minntelligence
What is AI safety?
AI safety is the work of making sure AI systems do what they're supposed to do — reliably, honestly, and without causing harm — as they get more capable and more widely used. It's part engineering, part research, part policy.
What people mean by it
- Problems that already exist. Today's systems make things up with confidence, repeat biases from their training data, leak private information, and get used for scams, deepfakes, and spam. A lot of safety work is just measuring and reducing this.
- Problems that get harder as systems get stronger. A system can score perfectly on the test you gave it while missing what you actually wanted. Getting a system to genuinely pursue the right goal — not just look good on the metric — is called alignment, and it gets harder to check as systems act more autonomously.
- Who decides. What gets tested before release, what standards apply, and who is accountable when something goes wrong. That's the governance side: evaluations, standards, and oversight, mostly before there's a problem rather than after.
What safety work looks like
In practice: testing systems before and after release (evaluations), trying to break them on purpose (red-teaming), figuring out how they work inside (interpretability), watching for real-world incidents, and writing the standards that say what "tested enough" means. Our own small piece of this is measurement — for example, checking whether public comments to government were AI-generated.
People in this field disagree — a lot — about how big the long-term risks are and how fast things will move. You don't need to buy any particular forecast to think "test it before you ship it, and keep a human able to check the work" is a good idea. That's most of AI safety.
Learn more — talk to a person
Curious, confused, or worried about what any of this means for you or your work? Talk to a person about it. No sales pitch, no hype, no doom. Most conversations for individuals, students, and non-profits are $0.
Start here: contact form or amin@minntelligence.fyi. To meet people locally, see AI safety groups (MN / Midwest).
Resources
Start here:
- 80,000 Hours — AI as a problem profile: the clearest overview of why people work on this and where the work matters most.
- AISafety.com: an open directory of AI safety communities, organizations, events, jobs, and courses.
- BlueDot Impact: free online courses on AI safety and governance, from a two-hour intro to deeper technical tracks.
- EA Forum — AI safety: current community discussion, readable without a technical background.
Go deeper:
Local: