Jaron Lanier says 'AI' is the wrong story

Jaron Lanier argues that the term “A.I.” is not just imprecise but actively misleading. In his April 20, 2023 essay for The New Yorker, the computer scientist says mythologizing systems like GPT-4 makes them harder to govern well, especially as parts of the industry talk openly about existential risk.

×

Lanier’s core claim is straightforward: today’s systems should be treated as tools, not creatures. He says large models are not independent minds but statistical systems built from human-created text, images, and other material. A model such as OpenAI’s GPT-4, in his view, resembles a vast, recombined record of human work rather than a new intelligence.

That framing matters because, he argues, the current debate is distorted by science-fiction thinking. Lanier notes that even prominent figures including Sam Altman have warned about severe harm, and he cites a recent poll in which half of A.I. scientists said there was at least a ten-per-cent chance that humanity could be destroyed by A.I. But he says those fears are often discussed in vague, quasi-religious terms rather than as concrete engineering and policy problems.

Lanier does not dismiss the risks. He argues that flexible, non-rigid systems can be useful, including in interfaces that adapt to people instead of forcing people to adapt to software. At the same time, he says those same systems can manipulate users, generate convincing deepfakes, and further concentrate power if left unchecked.

Recommended reading

Anthropic’s $1.5 billion author settlement gets court approval

His preferred policy direction is more specific than broad calls for a pause in development. Lanier says there is broad agreement that:

• deepfakes should be clearly labeled
• communications from artificial people should be labeled
• automated interactions designed to manipulate human thinking or actions should be labeled
• those labels should be paired with meaningful user choices

Data dignity and opening the black box

Lanier says the deeper challenge is the black-box nature of large models. He is skeptical that vague concepts such as alignment, safety, and fairness are enough on their own, because they are hard to define precisely and easy to turn into procedural box-checking.

His alternative is “data dignity,” also known as “data as labor” or “plurality research.” The idea is to trace model outputs back to the human creators whose work helped produce them, making the people inside the model more visible. In some versions, those people could also be paid when their contributions are filtered and recombined through large models.

For Lanier, that would be a cleaner answer to both manipulation and labor concerns than pretending current systems are mysterious digital beings. Big-model systems, he writes, are made of people—and opening the black box means revealing them.