# AI Safety and Governance

Safety research and institutional governance seek to manage AI's risks. This sub-article expands [Section 3.4 and 4.3 of Philosophy](../philosophy.md) and [Section 4 of Recent Topics](../recent.md).

## Types of Risk
- **Misuse** — harmful applications of powerful models.
- **Accidents** — failures from misaligned objectives or unexpected behavior.
- **Existential risk** — the possibility that a misaligned superintelligent system causes catastrophic harm.

## Safety Research
- **Alignment** ([alignment.md](alignment.md)) and **interpretability** ([interpretability.md](interpretability.md)).
- Red-teaming, adversarial robustness, and capability/risk evaluations.
- Preparedness frameworks in frontier labs.

## Governance Instruments
- **EU AI Act** (2024) tiers obligations by risk.
- National strategies, executive orders, and international summits (Bletchley 2023, Seoul 2024).
- Compute governance, licensing, and export controls.

## Open Questions
- How to verify safety at scale?
- Who audits frontier models?
- How do we coordinate globally without centralizing power?

## Why It Matters
Safety and governance determine whether powerful AI benefits or harms.

## Sources
- Bostrom, N. (2014). *Superintelligence.*
- European Parliament. *AI Act* (2024).
- Bletchley Declaration (2023).

*Back to [Philosophy](../philosophy.md) · See [Recent Topics](../recent.md).*