Jaron Lanier: Stop calling it AI—it's collaborative software

Jaron Lanier opens by rejecting the term "artificial intelligence" as misleading and potentially dangerous, especially at a moment when the field is shaping the technological future. He notes that the term was coined in the 1950s and has been reinforced by decades of science fiction—films like "The Terminator" and "The Matrix," and characters like Commander Data from "Star Trek: The Next Generation." These cultural narratives have become almost religious mythology within tech culture, driving researchers to pursue the dream of creating true AI.

Yet Lanier identifies a stark paradox: many of the same scientists building these systems express existential fear about them. In a recent poll, half of AI scientists agreed there was at least a ten-percent chance that the human race would be destroyed by AI. Even Sam Altman, who runs OpenAI, has made similar warnings. This creates a strange debate in tech circles: some argue the new code is just code and humans remain in charge; others insist anyone who thinks this simply doesn't understand how profound the technology is. When Lanier asks his most fearful colleagues to spell out how an AI apocalypse might actually happen, they often fail to articulate a coherent mechanism, resorting instead to vague claims like "Accelerating progress will fly past us and we will not be able to conceive of what is happening."

Lanier's own position is pragmatic: think of these systems as tools, not creatures. This doesn't eliminate the possibility of peril—badly designed or operated technology can still hurt us—but it rejects the mythologizing that clouds judgment. To understand what we are actually building, Lanier proposes a different framing: innovative forms of social collaboration. GPT-4, which writes sentences to order, is like a version of Wikipedia with vastly more data, combined using statistics. Image-generation programs are like advanced online image search, but with a system for combining pictures. In both cases, real people wrote the text and furnished the images; the new programs mash up human work using statistical guidance. This is a significant achievement, Lanier argues, but it illuminates hidden concordances between human creations rather than inventing a new mind.

The flexibility and unpredictability of these systems makes them feel intelligent, but Lanier explains this arises from simple mathematics. A large language model contains a cumulative record of how words coincide in vast amounts of text. This gargantuan tabulation causes the system to approximate grammar patterns and authorial style. When a user enters a query, it is correlated with what's in the model; results vary each time because of the complexity of correlating billions of entries. The non-repeating nature of this process can feel lively and can actually be human-centered—users get multiple options and must choose among them, introducing a moment of human agency into the process.

Lanier sees upside in this flexibility. Digital systems have historically been brittle, forcing people to conform to rigid designs rather than assess them—he cites the image of someone struggling at a doctor's office front-desk screen, their humanity undermined by the need to comply with the machine. AI tools could invert this: imagine a website that reformulates itself on the fly for someone who is color-blind, or that tailors itself to someone's particular cognitive abilities and styles. For a humanist like Lanier, this offers the possibility of more human control and agency.

But Lanier also acknowledges legitimate worries that the technology will push people around in ways they don't like or understand. He notes that some colleagues circulated a petition asking for a pause on ambitious AI development to allow time for policy work, though he found the notion too hazy—what level of progress would allow the pause to end? He also voices skepticism about existing AI-policy efforts, drawing on his years working on EU privacy policy. The problem, he argues, is that we don't know what privacy is well enough to generalize, and the same is true of AI concepts like "alignment" (matching an AI's goals to human goals), "safety" (foreseeing guardrails), and "fairness" (preventing bias). These efforts do good, but the vagueness of the terms turns policy into a game of trying to outwit a genie, leaving the AI research community communicating dire warnings while proposing ever more urgent but undefined deliberative processes.

In an informal experiment, Lanier called colleagues to see what specific points they could agree on. He found surprising consensus. Nearly everyone agrees that deepfakes and automated manipulative communications should be labeled as such by the programs creating them, with actions people can take. There is also near-unanimity that the black-box nature of current AI tools must end and the systems must be made more transparent. The challenge is that large-model systems aren't made of explicit ideas; there is no definite representation of what the system "wants" or when it is manipulating a person. Instead, there is a vast mathematical mixing.

Lanier sketches one obstacle: attempts to open the black box by forcing the system to produce otherwise unnecessary items like scripts or sketches would involve building another black box to interpret the first—an infinite regress. But he argues it's not necessary. The interior of a big model doesn't have to remain a trackless wilderness. Instead, the way to open the black box is to reveal the people. A real person created an illustration that was input as data, and in combination with contributions from others, it was transformed into a fresh image. Big-model AI is made of people—and transparency means revealing them.

This concept is called "data dignity," which Lanier has contributed to developing. It emerged as an alternative to the familiar arrangement where people give their data for free in exchange for free services. That arrangement, he notes, has had a dark side: network effects allow a few platforms to dominate, eliminating smaller players like local newspapers. Since the online experience is supposed to be free, the only remaining business model is hawking influence. Data dignity offers a different path: acknowledging and potentially compensating the real people whose creations were transformed into new outputs by these systems.

Lanier's essay contests a foundational myth in tech culture: that computer scientists are building genuinely intelligent entities. He traces the AI framing back to the 1950s coining of the term and to cultural touchstones like "The Terminator," "The Matrix," and "Star Trek," which have shaped researchers' aspirations and anxieties alike. The paradox he identifies is striking—many of the scientists building these systems simultaneously warn that they could annihilate humanity—yet when pressed for specifics on how an AI apocalypse might occur, their reasoning breaks down into vagueness. Lanier argues this contradiction arises because the underlying framing is wrong: these are not independent minds but statistical systems that remix human work.

The practical stakes of this reframing are high. Lanier contends that mythologizing technology makes it more likely we'll manage it poorly, because mystery paralyzes clear thinking. By contrast, understanding these systems as collaborative tools lets us focus on concrete, actionable safeguards—labeling synthetic media, making systems transparent, and crediting the people whose creations powered them. His frustration with existing policy efforts (alignment, safety, fairness) is that they rest on undefined terms and generate only vague deliberative processes, not real constraints. Data dignity, by contrast, offers a material anchor: the identifiable humans whose work is embedded in every output.