HOW DID WE FORGET - Part 3
The Machine
This publication is free. This year I’m adding audio versions so you can listen when reading isn’t an option. If the work means something to you, you can buy me a coffee to help keep it going. That’s it. Thanks for being here.
The audio:
In November 2024, President Biden met with Chinese President Xi Jinping on the sidelines of an economic summit in Lima, Peru. It was their last meeting before Biden left office. Among the topics they discussed was one that would have seemed like science fiction a generation ago: they agreed that the decision to launch nuclear weapons must remain under human control, not artificial intelligence.
It was the first time China had ever made such a statement publicly.
“It’s an important statement about AI and nuclear doctrine,” said Jake Sullivan, Biden’s national security adviser. An important first step between two nuclear powers to address a long-term strategic risk.
The agreement was nonbinding. There were no verification mechanisms, no enforcement provisions. But it represented something: an acknowledgment by the world’s two most powerful nations that a line existed that should not be crossed.
Here is what neither leader mentioned: by the time they shook hands in Lima, that line had already been erased.
THE BOTTLENECK
In 2021, a book appeared with an unwieldy title: “The Human-Machine Team: How to Create Synergy Between Human and Artificial Intelligence That Will Revolutionize Our World.” The author was listed under the pen name “Brigadier General Y.S.”
The author was Yossi Sariel, commander of Unit 8200—Israel’s equivalent of the NSA.
Sariel’s book described a problem. In modern warfare, he argued, there is too much data for humans to process. Too many potential targets. Too many variables. Humans are slow. Humans get tired. Humans hesitate.
In Sariel’s framework, humans weren’t an asset. They were a bottleneck.
His solution was a machine that could “rapidly process massive amounts of data to generate thousands of potential ‘targets’ for military strikes in the heat of a war.” A machine that would resolve what he called “the human bottleneck for both locating the new targets and decision-making to approve the targets.”
He wasn’t describing something theoretical. He was describing something that already existed.
TWENTY SECONDS
The machine is called Lavender.
What follows comes from six Israeli intelligence officers who spoke to journalists at +972 Magazine. All six served during the current Gaza war. All six had direct experience with the system. They spoke on condition of anonymity because what they described contradicts Israel’s official narrative about precision warfare and human oversight.
Lavender works like this:
Israeli intelligence maintains continuous surveillance on 2.3 million residents of Gaza. The system ingests everything: cell phone metadata, WhatsApp group memberships, social network connections, movement patterns, address changes. If you’re in a WhatsApp group with someone the system finds suspicious, that’s a data point. If you change phones frequently—which you might do if your home was bombed and you lost your belongings—that’s a data point. If you work for the Hamas-run civil administration, even as a clerk or traffic cop, that’s a data point. Hamas governs Gaza. Someone has to keep the lights on.
Lavender processes these inputs and assigns every person in Gaza a score from 1 to 100, expressing “how likely it is that they are a militant.”
37,000 Palestinians crossed whatever threshold the system set. They were marked for death.
The military checked the system’s accuracy by manually reviewing a sample. They found it was right about 90% of the time. Which means approximately 3,700 people on the kill list had loose or no connection to militant groups.
They deployed the system anyway.
Here is what happened next. A second AI system—named, with grotesque irony, “Where’s Daddy?”—tracked each marked individual until they returned home to their families. The system was designed to find targets at night, when everyone would be inside together.
One intelligence source explained the logic: “We were not interested in killing operatives only when they were in a military building or engaged in a military activity. On the contrary, the IDF bombed them in homes without hesitation, as a first option. It’s much easier to bomb a family’s home. The system is built to look for them in these situations.”
It’s much easier to bomb a family’s home.
When the target was located, the file went to a human operator for approval. This is the part of the process designed to satisfy legal requirements—the “human in the loop” that distinguishes an automated killing system from an autonomous one.
An intelligence officer designated “B.” described what that approval process looked like:
“I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added value as a human, apart from being a stamp of approval. It saved a lot of time.”
Twenty seconds. The only verification required was confirming the target’s gender. If the machine said militant and the operator confirmed male, the strike was authorized.
For junior militants, they used unguided 2,000-pound bombs, “dumb bombs” that destroy entire buildings. Why not precision weapons? “You don’t want to waste expensive bombs on unimportant people.”
The bomb fell. The building collapsed. No assessment was conducted afterward to determine whether the target had actually been killed, or to count the civilian dead. The operator moved on to the next file. There were always more files. “Because of the system,” one source explained, “the targets never end. You have another 36,000 waiting.”
Sometimes the target wasn’t even home when the bomb fell. The system had tracked him there, but he’d moved elsewhere.
“It happened to me many times that we attacked a house, but the person wasn’t even home. The result is that you killed a family for no reason.”
THE PATTERN
Read that sequence again. Surveillance of an entire population. Algorithmic scoring based on patterns no human fully understands. Mass flagging—37,000 people marked, thousands wrongly. Minimal human review: twenty seconds, gender confirmation only. Automated tracking to maximize casualties. No accountability, no assessment, no pause.
Now watch the same pattern appear somewhere else.
In February 2024—nine months before Biden and Xi agreed that humans must control nuclear weapons—the Pentagon confirmed that AI-developed targeting recommendations had been used in over 85 airstrikes in Iraq and Syria.
The system is called Maven. It started as a Pentagon initiative in 2017 to apply machine learning to drone surveillance footage. Google employees protested, and Google dropped the contract. Palantir picked it up.
Today Maven is a full-spectrum targeting system. It fuses satellite imagery, communications intercepts, and intelligence feeds into a single interface. Yellow boxes mark potential targets. Blue boxes mark friendly forces and no-strike zones. A senior targeting officer estimated that with Maven, he can process 80 targets per hour, versus 30 without it.
A U.S. Army officer learning the system described the rhythm of concurring with the algorithm’s recommendations: “Accept. Accept. Accept.”
Twenty seconds in Gaza. Accept, accept, accept in the Pentagon.
The machines recommend. The humans confirm. The bombs fall.
At the United Nations, the Pentagon has argued that human control of autonomous weapons is not required by international law.
COMING HOME
Here is where the story arrives at your door.
In 2012, Palantir Technologies—the company that now runs Project Maven—quietly embedded itself in the New Orleans Police Department. There was no public announcement, no contract reviewed by the city council. They partnered with a local nonprofit to sidestep oversight.
“We very much like to not be publicly known,” a Palantir engineer wrote in an internal email later obtained by journalists.
The system they built mapped webs of human relationships across the city. It assigned residents algorithmic risk scores. It helped police generate target lists—not for airstrikes, but for surveillance, for stops, for the kind of attention that follows you through your life.
New Orleans eventually shut it down after public outcry. So did Los Angeles, where Palantir’s software had been designating “chronic offenders”—disproportionately in minority neighborhoods.
But the appeal of the machine never faded. It just moved somewhere else.
In Chicago, they called it the Heat List.
Starting in 2012, the Chicago Police Department assigned every person arrested in the city a score from 1 to 500, estimating how likely they were to be involved in a shooting—as perpetrator or victim. The system tracked 400,000 people. Everyone fingerprinted since 2013 was in the database, scored by an algorithm.
If your score was high enough, police showed up at your home. Unannounced. They told you that you were “on the department’s radar.” They warned that “further criminal activity, even for the most petty offenses, will result in the full force of the law being brought down.”
These weren’t visits to people who had committed violent crimes. These were visits to people an algorithm predicted might commit violent crimes, or might become victims of them. People whose offense was having the wrong friends, living in the wrong neighborhood, being arrested once for something minor.
85% of the people with the highest risk scores were African American.
The system created its own feedback loop. People in overpoliced neighborhoods were more likely to be stopped. Being stopped raised their score. A higher score justified more surveillance. More surveillance meant more stops. The algorithm learned that being Black in certain zip codes was a risk factor, because being Black in certain zip codes meant more contact with police, and more contact with police was how you got into the database in the first place.
In 2019, a RAND Corporation study found the Heat List hadn’t reduced violence at all. Chicago quietly discontinued the program. But by then, hundreds of thousands of people had been scored, visited, warned—treated as future criminals by a machine that couldn’t explain its reasoning.
In New York, they built something even larger.
Clearview AI scraped more than 60 billion photographs from the internet—from Facebook, Instagram, LinkedIn, news sites, anywhere a face appeared online. They built a facial recognition system and sold it to police departments across the country. If you’ve ever posted a photo of yourself online, you’re probably in their database.
More than 7,000 people from nearly 2,000 public agencies have used Clearview to search through millions of American faces. The NYPD alone has run over 5,100 searches. Officers downloaded the app onto their personal phones, in violation of department policy, and ran searches on their own time—looking for suspects, but also, according to leaked records, looking up their own friends and family members.
In August 2020, the NYPD used facial recognition to identify Derrick Ingram, a 28-year-old Black man who co-founded a protest group called Warriors in the Garden. His alleged crime was shouting into a police officer’s ear with a bullhorn during a George Floyd demonstration two months earlier.
Fifty officers in riot gear surrounded his Hell’s Kitchen apartment. NYPD helicopters hovered overhead. An officer stood outside his door holding a facial recognition report with Ingram’s photo, captured by a photographer and shared online.
For a bullhorn.
“It’s so alarming,” said Albert Fox Cahn of the Surveillance Technology Oversight Project, “when we see the tools that are sold to the public as a way to fight terrorism and violent crime being used to silence dissent.”
That same summer, six federal agencies used facial recognition technology on George Floyd protesters: the FBI, the U.S. Marshals Service, the Bureau of Alcohol, Tobacco, Firearms and Explosives, the U.S. Capitol Police, the U.S. Park Police, and the U.S. Postal Inspection Service.
The Postal Inspection Service. Hunting protesters.
In Pittsburgh, police officers signed up for Clearview trial accounts during the protests, violating city policy and in some cases city law. Emails obtained through public records requests showed officers asking each other whether to hide their use of the technology. “When you write reports about identifying suspects through Clearview AI,” one asked, “do you name the program, etc.? Or is it better to keep it vague?”
THE DATABASE
In April 2025, Immigration and Customs Enforcement awarded Palantir a $30 million contract to build something called ImmigrationOS.
The system has three functions. First, it decides who gets deported and in what order. Second, it tracks whether people are “self-deporting”—leaving the country voluntarily—in near real-time. Third, it manages the entire deportation process from identification to removal.
It pulls data from passport records, Social Security files, IRS tax returns, license plate readers, biometric databases. It creates comprehensive profiles of individuals the government wants to find.
One ICE official described it as “Amazon Prime, but with human beings.”
But ImmigrationOS is just one piece of something larger.
In May 2025, The New York Times reported that the Trump administration was working with Palantir to build a centralized database linking information across federal agencies—IRS tax records, Social Security files, Medicare claims, student debt, bank account numbers, disability status. A single searchable system containing the government’s accumulated knowledge about every person in the country.
The work began under Elon Musk’s Department of Government Efficiency. Several DOGE staffers are former Palantir employees. The federal chief information officer is a former Palantir employee. In March 2025, Trump signed an executive order directing agencies to “eliminate information silos” and share data with each other.
“The unprecedented possibility of a searchable, ‘mega-database’ of tax returns and other data that will potentially be shared with or accessed by other federal agencies is a surveillance nightmare,” ten Democratic lawmakers wrote in a letter to Palantir’s CEO.
Palantir denied it. They said they were “not building a master database” and “neither conducting nor enabling mass surveillance of American citizens.”
But consider what already exists. Palantir’s software already runs at the Department of Homeland Security, at Health and Human Services, at multiple other agencies. ICE’s existing system already “enables users to make connections between subject records across various systems.” The company has received over $113 million in federal contracts since January 2025, plus a separate $795 million Pentagon contract.
“The ultimate concern,” said Cody Venzke of the ACLU, “is a panopticon of a single federal database with everything that the government knows about every single person in this country. What we are seeing is likely the first step in creating that centralized dossier on everyone in this country.”
The infrastructure doesn’t distinguish between immigrants and citizens. Between protesters and criminals. Between people who’ve done something wrong and people who’ve done nothing at all.
Once it exists, it can be used for anything. I wrote about this earlier in June:
Our Tech Overlords, Minority Report is Here
With each day we learn more and more about what DOGE was really up to, it was not finding fraud/abuse, saving money, it was just a data heist. After being feed years of lies about the lack of efficiency and abuses of the federal government, MAGA then cheered Musk and others on, never realizing they were giving up their free will to the technocrats. The …
Our Tech Overlords, Minority Report is Here
THE CONTRADICTION
Return to Lima.
In November 2024, Biden and Xi agreed that nuclear weapons must remain under human control. It was a statement of principle, a recognition that some decisions are too consequential to delegate to machines.
Two months earlier, at a summit in Seoul, nearly 100 countries had adopted a “Blueprint for Action” calling for human control over AI in military systems, particularly nuclear weapons. China refused to sign.
Two months later, in Lima, Xi changed course—or appeared to. The agreement with Biden was nonbinding, unenforceable, purely aspirational. But it established a norm: humans decide, machines advise.
And yet.
At the moment Biden and Xi shook hands, the Pentagon had already used AI targeting recommendations in 85 airstrikes. Project Maven was processing 80 targets per hour. In Gaza, Lavender was generating kill lists and humans were approving them in 20-second intervals. Across America, police departments were using algorithmic risk scores to decide who to visit, who to watch, who to treat as a threat before any crime occurred.
The principle they affirmed—that humans must remain in control—was already a fiction.
What does “human control” mean when a human spends 20 seconds confirming a machine’s recommendation? What does it mean when the approval process is Accept, Accept, Accept? What does it mean when a RAND study finds that an algorithmic system “did not reduce violence” but was deployed for seven years anyway, scoring hundreds of thousands of people, sending police to their doors?
Human control becomes human theater. A performance staged for legal and diplomatic purposes while machines make the actual decisions.
The system recommends. The human confirms. The bomb falls, or the police arrive, or the deportation file advances.
The bottleneck has been removed.
THE HUMAN or THE MACHINE
Every near-miss in nuclear history was averted because a human felt something was wrong.
Stanislav Petrov in 1983, watching his screen light up with what looked like five incoming American missiles. His training said launch. His system said launch. But something felt wrong. He hesitated. He was right—it was sunlight reflecting off clouds.
Boris Yeltsin in 1995, nuclear briefcase open, three minutes to decide. His generals said launch. His system said launch. But something felt wrong. He waited. He was right—it was a Norwegian research rocket.
The margin between existence and extinction wasn’t technology. It was hesitation. Human judgment. The irreducible moment when a person looks at what the machine is telling them and thinks: wait.
Petrov couldn’t explain why he hesitated. In interviews afterward, he talked about intuition, about something that didn’t add up. Five missiles didn’t make sense for a first strike. It wasn’t a logical deduction—it was a feeling.
The machine can’t feel that something is wrong.
The machine can calculate probabilities. It can identify patterns. It can flag anomalies. But it can’t exercise judgment in the way Petrov did—integrating technical data with strategic logic with something that might be called wisdom or might be called doubt.
The machine can’t hesitate.
And hesitation is the only thing that has ever saved us.
We are building a world without hesitation.
In military targeting, we call it efficiency. In policing, we call it prediction. In immigration enforcement, we call it streamlining. In nuclear command, we call it reducing decision time.
All of it points the same direction: toward systems where machines generate recommendations and humans provide signatures. Where the role of the person in the loop is not to decide but to authorize. Where twenty seconds is enough time. Where Accept, Accept, Accept is the rhythm of judgment.
The agreement in Lima was supposed to draw a line. Humans control nuclear weapons. AI advises but does not decide.
But look at what we’re building. Look at the kill chains and the risk scores and the targeting systems and the mega-databases. Look at the 20-second approvals and the officers asking whether to keep their methods vague.
The line doesn’t hold because we’ve already crossed it everywhere else. We’ve trained ourselves—trained our institutions, trained our reflexes—to trust the machine and provide the signature.
When the moment comes that actually matters, when the screen lights up with what looks like incoming missiles, when someone has six minutes or three minutes or thirty seconds to decide the fate of civilization—
What makes us think we’ll hesitate then?
What makes us think we’ll even remember how?
Part 4 will examine who built this world and why—the ideology that treats hesitation as inefficiency, human judgment as a bug to be fixed, and the elimination of friction as the highest value.
SOURCES
Biden-Xi Agreement / AI Nuclear Policy
NPR: “Biden and Xi take a first step to limit AI and nuclear decisions at their last meeting” (November 2024)https://www.npr.org/2024/11/16/nx-s1-5193893/xi-trump-biden-ai-export-controls-tariffs
Arms Control Association: “Artificial Intelligence and Nuclear Command and Control: It’s Even More Complicated Than You Think” (September 2025) https://www.armscontrol.org/act/2025-09/features/artificial-intelligence-and-nuclear-command-and-control-its-even-more
Brookings: “Steps toward AI governance in the military domain” (November 2025)https://www.brookings.edu/articles/steps-toward-ai-governance-in-the-military-domain/
Yahoo/Fortune: “China refuses to sign agreement to ban AI from controlling nuclear weapons” (September 2024 - REAIM Seoul summit) https://finance.yahoo.com/news/china-refuses-sign-agreement-ban-111948789.html
Texas National Security Review: “Artificial Intelligence and Nuclear Weapons: A Commonsense Approach” (July 2025) https://tnsr.org/2025/06/artificial-intelligence-and-nuclear-weapons-a-commonsense-approach-to-understanding-costs-and-benefits/
Lavender/Gaza AI Targeting
+972 Magazine: “’Lavender’: The AI machine directing Israel’s bombing spree in Gaza” (April 2024)https://www.972mag.com/lavender-ai-israeli-army-gaza/
The Guardian: “’The machine did it coldly’: Israel used AI to identify 37,000 Hamas targets”https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes
Project Maven/US Military AI
Bloomberg: “US Says It Used AI to Help Find Targets It Hit in Iraq, Syria and Yemen” (February 2024)https://www.bloomberg.com/news/articles/2024-02-20/us-says-it-used-ai-to-help-find-targets-it-hit-in-iraq-syria-and-yemen
Palantir/Domestic Surveillance
The Verge: “Palantir has secretly been using New Orleans to test its predictive policing technology”https://www.theverge.com/2018/2/27/17054740/palantir-predictive-policing-tool-new-orleans-nopd
The Intercept: “How the LAPD and Palantir Use Data to Justify Racist Policing” (January 2021)https://theintercept.com/2021/01/30/lapd-palantir-data-driven-policing/
New York Times: “Trump Taps Palantir to Compile Data on Americans” (May 2025) Referenced via https://www.newsweek.com/donald-trump-database-palantir-dystopian-alarm-2079688
Wired: “DOGE Is Building a Master Database to Surveil and Track Immigrants” (April 2025) Referenced via https://www.democracynow.org/2025/6/3/makena_kelly
American Immigration Council: “ICE to Use ImmigrationOS by Palantir” (August 2025)https://www.americanimmigrationcouncil.org/blog/ice-immigrationos-palantir-ai-track-immigrants/
Axios Denver: “ICE pays Palantir $30M to build new tool to track and deport immigrants” (May 2025)https://www.axios.com/local/denver/2025/05/01/palantir-deportations-ice-immigration-trump
Migration Policy Institute: “Seeking to Ramp Up Deportations, the Trump Administration Quietly Expands a Vast Web of Data” (June 2025) https://www.migrationpolicy.org/article/trump-ice-data-surveillance
AIM Media House: “Palantir Started By Spying on a City Now Sells AI for War” (April 2025)https://aimmediahouse.com/ai-startups/palantir-started-by-spying-on-a-city-now-sells-ai-for-war
Chicago Heat List
Chicago Tribune: “For years Chicago police rated the risk of tens of thousands being caught up in violence. That controversial effort has quietly been ended.” (January 2020) https://www.chicagotribune.com/news/criminal-justice/ct-chicago-police-strategic-subject-list-ended-20200125-spn4kjmrxrh4tmktdjckhtox4i-story.html
Filter Magazine: “The Lessons of Chicago’s Disastrous ‘Crime Prediction’ Experiment” (March 2020)https://filtermag.org/chicago-crime-prediction/
University of Chicago Legal Forum: “Constraining Big Brother: The Legal Deficiencies Surrounding Chicago’s Use of the Strategic Subject List” https://legal-forum.uchicago.edu/print-archive/constraining-big-brother-legal-deficiencies-surrounding-chicagos-use-strategic
Chicago Magazine: “The Contradictions of Chicago Police’s Secretive List” (August 2017)https://www.chicagomag.com/city-life/august-2017/chicago-police-strategic-subject-list/
Clearview AI / Facial Recognition
BuzzFeed News: “How A Facial Recognition Tool Found Its Way Into Hundreds Of US Police Departments” (April 2021) https://www.buzzfeednews.com/article/ryanmac/clearview-ai-local-police-facial-recognition
Gothamist: “NYPD Used Facial Recognition Technology In Siege Of Black Lives Matter Activist’s Apartment” (August 2020) https://gothamist.com/news/nypd-used-facial-recognition-unit-in-siege-of-black-lives-matter-activists-apartment
Vice: “Six Federal Agencies Used Facial Recognition On George Floyd Protestors” (July 2024)https://www.vice.com/en/article/six-federal-agencies-used-facial-recognition-on-george-floyd-protestors/
PublicSource: “Emails show Pittsburgh police officers accessed Clearview facial recognition after BLM protests” (May 2021) https://www.publicsource.org/pittsburgh-police-facial-recognition-blm-protests-clearview/
MIT Technology Review: “The NYPD used Clearview’s controversial facial recognition tool” (April 2021)https://www.technologyreview.com/2021/04/09/1022240/clearview-ai-nypd-emails/
Center for Democracy and Technology: “Limiting Face Recognition Surveillance: Progress and Paths Forward” (October 2022) https://cdt.org/insights/limiting-face-recognition-surveillance-progress-and-paths-forward/
Wikipedia: Clearview AI https://en.wikipedia.org/wiki/Clearview_AI
FBI/Domestic Surveillance
Rutherford Institute: “Trump’s Palantir-Powered Surveillance Is Turning America Into a Digital Prison” (June 2025)https://www.rutherford.org/publications_resources/john_whiteheads_commentary/trumps_palantir_powered_surveillance_is_turning_america_into_a_digital_prison
Harvard Civil Rights-Civil Liberties Law Review: “Minority Report: Why We Should Question Predictive Policing” https://journals.law.harvard.edu/crcl/minority-report-why-we-should-question-predictive-policing/
Nuclear Near-Misses
Stanislav Petrov incident (1983) - extensively documented
Boris Yeltsin incident (1995) - extensively documented
Annie Jacobsen, Nuclear War: A Scenario (2024)