Watchtower
Watchtower tracks when AI companies change their safety policies — and when they break them.

Changed terms of service concerning use of user data. Did not announce or report that change was made.
Cognition
Dec 11, 2024
On December 10, coinciding with the general release of their coding agent Devin, Cognition updated the terms of service for their platform to add more details concerning the use of user data. As far as we can tell, they did not announce these changes, and they also have not reflected that changes were made in the "last updated" variable on their ToS page.
Previously, the relevant section read:
3.3 Use of Customer Data: Cognition may use Customer Data to provide, maintain, develop, and improve our Services, enforce our terms and policies, and keep our Services safe.
It now reads:
3.3 Use of Customer Data: Cognition may use Customer Data to provide, maintain, develop, and improve our Services, enforce our terms and policies, and keep our Services safe. Through the Data Controls settings in the web application, you may opt out of the use of your input and output data for model training purposes ("Training Data"). Notwithstanding your Data Controls settings, Cognition reserves the right to collect, analyze, and utilize usage patterns, metadata, and other non-Training Data associated with your use of the Services for operational, analytical, and improvement purposes.
You can read the full page here.
OpenAI
Aug 18, 2026
OpenAI updated its Model Spec, the document outlining intended model behavior
xAI
Aug 17, 2026
xAI updated Grok 4.6's model card after release. A changelog was included.
Aug 14, 2026
Google updated its Gemini 3.7 Flash model card after publication, making changes to language in the “Key Results for Gemini 3.7 Flash” column in the Frontier Safety Assessment section
OpenAI
Aug 3, 2026
OpenAI updated its system card for two of its models: GPT 5.6 and GPT Live.
xAI
Jul 20, 2026
xAI made changes throughout the model card for Grok 4.5, including some that appear to be persistent errors
Frequently asked questions
Have more questions? Our team is happy to help, contact us.
We engage in a combination of research, outreach, and public advocacy to ensure that AI companies are meeting public expectations and living up to their past promises, in order to ensure responsible AI development and deployment.
We review technical literature, regulatory guidance, and case studies to distill concrete measures that will meaningfully improve public safety — such as frontier-model risk assessments, red-teaming requirements, and whistle-blower protections — and advocate for the most important voluntary steps that companies can take today to ensure they are acting responsibly.
We also monitor whether companies follow their stated policies and industry norms. When we find evidence of back-tracking or inadequate risk controls, we document it and call for corrective action — mobilizing employees, customers, and civil-society allies until the company adopts the necessary safeguards.
Finally, we publicize our research to inform the public of how AI companies stack up on safety and responsibility. We release our work in the form of scorecards, independent reports, open letters, and long-form writing so that regulators, investors, and the wider public can see how individual developers perform on safety and responsibility.
Various AI experts including Nick Bostrom and Stuart Russell have compared the development of advanced AI to the myth of King Midas.
According to legend, King Midas was once granted one wish by the god Dionysus: that everything he touched would turn to gold. At first, he was thrilled with his new powers. But the King soon discovered that he couldn’t touch food, water, or even his family without instantly turning them to metal. In other words, he got exactly what he wanted in pursuit of immense wealth — and it turned out it wasn’t what he wanted at all.
Much like King Midas, AI companies are now eagerly pursuing incredible wealth and power by developing increasingly powerful AI systems. But ensuring that these systems act in alignment with our values is still an unsolved technical problem. If we misspecify even a single goal for these systems, how will we prevent them from causing an incredible catastrophe – if they follow our instructions at all?
In the words of Stuart Russell, “If you continue on the current path, the better AI gets, the worse things get for us. For any given incorrectly stated objective, the better a system achieves that objective, the worse it is.” The Midas Project exists to ensure that AI companies do not take this extraordinary gamble without public accountability and oversight.
The Midas Project is a nonprofit organization founded in early 2024 by Tyler Johnston. Our work is supported by a small core team and a wider base of volunteers and supporters. We are a nonprofit, tax-exempt, 501(c)(3) organization that relies on donations from the public.
No. One of our central values is being pro-technology.
Progress in technology has improved lives for millions of people around the globe (after all, without it, we wouldn’t have penicillin, air conditioning, or the internet). Artificial intelligence is already being used to help improve medicine, education, and overall living standards. We believe this progress should continue, and we hope AI will be a positive force in the world.
But we may not be on track to realize this future. Without technical breakthroughs, we risk developing powerful AI systems that act against user intent, or can be misused by bad actors to cause tremendous harm. Powerful AI could also concentrate unprecedented power among a handful of AI companies and exacerbate social inequality. To avoid these downsides, AI must be developed with caution, transparency, and public oversight. That’s why The Midas Project is committed to raising awareness about the risks of AI and ensuring that everyone is given a chance to make their voice heard.
If you’d like to get involved, consider signing up for our newsletter, joining as an official volunteer, or making a charitable donation today.
You can email us at info@themidasproject.com, or reach out via the form on our contact page.