What is the open letter asking to pause giant AI experiments?

The letter is titled "Pause Giant AI Experiments." The letter's page dates it March 22; it was published with its first signatures at the end of March and widely reported on March 29. Its central request fits in one sentence: AI labs should immediately pause, for at least six months, the training of any AI system more powerful than GPT-4, the model OpenAI released on March 14.

The letter matters less for what it can force, since it has no legal power, than for the debate it has opened. It has put a question usually discussed among specialists in front of the general public: how fast should this technology move, and who decides?

What the letter asks for

The letter makes several specific requests:

• A pause on the largest training runs. Everyone should be able to check that it is happening, and every major lab should take part. If labs do not agree quickly, the letter says governments should step in with a moratorium, a legally required halt.
• Shared safety protocols. Labs and independent experts should use the pause to write safety rules for advanced AI, checked by outside auditors.
• New governance. It asks policymakers for regulators dedicated to AI, tracking of very capable systems and large pools of computing power, watermarking to help tell real content from AI-generated content, liability for harm caused by AI, and public funding for safety research.

The letter is careful to say it does not ask for a halt to AI development in general. Its target is the race toward ever larger models whose behavior even their builders struggle to predict, often called "black boxes" because you cannot see why they produce a given answer.

Who signed, and why

The Verge reported early signatories including Steve Wozniak, who co-founded Apple, Jaan Tallinn, a co-founder of Skype, author Yuval Noah Harari, politician Andrew Yang, Elon Musk, and AI researchers and executives such as Yoshua Bengio, Stuart Russell, Gary Marcus and Emad Mostaque. By the end of March, more than 2,000 people had signed. The list has had problems: some names were added as jokes, including OpenAI's chief executive, Sam Altman, so not every name on it can be taken at face value.

Supporters make the case laid out in the letter. They argue that labs are locked in a competition to release more powerful systems faster than anyone can understand or control them. The letter asks whether we should let machines flood our information with propaganda, automate away fulfilling jobs, or risk losing control, and it says such choices should not be left to unelected technology leaders. It also points out that OpenAI itself has written that independent review before training future systems may become important. The signers say that moment has arrived.

What the critics say

Criticism has come from several directions. One of the most detailed responses came on March 31 from Timnit Gebru, Emily M. Bender and Angelina McMillan-Major, authors of a 2021 research paper on the risks of large language models, known as "Stochastic Parrots," which the letter itself cites.

They agree with some proposals, such as watermarking synthetic media. But they argue the letter focuses on imagined risks from powerful "digital minds" while ignoring harms happening now: the exploitation of workers who label and filter data, the use of people's work without permission, the spread of synthetic media, and power concentrating in a few companies. Instead of a pause, they call for regulation that requires transparency: clear labels on AI-generated content, disclosure of training data and model designs, and companies held responsible for what their products produce. They also want the people most affected by AI to have a say, not only executives and academics.

Others doubt the pause will happen at all. The Verge judged the letter unlikely to change much in a market where large tech companies are rushing out products, while seeing it as a sign of growing pressure for regulation.

Here is a side-by-side view of the main positions:

Should AI development slow down?

There is no settled answer, and thoughtful people disagree. Those who favor slowing down point to how quickly systems like GPT-4 have arrived and how little outside oversight exists. A pause, in their view, buys time for safety research and for laws to catch up.

Those who disagree raise different concerns. Some think the danger lies less in future models than in how today's systems are already used, and that a pause would distract from that. Others doubt that companies competing to ship products would actually stop, or that a voluntary letter could make them.

The two sides share some ground. Both the letter and its Stochastic Parrots critics call for ways to label AI-generated content and for holding companies responsible for harm their systems cause. As a reader or user, you can follow the debate by asking concrete questions of any proposal: what exactly would stop, who would check it, and which harms would it address?