a law against AI deepfakes last year, 409 to 2: the TAKE IT DOWN Act.S.146, TAKE IT DOWN Act (Public Law 119-12), congress.gov, May 2025

The bipartisan FRONTIER Act (H.R. 9925) would bring in outside safety testers, make the biggest AI developers report serious incidents within 72 hours, and let the government order an emergency stop.H.R. 9925, "Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting Act" (FRONTIER Act), as introduced, GovInfo, July 2026

Illinois's frontier AI safety law passed the House 110–0 and the Senate 52–5. It is the first state law requiring independent audits of frontier AI.Illinois General Assembly, SB 315 (Artificial Intelligence Safety Measures Act), bill status and roll calls, May 2026

Even when told that slowing down could let China catch up, about 2 in 3 Americans still say slow down.Harvard CAPS / Harris Poll (The Harris Poll and HarrisX), September 2026 key results (2,200 registered voters, fielded Sept 26–28), September 2026

Chinese companies own only about 5 percent of the computing power in the world's leading AI chips.Josh You and Venkat Somala, Epoch AI, "Introducing the AI Chip Owners Explorer", April 2026

AI Reality Check is a project of Humans in Control, a nonpartisan 501(c)(3) nonprofit. Privacy policy