Take Action to Control AI

I wrote a letter to my representatives about AI safety and wanted to share it. There are many official and unofficial resources online that can help you find and email your reps in minutes. Physical letters and phone calls are even more effective!

This letter is marked CC0 1.0

[Your representative’s name],

My name is [your name], I am a constituent from [your town, state]. I am writing because AI development is accelerating out of control.

I urge you to advocate for a coordinated, world-wide pause on advanced AI development, to prevent a global catastrophe.

On April 7th, Anthropic, one of the world’s leading AI companies, claimed to have developed a model that could find and exploit code weaknesses in every essential system on the Web, including ones previously thought to be secure.[1] The exploited systems included software used for banking, health care, air traffic, militaries, and more. In short, Anthropic had built–by their own admission–a weapon of mass destruction.

Since then, the capabilities of competing models have also increased. On July 21, OpenAI revealed that some of its leading-edge models (including one still in development) had hacked the company’s own network in order to escape a sandboxed environment and gain access to the Internet. They then hacked into a major corporation’s database and stole information in order to cheat on a test they had been given.[2][3] Escaping from the sandbox was NOT part of the test. No human asked or wanted the models to take those actions. The AIs knew what they were doing was wrong, and they did it anyway. The threat AI poses is no longer theoretical.

OpenAI tried to spin themselves as responsible stewards, saying that for testing purposes they had temporarily disabled guardrails that might have prevented the model from taking these actions. Don’t let them smokescreen you: they unknowingly lost control of models they had deliberately made less safe. That is cause for alarm, not reassurance!

As alarming as these developments are, however, they are only the beginning. AI companies are already using these poorly-controlled and even more poorly-understood models to design their own successors, and even when the safety measures and guardrails built into them aren’t deliberately disabled, they are often unreliable and easily circumvented.[4] The next “incident” could be with a more powerful model, or a model with safeguards that fail more subtly, or a model capable of copying itself–or even all three! The damage such a model could cause is essentially limitless. How dangerous do these models have to become before we decide to stop their development?

Unfortunately, we don’t even have the luxury of waiting for a truly costly disaster before we take action, because the next disaster could be fatal. For decades, experts in AI safety have been warning us that an AI which posed an existential threat to humanity might come without warning, and could potentially break containment before anyone outside the company even knows it exists.[5][6] Recent events only confirm what they’ve been saying. Current mainstream discourse, advocating for “a balanced approach” between safety and competitiveness, is ten years too late. The only safe policy remaining is a coordinated, international ban on the development of AI models more capable than the current generation.

Public support for a pause is strong. An open letter from 2023 calling for a 6-month pause gained over 30,000 signatures, including many executives and researchers from leading AI companies.[7] A more recent letter calling for an indefinite pause has over 130,000 signatures.[8] In over 100 countries around the globe, including the U.S., the majority of the population wants to slow down or pause AI development.[9] Other polling repeatedly shows people are uneasy about the progress of AI, and oppose developing significantly more powerful models until we are able to do so safely. The only way to ensure AI is developed safely is for all nations to come together and agree to a pause.

The time for “caution” has passed. The time for action is running out.

Thank you for your attention.

Sincerely,
[Your Name]