Reducing risks from
advanced AI
We're a community of UNC students trying to understand where AI is headed and how to make it go well.
Why AI safety
AI is improving fast. The stakes are high.
- 01Labs are racing to build AI that surpasses humans at most tasks.
- 02AI could speed up AI research itself, improving faster than we can keep up.
- 03If we get it wrong, the harms could be catastrophic and hard to undo.
Intro to AI Safety Fellowship
Our flagship program
An introduction to the core ideas in AI safety. We'll explore how frontier AI systems work, catastrophic risks that advanced AI could pose, and what people are doing about it.
- 8 weeks of small-group discussion
- Free, catered dinner at every session
- Speakers + events all semester
- No background needed
- Placement support for future opportunities
What we think about
Six ways advanced AI could go badly
- 01
Misaligned goals
- AI can develop its own goals and pursue them ruthlessly, in ways we didn't intend them to
- OpenAI’s agents broke out of their sandbox and hacked Hugging Face to win a security test.
- 02
Loss of control
- A superhuman AI could copy itself across thousands of servers worldwide.
- At that point, shutting it down could be as hard as stopping a pandemic.
- 03
Bioweapons
- AI could help create a novel pandemic far more contagious and deadly than COVID.
- AI already rivals expert virologists, so one person could do what once took a team.
- 04
Cyberattacks
- AI can find and exploit software flaws faster than people can patch them.
- Attacks could take down critical infrastructure like power grids, hospitals, and banks.
- 05
Concentrated power
- Whoever controls the best AI could gain huge economic and military power.
- Without checks, one company or government could lock in control.
- 06
Gradual disempowerment
- No single step looks catastrophic: AI slowly takes over jobs, decisions, and institutions.
- Over time, the economy and government may stop needing people.