Written by: Claude AI.
Curator/Editor: Học Trò.
In January 1951 a German mathematician sat in Paris with a pencil and counted the people at a conference called "Computing Machines and Human Thought": 184 French, 40 British, five Americans, one German, himself. In May 2026 a journalist stood on a stage in Mountain View and described a year spent letting artificial intelligence run her life. The Computer History Museum keeps both moments, along with a 1770 chess-playing fraud, a robot squirrel, a talking Furby, an analog neural chip Intel didn't know what to do with, and a drawing of an assistant called OLIVER that was meant to live in the ARPAnet. Its online playlist "Artificial Intelligence at CHM" gathers thirty-three such objects, talks and films. Read one at a time, they are museum labels. Read together, they make a quiet argument: almost every promise and every fear in today's AI news already has an object in the museum. The useful question is not whether AI is new. It is which parts are new.
Visiting the museum. Computer History Museum, 1401 N. Shoreline Blvd., Mountain View, CA 94043 · (650) 810-1010 · computerhistory.org. Hours, tickets and directions: computerhistory.org/visit. When this was written (November 2026), that page listed the museum as open Wednesday–Sunday, 10 a.m.–5 p.m., and closed Monday and Tuesday. Check the page before you go. Many of the objects below are in the main exhibition, Revolution: The First 2000 Years of Computing, whose Artificial Intelligence & Robotics gallery can also be browsed online. Everything in the playlist can be read or watched from home.
0. Two people counting
Alwin Walther was fifty-three in January 1951. He ran the Institute for Practical Mathematics at the Technische Hochschule Darmstadt, and his wartime record was not clean. His institute had done calculations for Wernher von Braun's V-2 rockets, and, according to the museum's curator David C. Brock, Walther was evidently involved in a scheme with the SS to make Jewish prisoners at Sachsenhausen do hand calculations. Whether that scheme ever ran is unclear; Allied bombing destroyed most of Darmstadt's university in September 1944. Six years later Walther took the train to Paris for a six-day international conference on computers and the mind. On the attendee list, he did something only a mathematician would do on a train: he counted. France 184, United Kingdom 40, Netherlands 8, Belgium 6, United States 5, Spain 5, Italy 4, Sweden 4, Switzerland 1, Brazil 1, Germany 1. Next to the last number he wrote "1 Allemande." That was him.
Seventy-five years later, on 19 May 2026, the journalist Joanna Stern sat on the stage of the same museum that now owns Walther's annotated list. She had just published I Am Not a Robot: My Year Using AI to Do (Almost) Everything, and she was talking with Nilay Patel of The Verge about what she learned. One line stands out: for AI to be most useful, people often have to adapt themselves to it, "to essentially become a data set." And what worried her most was not losing jobs to AI but "when we outsource our own thinking."
Between those two people counting, one with a pencil and one with a book, sits the whole history the playlist covers. The playlist is short: thirty-three items, a mix of blog posts about recent public talks, archival films, catalog records, an online exhibition called Chatbots Decoded, and a few physical objects. The museum introduces it with the questions its archivists ask before acquiring anything: Is it rare, unique, or precious? Does it represent an influential product, key change, or instructive failure? Was it created by a seminal inventor, or someone who has been overlooked? Does it fill a gap in knowledge?
Those questions are a good way to read AI too. This essay walks through the collection in nine stops: origins, imagination and fakery, understanding, method, booms, agents, costs, people, and memory. At each stop it asks the same pair of questions: what here is old, and what is genuinely new? The answers differ more than the hype on either side suggests.
A timeline drawn only from the collection
| Year | Object or event in the playlist |
|---|---|
| 1770 | Wolfgang von Kempelen's chess "automaton," the Mechanical Turk, with a human hidden inside |
| 1907 | Tik-Tok, the "Perfect-Talking" copper man in L. Frank Baum's Ozma of Oz |
| 1920 | Karel Čapek's play R.U.R. introduces the word "robot" |
| 1938 | The coin-operated Talkie Horoscope: an astrologer's voice on hidden records |
| 1943 | McCulloch and Pitts publish the first mathematical model of a neural network |
| 1950 | Turing proposes the "imitation game" |
| 1951 | Paris: Les machines à calculer et la pensée humaine |
| 1956 | Dartmouth summer workshop names "artificial intelligence" |
| 1958 | Rosenblatt's Perceptron Mark I |
| 1960 | ADALINE, built by Bernard Widrow and his student Ted Hoff |
| mid-1960s | Weizenbaum's ELIZA |
| 1968 | Licklider and Taylor imagine OLIVER, an assistant living in the network |
| 1969–71 | Stanford AI Lab films: speech commands to a robot arm; a vision system solving Instant Insanity |
| 1987 | Apple's Knowledge Navigator concept video |
| 1989 | Intel's ETANN 80170 analog neural network chip |
| 1990 | General Magic is founded to build a wireless "agent" handheld |
| 1993 | Edward Feigenbaum's lecture Tiger in a Cage |
| 1998 | Furby |
| 2006 | Fei-Fei Li's team begins building ImageNet |
| 2011 | Siri ships inside the iPhone 4S |
| 2012 | AlexNet wins the ImageNet challenge |
| 2017 | Google researchers publish the transformer |
| 2022 | ChatGPT |
| 2024 | Chatbots Decoded opens at CHM |
| 2025 | "The Great Chatbot Debate": Emily Bender vs Sébastien Bubeck |
| 2026 | Joanna Stern's year with AI |
1. Where it began, and which way the arrow pointed
Ask anyone where artificial intelligence was born and you will hear the same answer: Dartmouth College, summer 1956. John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon organized a summer research project, about two dozen people came and went, and the field got its name. The museum's exhibition reproduces the famous photograph of seven of them on the lawn: Oliver Selfridge, Rochester, Ray Solomonoff, Minsky, Peter Milner, McCarthy and Shannon. Its caption sums up the meeting in two sentences that have aged well: "They agreed to call it 'artificial intelligence.' They didn't agree on how to achieve it." The playlist also holds McCarthy himself, interviewed at eighty by his Stanford colleague Nils Nilsson in 2007, explaining how he organized that summer and how his search for a language suited to the new field became LISP.
But the playlist also holds the Paris conference, and it complicates the story.
From Monday 8 to Saturday 13 January 1951, Louis Couffignal, director of the digital computing laboratory of the Institut Blaise Pascal, hosted 259 people a few minutes' walk from his lab in the Latin Quarter. The money came from France's national research agency, the CNRS, with help from the Rockefeller Foundation. The first two days were about machines. Howard Aiken spoke about his Harvard Mark II, III and IV. Andrew Booth described his computers at Birkbeck. Eduard Stiefel of ETH Zurich reported on Konrad Zuse's Z4, then in his care. F. C. Colebrook described the Pilot ACE, built to Alan Turing's design. Freddie Williams reviewed Manchester's work, and Douglas Hartree and Maurice Wilkes spoke about programming Cambridge's EDSAC. Couffignal demonstrated his own "pilot-machine" on the first evening, its world debut.
The last two days were about minds. On Friday morning Gonzalo Torres y Quevedo showed his late father's electromagnetic chess-playing machine from the 1920s, and Norbert Wiener played it. That afternoon the British psychiatrist W. Ross Ashby presented his Homeostat, a machine designed to adapt to its environment, and the neurophysiologist Grey Walter showed his light-seeking robot "tortoises." Wiener spoke about machines giving humans new kinds of perception. Warren McCulloch, who with Walter Pitts had invented the idea of "nervous nets" in 1943, offered the line of the week: "Brains are calculating machines, but man-made calculating machines are not yet brains."
In 2017 the Swiss computer scientist and historian Herbert Bruderer argued in a Communications of the ACM blog post that "this well-documented event could also be regarded as the first major conference on artificial intelligence." Brock, writing for the museum in 2021, is more careful. He says the answer depends on what one means by "artificial intelligence," and that what Paris certainly did was build a community "animated by thinking about minds and machines together." Both readings are fair. Dartmouth gave the field its name and its research program. Paris put working computers and brain-inspired machines in the same rooms, five years earlier, and in Europe.
One reason Dartmouth wins the textbook contest is simple, and the museum's own account points to it. The full Paris proceedings, more than 560 pages published in French by the CNRS in 1953, were not freely available on the web when Brock searched in 2021. History favors what can be read.
The most interesting thing about Paris, though, is not who came first. It is the direction of the arrow. McCulloch wanted to build machines on the model of brains. Couffignal, closing the conference on Saturday morning, argued the opposite. He proposed studying the brain as "a machine where thoughts are elaborated," and comparing its actual working processes to the "ideal logic performed with computing machines." The goal, in Brock's summary, was "not to make computers think like humans, but for humans to think like computers." Couffignal thought this would be good for people. It would break the "few steady ideas" humanity reasoned with and raise its "intellectual potential."
Couffignal's own career did not go well. He had designed a French computer that broke with the American model: large parallel units of complex logic and almost no memory, organized, he said, like "a factory assembly line." The contractor, Logabax, went bankrupt in 1952, and the big machine was never built. His laboratory ended up buying a standard British Elliott 402, the very design he had rejected. In 1959 he was fired.
Yet his idea comes back in the collection's newest item. Stern's worry that people must "become a data set" for AI to work, and Kate Crawford's observation (in a 2021 CHM talk) that too little attention goes to "the way that people are being treated like robots," are Couffignal's dream seen from the other side. He hoped people would reason more like machines and grow. The 2020s version is quieter: people adjusting themselves to machines because that is what makes the machines useful.
The playlist adds two smaller origin stories. One is a language. LISP, which McCarthy designed at MIT in the late 1950s, spread around the world as institutions re-implemented it for their own computers. In East Germany a young philosophy PhD named Herbert Stoyan, working in an AI group at the state electronics firm Robotron, implemented the LISP system used for all AI work in the country, working only from a single book, The Programming Language LISP: Its Operation and Applications. He moved to West Germany in 1981, became a professor of artificial intelligence at Erlangen, and spent decades collecting LISP's history. His collection at the museum fills 160 boxes. A field crossed the Iron Curtain as a printed book, and the man who rebuilt it became its archivist.
The other story is a road the field did not take. Before "artificial intelligence" won the name, the brain side of computing was called cybernetics: feedback, adaptation, small machines responding to light and touch. The museum's robotics gallery keeps its relics. There is Edmund Berkeley's SQUEE, a robot squirrel that hunted tennis-ball "nuts" with two light sensors and two contact switches; it was described as "75% reliable" but worked well only in a very dark room. There is the Johns Hopkins "Beast," which searched for black wall outlets to recharge itself. And there is Grey Walter's Elsie, which sought moderate light and was, the label notes, "peculiarly attracted to women's stockings." At Dartmouth, the exhibition says, "some advocated cybernetics." Symbolic AI won, and cybernetics faded into robotics and control engineering. It never quite died. Rodney Brooks's insect robots at MIT used reflexes "instead of a central 'brain'," and Brooks went on to co-found iRobot, maker of the Roomba. And at a 2025 CHM panel, MIT's Daniela Rus described her company Liquid AI's models as inspired by a worm with very few neurons. By her account, a lane-keeping task that needs about 100,000 neurons in a classical model needs only 19 in hers. That is her claim, not an independent measurement, but its spirit is pure 1951.
Old: the dream of machines that think, and the rival dream of humans who think like machines. New: the second dream is no longer a speech at a banquet. It is how ordinary products are designed.
2. The long imagining, and the long con
Long before anyone could build a thinking machine, people told stories about them, and the museum's Chatbots Decoded exhibition starts there. Talos, the bronze giant of Greek myth, was made by the god Hephaestus. The Golem of Jewish folklore is clay that cannot speak but is brought to life by Hebrew words written on paper, a creature animated by text. Pinocchio, the exhibit points out, "gains the power of speech but, like today's chatbots, has trouble with truthfulness." Baum's Tik-Tok wears a label reading "Thinks, Speaks, Acts, and Does Everything but Live." In Čapek's R.U.R. the manufactured workers revolt, and their makers regret not giving them different languages so they could have been turned against one another. Asimov's I, Robot, the exhibition notes, is largely about conversation as the way to control robots. Astro Boy, the Japanese robot child, was created by a man who had lost his son.
Each story carries one of today's arguments in miniature: deception, control through words, servitude, rebellion, grief. They also explain why talking machines affect us so strongly before they do anything useful.
The next room of the exhibition is less flattering. It is about fakes.
In 1770 the Hungarian inventor Wolfgang von Kempelen unveiled a mechanical figure that seemed to play chess. The "Mechanical Turk" toured Europe and America for eight decades, and many people correctly guessed that a skilled human player was hidden inside. In 1938 the coin-operated Talkie Horoscope spoke fortunes in the voice of a real astrologer, recorded on vinyl records hidden in the cabinet. Chatty Cathy, a best-selling doll of the 1960s, talked through a tiny phonograph; Teddy Ruxpin, in the 1980s, through a cassette. The exhibition's verdict on all of them is the same: "talking was an illusion."
That much is fairground history. The uncomfortable part is that the pattern did not end with the fairground. In 2001 the instant-messenger chatbot SmarterChild became famous for its quick replies; the exhibition notes it was "backed by a team of writers." In her 2021 CHM talk about Atlas of AI, Kate Crawford described an "AI" scheduling assistant whose work was actually done by people paid a few cents per task. She used the writer Astra Taylor's word for this: "fauxtomation." The New York Times reporter Kashmir Hill, interviewing her, described her own investigation of an app called Invisible Girlfriend. And the exhibition's final section, "How Much Does a Free Chatbot Cost?", describes millions of low-paid "taskers" or "clickworkers," some in developing countries and some, it says, in European prisons, labeling images and text "from earlobes to irony," sometimes twelve hours a day.
Why does the trick keep coming back? Crawford's answer is that the appearance of full automation attracts funders and customers, while the reality is "built on repetitive human actions." The Mechanical Turk was a business too. Kempelen's machine made money as a spectacle because people believed, or half-believed, that a machine was thinking.
The museum's playlist also contains a more innocent ancestor of the chatbot: the psychology experiment. In the mid-1960s Joseph Weizenbaum at MIT wrote ELIZA, a program that answered typed sentences using simple rules, borrowing a therapist's habit of turning statements back into questions. The exhibition shows a 1966 printout of a session run by a computing specialist named Emily White. Weizenbaum was alarmed by how readily people poured feeling into conversations with a program they knew did not understand them. At Stanford, the psychiatrist Kenneth Colby built PARRY, a rule-based program that imitated a patient with paranoid schizophrenia; the museum holds a typed transcript of PARRY talking to a nine-year-old boy. At Yale, Robert Abelson built an "Ideology Machine" that modeled the beliefs of a Barry Goldwater supporter; a 1963 flowchart of its reasoning includes boxes for "denial" and "rationalization."
These early programs modeled a therapist, a patient and a political believer. Today's companion apps, therapy bots and worries about political persuasion are not new questions. They are 1960s questions, now asked about systems far more fluent and with far more users.
The web turned these experiments into a public pastime. A.L.I.C.E., a web chatbot that appeared in 2001, used pattern-matching rules much like ELIZA's and, the exhibition says, repeatedly fooled people into thinking they were chatting with a person. Its rules were open-source, so programmers around the world kept improving it. The director Spike Jonze, the exhibition notes, was inspired partly by chatting with A.L.I.C.E. in the early 2000s when he made Her, his 2013 film about a man who falls in love with his voice assistant. A rule-based program that understood nothing helped inspire the decade's best-known story about loving a machine. Weizenbaum would not have been surprised.
Then come the toys. Furby caused a holiday buying frenzy in 1998, with resale prices up to $300. It spoke "Furbish," forty-two words loosely based on German and Japanese, and Furbies talked to each other through infrared sensors between their eyes. A 2005 model, the exhibition notes, woke up to the phrase "Hey, Furby!" years before Siri and Alexa. Sony's AIBO robot dog cost $2,000 and answered more than a hundred voice commands. The 2010s brought "social robots" like Pepper and Jibo, both commercial failures by the early 2020s. The museum also keeps a three-hour oral history with Furby's creator, David Hampton, and two humbler items: a 1970s Pet Rock and a Tamagotchi, the pocket pet that "died" if neglected, sometimes upsetting its owners. The lesson of the toy shelf is that people will attach to almost anything that responds, and many things that don't.
Old: hidden humans behind "automatic" machines; our readiness to bond with anything that talks back. New: the scale. The Turk had one hidden player; today's systems rest on millions of workers, and their companions have hundreds of millions of users.
3. Does it understand?
The exhibition calls this "a big question experts are tackling," and states the two sides plainly. A 2023 Microsoft Research paper, Sparks of Artificial General Intelligence, argued that large language models show real intelligence, reasoning and knowledge. A 2021 paper, On the Dangers of Stochastic Parrots, argued they are closer to parrots, "repeating phrases they've picked up without really understanding them."
In March 2025 the museum, with IEEE Spectrum, put the lead authors on one stage. Emily M. Bender, a University of Washington linguist and co-author of the parrots paper, faced Sébastien Bubeck, lead author of Sparks, formerly of Microsoft Research and by then at OpenAI. The editor Eliza Strickland moderated. The event title asked the question directly: Do LLMs Really Understand?
A debate like this is easy to score as a contest. It is more useful to notice that the two sides use different evidence. Bubeck's side looks at behavior. Give the system hard, unfamiliar problems and see whether it solves them in ways that require something like reasoning. Bender's side looks at how the system was built. A model trained to predict text from text has, by construction, never touched the world the words refer to, so whatever it does, it cannot mean what we mean. Each side's evidence is close to invisible to the other. That is why the argument does not end.
The collection shows how old this split is. McCulloch said in 1951 that calculating machines were "not yet brains." Turing, in 1950, had avoided the question altogether: if a machine could converse so well that people could not tell it from a person, he proposed, it was "reasonable" to say it could think. The exhibition also quotes the Dutch computer scientist Edsger Dijkstra, from around 1985: "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Dijkstra's point is that "think" and "understand" may be the wrong words, and that we keep asking because the machines talk.
The museum gives a room to the early critics, and they read well today. In 1972 the philosopher Hubert Dreyfus argued that much of human thinking is not symbolic, so an AI built entirely on symbols and rules was on the wrong track. In 1976 Weizenbaum, ELIZA's creator, wrote: "Since we do not now have any ways of making computers wise, we ought not now to give computers tasks that demand wisdom." The anthropologist Lucy Suchman argued that decisions take shape in social situations, which computers do not live in.
Sorting these critiques shows what has changed. Dreyfus was largely right about his target. The symbolic AI of his day did hit the wall he described, and the systems that replaced it learn from examples rather than rules, which is roughly what he said thinking requires. In that sense deep learning half vindicated its harshest critic. Weizenbaum's warning is untouched by any of this. Nothing about scale makes a system wise, and the question of which tasks demand wisdom is now asked every day, about hiring, medicine and courts. Suchman's point lives on in a different form: the systems now live in social situations after all, millions of them a day, and what they learn there is not always what anyone intended.
One more strand belongs here because the exhibition places it under conversation: gender. Chatbots have no gender, the exhibition says, "yet we often give them one," usually female, reflecting stereotypes about assistants. Chatbot makers report that more than one in five interactions involve sexual language or content. The 1997 chatbot Jabberwacky was a rare exception, with an explicitly male persona modeled on its developer. The name Siri, the museum's Siri story notes, means "beautiful woman who leads you to victory" in Norwegian. And when Joanna Stern explored AI relationships for her book, she had an AI "boyfriend," whom she did not introduce to her children. Whatever a chatbot does or does not understand, people clearly understand it as someone.
Old: the argument itself, which is at least as old as Turing and McCulloch. New: the systems now pass, every day and for ordinary users, the conversational test that was meant to settle it, and it has not settled anything.
4. Three schools and a pendulum
The exhibition offers a clean map of the field's methods, and it is worth keeping in mind for the rest of the tour. Symbolic AI represents the world as symbols and applies rules and search to them; it dominated from the 1960s into the 1990s and produced chess programs and "expert systems." Neural networks are loosely modeled on brains: layers of connected numbers whose connection weights are adjusted, through training on examples, until the network does a task. Statistical machine learning finds patterns in data and makes predictions from them. It took off in the 1990s, Fei-Fei Li said in her 2024 CHM talk, as a "quiet revolution," and the web then supplied it with endless data. In the 2010s neural networks returned as deep learning, with many layers, huge datasets and graphics chips doing the arithmetic. Since 2017 the transformer design has let language models train on trillions of words. The exhibition says reading that much would take about 50,000 years nonstop.
The pendulum image is the exhibition's own. It ends the section by asking whether the pendulum has finished swinging, and lists researchers who think not: Doug Lenat, who argued for combining symbolic AI with deep learning; Josh Tenenbaum at MIT, modeling how children learn from just a few examples; Jeff Hawkins, co-founder of Palm, whose company Numenta aims to reverse-engineer the cortex. "AI is still evolving," the label says. "Multiple paths remain open."
The hardware tells the same story in metal. Frank Rosenblatt's Perceptron Mark I of 1958 was a neural network built as electronic hardware, trained to recognize letters. His later speech-recognition network was named Tobermory, after a talking cat in a short story by Saki. In 1960 Bernard Widrow, who had been at Dartmouth, built ADALINE with his student Ted Hoff; it could be trained to recognize letters or balance a broom. Hoff, the exhibition notes, went on to play a key role in making the first microprocessor at Intel. The two great lines of computing hardware, the general-purpose processor and the neural network, share an engineer.
Intel itself tried a neural chip three decades before the boom. In 1989 it announced the 80170 ETANN, an "Electrically Trainable Analog Neural Network" with 64 analog neurons and 10,240 analog synapses, built on Intel's flash-memory process. A columnist for PC Magazine, John C. Dvorak, said it had "achieved the theoretical intelligence level of a cockroach" and that "nobody at Intel knows what to make of this thing." There was, the museum's curator Dag Spicer writes, "neither the computing power nor the understanding to deploy them at scale." The project manager, Mark Holler, donated two of the chips to the museum in 2024. The playlist's later chips (Google's tensor processors, IBM's brain-like TrueNorth, Cerebras's wafer-sized processor with 2.6 trillion transistors, Nvidia's H100 at around $30,000 each) are the same idea arriving with enough computing power behind it.
The most vivid evidence of how hard the early problems were is on film. In autumn 1968 the Stanford Artificial Intelligence Laboratory filmed Hear! Here!, showing a computer learning to recognize spoken phrases and then a computer-controlled arm carrying out commands such as "Pick up the small block in the lower left-hand corner." In 1971 another Stanford film, Instant Insanity, showed a vision system and a robot arm solving a century-old puzzle of four colored cubes, which can be arranged in 41,472 ways. The system found each cube by matching its edges to a model cube, turned cubes 45 degrees to see three faces, and read colors by looking through three color filters in turn. The museum also keeps one of the arms from that lab: a hydraulic manipulator built by Victor Scheinman and others between 1966 and 1968 to test whether a robot arm could keep up with the computer controlling it.
Watching those films in 2026 is strange in a specific way. The parts that seemed like the frontier, understanding a spoken sentence and recognizing objects, are now done by a phone. The parts that seemed like plumbing, picking up a block smoothly and handling an unfamiliar object, are still hard for robots. Researchers have long noticed this reversal (it is often called Moravec's paradox): what is hard for people can be easy for machines, and what a child does without thinking can be the hardest of all. The 1966 MIT Summer Vision Project, which Crawford mentions, planned to teach a computer to label any object in the world in one summer. As she says, the challenge is still not fully solved.
Old: every method in use today has a 1950s or 1960s ancestor, and each one has been declared the winner before. New: the hardware finally caught up with the oldest idea on the list, the neural network. That is the main reason this decade looks different from the last time.
5. Booms, busts, and the bars of the cage
In October 2025 the museum invited three founders from three generations of AI companies to answer one question: is today's boom different? The event's title gave a hint: This Time It's Different.
Jerry Kaplan, one of the first people to earn a PhD in AI, co-founded Teknowledge in the 1980s expert-systems boom. His company took expert systems from Stanford and sold them to corporations. The technology, he said, was not good enough to do what it promised, and the business model "was based on incorrect assumptions—just like the AI industry today." Asked whether this time is different, he answered: "It's different just like last time." He expects this bubble to burst, with serious consequences.
Adam Cheyer, co-founder of Siri, was more optimistic. He pointed to AI solving fifty-year-old math problems and advancing protein folding, and called it "the most powerful tool humanity has ever created." Daniela Rus first argued that bringing new ideas into the world matters more than timing, then agreed with the others that timing is critical. On jobs, Rus cited her MIT colleague David Autor's research. The CHM write-up garbles the name as "David Otter" and gives a figure of "more than 60%" of today's work not existing "before 1948." Autor and his co-authors' paper, New Frontiers: The Origins and Content of New Work, 1940–2018 (NBER, 2022), actually finds that "the majority of current employment is in new job specialties introduced after 1940." The point stands even after correction: most of today's work did not exist eighty years ago, and nobody in 1940 predicted it.
Kaplan's era is represented in the playlist by its most important figure. In August 1993 Edward Feigenbaum, the Stanford pioneer of expert systems, gave a lecture called Tiger in a Cage. Expert systems, the lecture description says, had partly realized the old dream of super-human performance, but only "for narrow ranges of human endeavor," and the field had then shifted toward "knowledge systems" that help people rather than replace them. Adoption in industry had been "cautious, slow, and linear (rather than exponential)." "The 'tiger' is in a cage," the description concludes, "and Feigenbaum sets the challenge of understanding the bars."
That phrase is the most useful thing in the playlist for anyone trying to judge the current boom. The tiger was capability; the bars were everything else: cost, integration with old systems, trust, liability, the people whose jobs had to change, the knowledge that had to be extracted and maintained. In 1993 the bars, not the tiger, set the pace. Today's tiger is far stronger. Whether the bars are weaker is a separate question, and it is the one investors are betting on.
Two older ideas from the exhibition help explain why AI history moves in waves. The first is the "AI effect," summed up in a line the museum attributes to Larry Tesler around 1970: "Intelligence is whatever machines haven't done yet." Once grammar checking, handwriting recognition or chess-playing works, it stops being called AI. Expert systems, the exhibition notes, did not vanish after the bust. They became routine, "powering every automated phone system you call." The second is the cycle itself. Gartner's 1995 "hype cycle" describes the rise from excitement to disappointment to quiet usefulness, and in AI the disappointments have had a name: winters, when funding dried up. Stoyan's LISP collection records one of them, noting that the AI winter at the end of the Cold War hit demand for AI applications and interest in LISP.
The museum had already framed the question of fear this way in February 2021, almost two years before ChatGPT, in a virtual town hall called Should We Fear AI? Its announcement set the extremes side by side. On one side, drones hunting people down from their social-media profiles. On the other, a personal robot that brings tea, picks up the children's socks and helps brainstorm the end of a novel. Then it pointed out how old both feelings are. "Smarter machines have been putting people out of work since the printing press," it said, yet automation also freed most people from farm labor, turned "doctoral-level math into routine calculation," and gave us dishwashers that "would have been fairy tales to our ancestors." Its central question is the right one for any boom: "Are we even afraid of the right things?"
The honest conclusion is a split one. Kaplan may be right that the money is ahead of the business. Cheyer may be right that the tool is real. Both happened last time too: the expert-systems companies failed, and the technology lived on in every phone menu.
Old: the cycle of hype, bust and quiet absorption, and the gap between what a technology can do and how fast organizations can use it. New: the tiger. Whatever happens to the companies, the capability is not going back in the cage.
6. The unfinished "do engine"
In April 1968 J. C. R. Licklider and Robert W. Taylor, the two men most responsible for funding the network that became the ARPAnet, published an essay called "The Computer as a Communication Device." In it they imagined an assistant named OLIVER, after Oliver Selfridge, the pioneer who had suggested the idea and had helped organize Dartmouth. OLIVER would be "a complex of computer programs and data that resides within the network and acts on behalf of its principal, taking care of many minor matters that do not require his personal attention." They anticipated the obvious objection. "'You are describing a secretary,' you will say. But no! Secretaries will have OLIVERS."
Licklider was in no hurry. Weber notes that he believed human-level AI was possible but thought it might take "between 20 and 500 years," and he expected the period of human-computer partnership before then to be, in his words, "intellectually the most creative and exciting in the history of mankind." Taylor went on to help bring the personal computer into being at Xerox PARC, where the Alto was built. Both men are now remembered mainly as champions of augmenting human intelligence rather than replacing it, and OLIVER fits that tradition. It was an assistant, not a substitute.
The museum's curator Marc Weber tells the story of what happened to that idea in a long piece called "Secrets of Siri," built around an oral history with Siri's founders that the museum recorded in 2013 and agreed not to release until they gave permission. His timeline has four stops.
The first is OLIVER itself. The second is 1987, when Apple's CEO John Sculley, needing a vision of his own after pushing out Steve Jobs, had Apple's Advanced Technology Group make a concept video called Knowledge Navigator. A professor uses a tablet with a touch screen, and a bow-tied animated assistant manages his calls, finds a colleague by video and searches something like a future web. The video is set in 2011. The third is 1990, when a group from Apple left to found General Magic, whose handheld operating system, Magic Cap, was built around software "agents" that could go off and order lunch or find a cheap flight over a wireless network. Its leader, Marc Porat, described a world in which people would be connected all the time to a "cloud" of services through devices in their pockets. The vision was close to today's smartphone. The hardware was not: Magic Cap devices from Sony and Motorola reached the market in 1994 and sold poorly. (The museum's caption on a Magic Cap screenshot dates it to 1984, which cannot be right for a company founded in 1990.)
The fourth stop is Siri. It came out of years of government- and industry-funded research at SRI International, where Adam Cheyer, chief scientist Didier Guzzoni and others had worked on getting many AI components to cooperate. Steve Jobs personally courted the founders and bought the small company in 2010 for an estimated $200 million. Apple announced the iPhone 4S, with Siri built in, on 4 October 2011. Jobs died the next day. Tom Gruber remembers Jobs's pitch: "you won't have to tap and type all the time anymore." In the 1987 video the assistant appears in 2011. Siri arrived in 2011.
What the founders had built was more ambitious than what Apple shipped. At the 2025 CHM panel Cheyer explained that they never pitched Siri as AI. They called it a "do engine," to distinguish it from a search engine: something that would combine knowledge and action, book the taxi, buy the tickets. Before Apple, Siri's business model fit that vision. Instead of selling ads that might lead someone to book a hotel, Siri would take a commission on bookings it actually completed. Apple built only part of this. Cheyer, Dag Kittlaus and Chris Brigham then founded Viv Labs to try again, sold it to Samsung in 2016, and parts of it live on in Samsung's Bixby.
Weber's conclusion is the sharpest sentence in the playlist about the current agent boom. Generative AI has made the components "smarter and much, much more fluent," but it is "only beginning to address the central systems integration challenge": how to coordinate many services into one task, start to finish. His example is deliberately ordinary. An agent that finds a good Pad Thai restaurant but doesn't check with a mapping service for travel time delivers dinner late. Gruber, in the oral history, compares building such a system to assembling a 747. Siri's team built a framework they called Active Ontologies for exactly this problem, so that separate AI modules could be defined and orchestrated within one system.
Anyone who has managed a large engineering project will recognize the problem. The difficult part is rarely any single component. It is the interfaces, the handoffs and the order of operations, which is where Weber says agentic AI still stands. Chatbots Decoded shows how many attempts there have been along the way: Microsoft's Clippy in 1997, which guessed when users needed help and was found "annoyingly intrusive" by many; SmarterChild in 2001; Amazon's Echo in 2014, with over half a billion Alexa devices sold; the Humane AI Pin of 2024, a screenless wearable that the exhibition presented as the latest step. Each made conversation easier. None made the "do" part reliable.
The money question has also not changed. The exhibition's "Follow the Money" section lists the three ways information technology is usually paid for: advertising, subscriptions and pay-per-use. It asks what happens to advertising when a chatbot answers directly instead of showing a list of links. Siri's founders had already given one answer in 2010: get paid for the completed transaction. Each model produces a different kind of agent. One paid by advertisers will persuade. One paid by commission will close deals. One paid by subscription will, perhaps, work for you.
Old: the dream, stated almost word for word in 1968, and the business question of who pays the assistant. New: the assistant can now talk. Whether it can reliably do is the open question, just as it was for Siri.
7. The bill: data, bodies, earth
Every technology has costs that do not appear on the box. The playlist spends more time on AI's costs than on any other theme. The museum ran public talks on bias (2021), on the material footprint of AI (2021), on journalism (2025) and on education (2025), and the Chatbots Decoded exhibition ends with a section asking "Should You Worry?"
The data. The playlist tells the story of one dataset twice. In Fei-Fei Li's 2024 talk, ImageNet is a story of persistence. As a young researcher she bet on assembling a massive labeled image collection when most of her senior colleagues worked with small data, and the open dataset, begun at Stanford in 2006, "jumpstarted" modern AI research. When a deep network called AlexNet, from Geoffrey Hinton's lab in Toronto, won the ImageNet challenge in 2012, it launched the deep-learning era. The museum even has a photograph of the home computer with graphics cards on which AlexNet was trained.
In Kate Crawford's 2021 talk, the same dataset is a warning. With the artist Trevor Paglen she spent two years opening up training datasets, a project called Excavating AI. In ImageNet's categories for people she found labels that were racist, misogynistic, offensive or simply illogical, and that had been used to train systems for over a decade. Many of those categories were later removed; the CHM account puts the number at 660,000 images. Li's own group published a paper on filtering and rebalancing ImageNet's "person" categories in 2019–20 (Yang, Qinami, Fei-Fei, Deng and Russakovsky, Towards Fairer Datasets). Crawford raises a further problem: systems already trained on the removed data are still running, but the evidence of what they learned from is gone. "When the original data is no longer viewable," the museum's summary asks, "how can we investigate the causes of these outcomes?"
Both stories are true, and the dataset needs both. Engineers, Crawford says, tend to treat datasets as "tools in a toolbox," applied without looking inside. Some of the data was gathered without consent. She cites mugshot collections used by the U.S. standards agency NIST, and a face-recognition company that scraped at least three billion images from the web.
The bias. The museum's 2021 event "Is AI Racist?" produced the playlist's most self-incriminating exhibit. When the museum ran automatic transcription on its own recording, the software rendered the name of the MIT researcher Joy Buolamwini as "joy blend weenie." Dr. Vilna Bashi Treitler became "Dr. Vilna bossy trailers," and Miriam Sweeney became "Miriam, sweetie." It had no trouble with White and Asian male names, "even 'Satya'." (The museum's write-up itself misspells Buolamwini's name.) The panelists then described failures with heavier consequences. Black men were arrested after being misidentified by police facial-recognition systems. An algorithm for organ transplants disproportionately rejected people of color. One startup's system flagged content uploaded by people of color as "unsafe" and removed it. That experience led the computer scientist Deborah Raji to the field of AI fairness, and to Buolamwini. Lili Cheng of Microsoft recalled Tay, a chatbot that users on Twitter taught to be racist and sexist within hours; the exhibition says it was pulled after sixteen hours. NYU's Charlton McIlwain, who has long studied the history of racism in technology, traced police-and-computer partnerships back to systems built in the late 1960s to deal with what was then called "the race problem" in poor urban areas.
Their conclusions were not technical. "There will not be an 'algorithmic fix'" to racism or the wealth gap, said the UCLA scholar Safiya Noble, and McIlwain agreed: "Fixing a line of code or a new app is not a fix." Noble's proposal was organizational: give social scientists, historians and humanists as much power over a product as its engineers. Raji's was legal: create real accountability for companies whose systems harm groups of people, and stop dismissing the marginalized people who usually see the harms first.
A 2017 talk in the playlist shows how close this sits to commercial products. Rana el Kaliouby's company Affectiva, spun out of the MIT Media Lab, had built software to read emotions from faces, trained on 4.8 million faces from 75 countries. Its customers were consumer brands measuring how audiences reacted to ads. Set beside the transcript that could not hear a Black woman's name, it raises an obvious question: whose faces does a system read well, and who decides what a face means?
The earth. Crawford's Atlas of AI begins with one object, an Amazon Echo, and traces its whole life. Its minerals are mined, smelted, assembled and shipped by workers. While it sits in a living room, it collects data used to train AI. When it is thrown away, it ends up in an e-waste dump, she says, in a place like Ghana or Pakistan. "AI is not artificial," the museum's summary puts it; it depends on materials, energy and people. Crawford warned that at current rates the world could run short of lithium as early as 2040. That is her projection, and it depends heavily on assumptions about mining and recycling. She and Alex Campolo named the contradiction "enchanted determinism": AI is marketed as almost magical, yet trusted to make decisions in criminal justice, education and hiring. The exhibition adds the data centers, which consume "more electricity than a small country," with cooling towers that can use millions of gallons of water a day.
The news. At a 2025 panel on journalism, Jason Koebler of 404 Media described spending months on an investigation into illegal material in an AI training dataset. Within minutes of publication it was scraped and republished by dozens of AI content farms, and those copies then ranked above the original in Google search. He called the result a collapse of discoverability: it is now so cheap to produce plausible content that it crowds out the people doing original work. A live poll of the audience produced the words "Scary," "Caution," "Untrustworthy," "Propaganda" and "Fake." The panel's advice was old-fashioned and still correct: read outside your echo chamber, pay for news you trust, and check where a story came from.
The image. The exhibition's section on generative AI makes the trust problem concrete with a small gallery of pictures made by the image generator Midjourney. One prompt asked for "a black and white photographic portrait from 1971 of the first woman on the Moon." The result looks like an archival photograph of an event that, as the label says, has never happened. Others render Alan Turing in the styles of Andy Warhol, Frida Kahlo and Kehinde Wiley, which shows at a glance why artists worry about their styles being absorbed. Next to these the museum places the opposite case: an aerial photograph of Golden Gate Park posted on Reddit, so beautiful that many users assumed it was AI-generated. The fact-checking site Snopes concluded it was real, and the photographer, Eric Thurber, confirmed it. The exhibition puts the consequence as a question: "What will happen if people feel that they can no longer believe what they read, see, or hear?" The fake that fools people is half the problem. The other half is the real photograph nobody believes.
The classroom. Sal Khan, at a 2025 talk on his book Brave New Words, described Khanmigo, his organization's AI tutor. About 450,000 students and teachers were piloting it, with 1.1 million expected the next school year. Teachers, against the stereotype of a slow sector, were the fastest adopters. The surprising number was another one: 80 to 85 percent of students, Khan said, did not know how to ask the tutor questions, just as they hesitate to raise their hands in class. Khan argued that educators are "in a little bit of denial" if they believe they can detect AI-written work, and that schools should instead ask what an assignment is for. The exhibition places this in a longer arms race: calculators in the 1970s, Wikipedia in the 2000s, now AI detectors against tools that rewrite text to seem more "human." The 80–85 percent figure is the real story. The skill that AI rewards most, asking a good question, is one that classrooms rarely teach.
The companions. The exhibition's section "Your New Digital Friend?" describes apps such as Replika and Character.ai, and Microsoft's Xiaoice persona in Asia, with hundreds of millions of users. Most affecting is James Vlahos, who, when his father died, built a chatbot based on him to keep their connection going, and then co-founded a company, HereAfter AI, that makes "versonas" of specific people. The exhibition asks whether such companions reflect our growing isolation "or contribute to it," and leaves the question open. One visitor's card on the museum's talk-back wall sees companionship for isolated people as AI's biggest near-term good. Therapy bots raise the same question with higher stakes: ELIZA in the 1960s, PARRY in the 1970s, Woebot in 2017, each promising help to people who cannot reach a human therapist, each raising questions of privacy and responsibility.
The work. "Will AI steal my job?" is the exhibition's plainest question, and it answers with history rather than prediction. The Luddites smashed the machines taking their skilled weaving jobs. Gutenberg's movable-type press, built from a wine press around 1450, let a pair of printers produce 3,600 pages a day. The scribe, a respected profession for nearly five thousand years, the label notes, "was among the first brainy jobs transformed by automation," and many European scribes moved into clerical work. In 1839 the painter Paul Delaroche, on seeing Daguerre's photographs, supposedly declared "Painting is dead." Painting changed and survived; the exhibition asks whether AI will do the same to art, or whether artists will become "more like patrons, commissioning rather than creating work."
The collection's most surprising object here is from 1960. Two MIT students, Doug Ross and Harrison "Dit" Morse, programmed the TX-0 computer to write the script for a short Western called Saga, made for a CBS television program about "The Thinking Machine." It was AI text generation sixty years before the phrase "generative AI." The newest objects are a Writers Guild of America strike sign from 2023. The strikers settled that September with protections against AI-written scripts, after a court ruled that works made by AI alone cannot be copyrighted. Programming gets its own panel, "The End of Programmers or Just Another Tool?" It traces the line from hand-coding in ones and zeros to high-level languages to AI "pair programmers" that suggest code as people type. They boost speed, stoke fears about job security, and "can also introduce errors that humans must debug." The writer Karl Sharro gets the section's sharpest line: "Humans doing the hard jobs on minimum wage while the robots write poetry and paint pictures is not the future I wanted."
The labor argument in the playlist is therefore not one argument but two. One is about replacement: which jobs disappear. The other, raised by Crawford and by the clickworker section, is about the jobs AI creates underneath itself, and whether those jobs are good ones. Daniela Rus's optimism and Karl Sharro's joke are both about the second question.
The rules. The exhibition was written in 2024, and its regulation section reads like a snapshot. It shows the 2023 letter calling for a pause in AI research, signed by Elon Musk, Steve Wozniak and some twenty thousand others, with a dry note: "However heartfelt, such warnings are also excellent marketing." It shows the U.S. Blueprint for an AI Bill of Rights, published a month before ChatGPT, Senate hearings, a presidential executive order, and the New York Times lawsuit against OpenAI and Microsoft over the use of its articles. It recalls the 2023 Writers Guild strike, which won protections against AI-generated scripts. And it offers a historical parallel. Section 230 of the Communications Decency Act of 1996 (the exhibition mistakenly calls it the "Telecommunications Decency Act") shielded online platforms from liability for what users posted, and that one rule shaped the entire social-media era. "Different rules," the label says, "would have brought different results." That sentence applies to AI now.
Old: almost every cost on this list. Biased records, hidden labor, mining, cheating panics and lonely people talking to machines all predate the technology. New: the speed and the scale, and the fact that the costs are now carried inside systems that people trust to be neutral.
8. The people in the room
The playlist is full of objects, but its recent talks show something objects cannot: the people building AI, describing what they think they are doing.
In April 2023 the museum marked the fiftieth anniversary of the Xerox Alto, the 1973 machine from Xerox's Palo Alto Research Center that showed the world the personal computer: mouse, windows, networking. Two of its designers, Butler Lampson and Charles Simonyi, and Alan Kay on video, remembered a lab culture built by Bob Taylor that got "lone wolves" to collaborate. They remembered how memory-starved the machine was (under 128K), so short that text would disappear from the bottom of the screen when the machine needed memory to format or print. Lampson argued that advances in computing, deep learning included, are limited mainly by hardware, especially memory. Kay argued that when research produces powerful tools, training and education must be part of the system, so that the tools are "used for good."
The second panel that evening was about AI. Ilya Sutskever, then OpenAI's chief scientist, said that all of computing now has one goal: "to build artificial intelligence that actually loves humanity." Eric Horvitz, Microsoft's chief scientific officer, said AI would affect not only jobs and education but "our sense of our uniqueness and identity as human beings." Both were frank that the disruption would be large. Sutskever said developers could build many layers of guardrails, such as refusing unacceptable requests, but that society also needed "wise, well-chosen" rules. Horvitz expected creative solutions and new problems to arrive at the same time. Sutskever's summary was "It will be not boring."
Fei-Fei Li, at her 2024 talk, was the most careful voice in the playlist. She avoids hype, she said, and she called artificial general intelligence a "North Star" that she does not think can be reached for "the holistic definition of human." Asked about AI's biggest risk, she gave a one-word answer: ignorance. Anyone who says AI is all good is "ignorant of the past," because every tool in history has been used to cause harm. But talking only about existential catastrophe ignores that AI is "a physical system" in data centers, "tethered to human society," and therefore something people can govern. Her last point was about public money: the best outcome of public investment in research is "not the technology that it helps to develop but the people."
The investor Reid Hoffman, co-founder of LinkedIn and of Inflection AI, came to the museum in October 2023 with the most literal version of human-machine partnership in the playlist: a book, Impromptu: Amplifying Our Humanity Through AI, co-written with GPT-4. The event's own framing put the choice plainly: is AI "a threat to humanity or a partner with the power to unlock our full potential?" Hoffman's title gives his answer. It also echoes an older idea. Licklider, Taylor and Doug Engelbart championed computers that augment human intelligence, and at PARC, Alan Kay recalled, the goal was computers as "personal intellectual amplifiers."
And then there is Joanna Stern, who used AI for nearly everything for a year and came back with boundaries. Personal communication should not be handed over; AI-written emails still read as generic, and real connection needed her own voice. An AI toothbrush was "not ready for prime time." She used AI to edit, analyze interview transcripts and set up the business behind her book, but she wrote the book herself. She thinks the smartphone, "the pinnacle computer," will not be replaced soon, and that today's models can already do enough. She also reported that students had begun booing AI advocates at graduation ceremonies. A year earlier, her own commencement speech about AI had drawn only yawns.
Old: the people building powerful tools have always argued about what the tools are for, and the best of them (Kay at PARC, Li at Stanford) insisted that education be part of the tool. New: the builders now say openly, on museum stages, that they are making something that could change what it means to be human.
9. How the past is kept
The final stop is the museum itself, and the strange paths objects take to reach it.
After the 1951 conference, Alwin Walther went home to Darmstadt. He rebuilt his institute, finished an electronic computer, the DERA, in 1957, and built up a library that became a key resource for German computing. For that library he had his Paris papers bound together: the four-page program, the three-page announcement, the nine-page attendee list with his pencil counts, and more than a hundred pages of abstracts in English and French. The volume sat on the shelf as item B8807 for years. Then the library discarded it. The rare-book collector Jeremy Norman bought it and used it in his 2002 documentary history The Origins of Cyberspace, and later donated it to the Computer History Museum. It now sits in the museum's archive in the San Francisco Bay Area, and, as Brock notes with some pleasure, a scan of it sits "in a Microsoft computer somewhere in the Azure infrastructure," downloadable by anyone. Wartime Darmstadt, a discarded library volume, a collector, a museum and a cloud server: that is the route by which we can still see who came to Paris.
Software has an even harder time surviving. The museum's Software Preservation Group began with a 2002 email from the software engineer Grady Booch titled "Preserving classic software products." Its first pilot project, started in 2004, tracked down the original FORTRAN compiler and succeeded within a year. Other projects followed, among them LISP, APL and Doug Engelbart's NLS system. Herbert Stoyan's 160 boxes of LISP history are part of the same effort: manuals, reports, source listings, tapes, letters.
The museum's acquisition questions include one that deserves more attention: does an object represent an "instructive failure"? The AI collection is full of them. There is Couffignal's machine that was never built and the ETANN chip that came twenty-five years early. There is General Magic's handheld that foresaw the smartphone, and Apple's Newton, which grew out of the Knowledge Navigator vision. There is Clippy, and Tay, and the 2005 "Emoto-Tronic" Furby that failed to revive the craze. There are Pepper and Jibo, and an educational robot whose vacuum attachment "sadly proved impractical." Line them up and the causes repeat: the hardware was not ready, the timing was wrong, users did not trust it, the price was too high. Cheyer and Kaplan, from opposite sides of the optimism question, agreed on one thing in 2025: timing makes or breaks a technology. A museum that keeps failures makes that lesson visible.
There is also a gap the museum itself admits. The 1953 CNRS proceedings of the Paris conference, the full record of what was said, were not freely online when Brock looked. What gets scanned gets remembered, and what gets remembered becomes the official origin story. That is one reason a five-year-earlier meeting with working computers and brain-like machines lost the title of AI's birthplace to a summer workshop at Dartmouth.
Old: the problem of deciding what to keep. New: almost nothing. The museum's questions (rare? influential? an instructive failure? someone overlooked?) work as well for 2026 as for 1951. AI is producing artifacts faster than anyone can judge them, which makes the questions more useful, not less.
10. Which way the arrow points
The museum's Chatbots Decoded ends with a wall where visitors answer a question: How will AI change the way we work and live in the next decade? The cards the museum chose to show range from environmental worry to computer games, from fear about graphic artists' jobs to hope for lonely people. Above them the exhibition offers two reminders about predicting inventions. Bubble wrap was invented as wallpaper and insulation. The Segway was predicted to be "bigger than the internet." An invention's impact, the label says, "ultimately depends on how regular people end up using it."
Thirty-three objects cannot predict the next decade either. What they can do is sort the present into the parts that repeat and the parts that do not.
Much repeats. The dream of a thinking machine is as old as Talos. Hiding people inside "automatic" systems goes back to the Mechanical Turk. Our readiness to love what talks back goes back to the first talking doll. The argument over understanding goes back to Turing and McCulloch, the boom-and-bust cycle to the expert-systems era, the assistant to OLIVER in 1968, and the bias problem to the records the machines learned from. Even the wish that humans would learn to think like computers has a date and a place: Saturday morning, 13 January 1951, a few minutes' walk from the Panthéon.
Some things are new. The systems really do converse now, and fluently. The hardware finally matches the oldest idea in the field. Costs that once stayed local (a rigged automaton, a biased file, a lonely user) now move at the scale of billions. And a tool that was supposed to adapt to people increasingly asks people to adapt to it.
That last point brings the essay back to the two people counting. Couffignal hoped people would think more like computers, and become more capable for it. Joanna Stern found, seventy-five years later, that making AI useful often meant making herself into data, and she drew lines around what she would not hand over. Neither of them was naive. The difference between them is the question this decade will answer: whether adapting to machines makes us larger, as Couffignal hoped, or smaller, as Stern feared.
Asked at the end of her talk what young people should do, Stern gave advice that would have fit on Walther's attendee list: keep reading, and keep thinking critically. A museum is a good place to do both.
Sources
Every link below was opened and checked on 18 November 2026. All are from the Computer History Museum's playlist Artificial Intelligence at CHM except those marked outside source.
CHM articles and exhibitions (read in full)
- David C. Brock, Thinking about Machines and Thinking (CHM blog, 28 Jan 2021). The Paris 1951 conference, Couffignal and Walther.
- Les machines à calculer et la pensée humaine, 1951 (CHM catalog record for Walther's bound volume).
- Chatbots Decoded: Exploring AI (CHM online exhibition, opened Nov 2024).
- Marc Weber, Secrets of Siri: An Unfinished Revolution in Agentic AI (CHM story).
- CHM Editorial, A Year With AI (22 May 2026). Joanna Stern and Nilay Patel.
- CHM Editorial, Is Today's AI Boom Different? (17 Oct 2025). Jerry Kaplan, Adam Cheyer, Daniela Rus.
- CHM Editorial, New News Tech (25 Apr 2025). Journalism panel.
- CHM Editorial, Sal Khan on AI + Education (18 Mar 2025).
- CHM Editorial, Fei-Fei Li Talks AI (23 Sep 2024).
- Dag Spicer, Neural Network Chip Joins the Collection (9 Aug 2024). Intel ETANN.
- CHM Editorial, From Alto to AI (4 May 2023).
- CHM Editorial, Decoding Racism in Technology (5 May 2021).
- Emily Parsons, Mining More Than Data (15 Jun 2021). Kate Crawford, Atlas of AI.
- Revolution exhibition: Artificial Intelligence & Robotics, AI in Action: Robotics (SQUEE, the Beast, Elsie), The Robots Around Us, Furby, Stanford hydraulic arm.
- Herbert Stoyan collection on LISP programming (CHM catalog) and the Software Preservation Group's History of LISP.
- Oral History of David Hampton (CHM catalog record, Furby's creator).
CHM videos (official descriptions read; the videos themselves were not used for any claim beyond their descriptions)
- The Great Chatbot Debate: Do LLMs Really Understand? (25 Mar 2025). Emily M. Bender, Sébastien Bubeck, moderator Eliza Strickland.
- AI + Humanity: A Conversation with Reid Hoffman (30 Oct 2023).
- Should We Fear AI? A Virtual Town Hall (19 Feb 2021).
- Oral History of John McCarthy (interviewed by Nils Nilsson, 2007).
- Hear! Here! Computer Speech Recognition (Stanford AI Lab, recorded 1968).
- Instant Insanity: Computer Vision & Robotics (Stanford AI Lab, 1971).
- Tiger in a Cage: Applications of Knowledge-Based Systems (Edward Feigenbaum, 1993).
- Engineering Emotional Intelligence (Rana el Kaliouby, 2017).
Outside sources (used to check or correct the CHM pages)
- Herbert Bruderer, The Birthplace of Artificial Intelligence?, BLOG@CACM, 3 Nov 2017.
- David Autor, Caroline Chin, Anna Salomons and Bryan Seegmiller, New Frontiers: The Origins and Content of New Work, 1940–2018, NBER Working Paper 30389 (2022).
- Kaiyu Yang, Klint Qinami, Li Fei-Fei, Jia Deng and Olga Russakovsky, Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy (FAT* 2020).
- Wikipedia, for dates only: iPhone 4S (announced 4 October 2011); Steve Jobs (died 5 October 2011); Magic Cap (Sony Magic Link and Motorola Envoy, 1994); Section 230 (Communications Decency Act of 1996).
A note on corrections. Museum pages read like settled fact, but a few of these contain slips, and the essay corrects them rather than repeating them: the economist's name is David Autor, not "Otter," and his paper's baseline year is 1940; Joy Buolamwini's name is misspelled on one page; the 1996 law is the Communications Decency Act; and Magic Cap could not have existed in 1984. None of these changes the substance of what the museum is showing.
Written by: Claude AI.
Curator/Editor: Học Trò.




No comments:
Post a Comment