The Longest Warning

At 3 a.m. on June 3, 1980, the national security adviser got a phone call telling him NORAD's computers were reporting 2,200 Soviet missiles inbound. He didn't wake the president, because minutes later it resolved as a false alarm. The cause, when the Pentagon briefed the press, was a failed integrated circuit in a Cheyenne Mountain computer. The New York Times headline that month reads like satire: "Missile Alerts Traced to 46¢ Item."

Hold that date. WarGames, the movie everyone remembers as a warning about computers and nuclear war, came out 3 years LATER. It didn't predict the near-miss. It dramatized one that had already happened, twice, and we know that because the incident reports are declassified and sitting on the National Security Archive's website.

That ordering, reality first, fiction second, policy third if at all, repeats for 183 years. I went looking for it because I kept hearing 2 versions of the AI story that can't both be true: "this all just started with ChatGPT" and "the government has had this forever." So I checked. 6 research lanes, 2 independent sources per claim, then an adversarial pass built to break my own findings. Here's what the record actually shows.

The proof, decade by decade

It starts before electricity was in homes. In 1843, Ada Lovelace published the first algorithm, for a machine that was never built, and wrote the first version of the debate we're still having: the Engine "has no pretensions whatever to originate anything." Turing thought that objection important enough to name it after her and answer it, in print, in 1950. The Pentagon later named its official programming language Ada, standard number MIL-STD-1815, her birth year. The military's own naming conventions know how old this story is.

The first computer was a secret for 30 years. Colossus was breaking German high-command ciphers in early 1944. It was programmable, electronic, and operational a half-decade before the machines in the textbooks. After the war the machines were dismantled and the program stayed classified until 1974. Every history of computing written before then was wrong about where the field started, and the people who knew couldn't say.

Then the state built the field, on purpose, on the record. In 1963 ARPA wrote MIT a $2.2M check and kept funding the MIT, Stanford, and CMU AI labs for decades. Federal money peaked at 66.8% of ALL US research and development in 1964. Shakey the robot (1966), speech recognition with a 1,011-word vocabulary (1976), the A* routing algorithm your phone still uses: all of it defense-funded, none of it hidden. This part was never secret. It was just ignored.

By the 1980s, autonomy was operational, not theoretical. The Phalanx gun system went to sea in 1980: it searches, tracks, evaluates, engages, and does its own kill assessment, and it is still deployed. Tomahawks were navigating themselves by terrain-matching in 1983. And on May 31, 1985, at a Martin Marietta test site in Colorado, a DARPA vehicle drove itself down a road. Self-driving is 41 years old. In 1995, a CMU neural network steered a minivan 2,850 miles from Pittsburgh to San Diego, 98% hands-off. That's not a demo I'm extrapolating from. That's a road trip that happened when gas was $1.15.

The 90s put AI into production on citizens, quietly. In March 1993 the Treasury switched on FAIS, an AI scoring roughly 200,000 wire transfers a week for money-laundering risk. It's documented in peer-reviewed AI conference papers from 1995. The same year, a counterdrug program called FERET started building the face-recognition benchmark that the entire modern industry grew from. Automated keyword filtering of global communications had been running since the 70s under ECHELON, exposed by a journalist in 1988 and confirmed, this is the part people forget, by a European Parliament investigation that adopted its final report 367 votes to 159 in 2001. A parliament. Not a forum thread.

The 2000s are when the pipeline reached your pocket. DARPA's CALO project (2003-2008) was the largest AI program in US history at the time, 300 researchers at SRI. SRI spun it out as a startup in 2007. Apple bought it in 2010. You know it as Siri. Eight years, defense lab to iPhone, and the lineage is stated in the first paragraph of Siri's own Wikipedia entry. Same decade: the CIA's venture arm seeded Palantir, PRISM started running 6 years before you heard the name, and a DEA program called Hemisphere was querying AT&T call records reaching back to 1987.

And the 2010s gave us the single detail I'd put on a billboard. From at least 2012, the NSA ran a machine-learning program over the phone metadata of 55 million people in Pakistan, scoring them on courier-like behavior. Internal slides revealed in 2015 show its name: SKYNET. The agency named its ML targeting program after the movie villain. When a Human Rights Data Analysis Group scientist reviewed the leaked methodology, his published assessment was that it was scientifically unsound; at the disclosed false-positive rates, thousands of innocent people would score as targets. Machine-learning target selection with disputed statistics wasn't a 2026 debate. It was running a decade ago.

The mirror runs both directions

So did we make movies about rumors, or build reality from movies? The dates say: both, in a loop, and every link is documented.

Fiction got there first only at the very beginning, before any computer existed. R.U.R., the 1921 play that coined the word "robot," ends with the robots destroying humanity, 25 years before ENIAC. Asimov published the Three Laws in March 1942, an alignment framework 14 years before the field had a name.

But in the cases everyone treats as prophecy, reality was already there. The real Colossus ran in 1944; the novel about a secret defense computer called Colossus seizing control was published in 1966, while the real one was still classified (no evidence the author knew; the sequencing stands either way). HAL was designed with Marvin Minsky, founder of the MIT AI lab, advising Kubrick: the field consulted on its own nightmare. NORAD's false alarms happened in 1979 and 1980; WarGames came in 1983. DARPA's billion-dollar autonomous-weapons roadmap was announced publicly in October 1983; The Terminator opened in October 1984.

Then fiction fed back into reality, and this part is the strangest because it's the best documented. Reagan watched WarGames at Camp David on June 4, 1983, asked the chairman of the Joint Chiefs whether it could really happen, and got told the problem was much worse than he thought. Fifteen months later he signed NSDD-145, the first national directive on computer security. A movie produced national policy. And 28 years after that, the NSA named its targeting program after Skynet. The signal you're looking for isn't hidden. Sometimes it's literally the program name.

The witnesses were the builders

Here's what's hard to square with the "nobody saw it coming" story. The warnings didn't come from Hollywood. They came from the people building the machines, in every generation, on the record.

Norbert Wiener, who founded cybernetics, 1950. Alan Turing, 1951, in a talk you can read today: once machine thinking started, "it would not take long to outstrip our feeble powers... we should have to expect the machines to take control." I.J. Good, who sat next to Turing at Bletchley breaking Enigma, published the intelligence-explosion argument in 1965. David Parnas walked away from a $1,000-a-day seat on the Star Wars advisory panel in 1985 and published exactly why software controlling autonomous defense could never be made trustworthy, in the field's flagship journals. And in May 2023, the CEOs of the 3 leading AI labs signed a one-sentence statement that mitigating extinction risk from AI should be a global priority, while the man who won the Turing Award for inventing the underlying technique quit Google so he could say it freely.

Bletchley to Turing to Good to Hinton. The chain of custody on this warning is 75 years long and it never once left the building.

Even the money structure is old

If you read part 2 of this series, the Capital Loop, here's the part that reframes it: the loop has a birth certificate. In the 1950s, SAGE, the air-defense computing project that cost roughly 4 times the Manhattan Project, generated 80% of IBM's computer revenues and handed IBM the architecture it resold to industry as the SABRE reservation system. State funds the machine, vendor becomes the industry, technology sells back to the market. In 1999 the CIA turned the pattern into doctrine with In-Q-Tel: the investor and the customer are the same institution, by design, and Palantir's own IPO filing shows the result: government work was 45 to 54% of its revenue across the disclosed periods. In 2001, the private-sector version of the loop, Lucent and Nortel financing their own customers' purchases, collapsed, which is the precedent analysts now cite about today's AI deals. And in 2025 the state closed the circle by taking a roughly 10% equity stake in Intel while a $500B AI infrastructure venture got announced from the White House. To be precise, because precision is the discipline here: the broad pattern is 70 years old; the genuinely new move is the government becoming a shareholder.

The gap, in 4 numbers

82 years since Colossus ran. 75 years since Wiener's warning. 12 years of UN talks on autonomous weapons without a binding treaty. 0 comprehensive federal AI laws in the United States, where a 2023 executive order on AI safety was rescinded 15 months after signing. Meanwhile the international declaration on frontier AI risk, non-binding, was signed at, of all places, Bletchley.

I'm not going to tell you what those numbers mean. The whole method of this series is that I show the ledger and you draw the conclusion. But I'll say what the record rules out: it rules out surprise. Nobody gets to claim this arrived suddenly. The capability is generations old, the deployments are documented, the warnings came from the inventors themselves, and the pattern that used to keep all of it behind clearances has inverted: the frontier now lives in commercial labs, funded by the capital loop from part 2, reaching the next 4 billion users from part 1. The capability moved from classified labs into the open market. The rules that would govern it are still being written, and the record above is what any judgment about that should rest on.

Sources: National Security Archive declassified false-alarm files; European Parliament report A5-0264/2001; Microsoft, Bloomberg, CNBC, Crunchbase and Goldman Sachs data for the modern figures; NIST, NSF, GAO, DoD IG, DARPA program archives; The Intercept's SKYNET reporting; Fred Kaplan's Dark Territory for the WarGames account; the NRC's Funding a Revolution for the early funding record; Palantir's S-1.

The full series essay is on Stay Level.