Alignment and safety

Alignment (in everyday AI talk) means making systems behave in line with intended human goals and constraints—not just maximizing a narrow score. Safety covers preventing harms from mistakes and misuse.

Labs publish safety policies and evaluations because models can be helpful *and* risky. You do not need conspiracy theories to care: scams, privacy leaks, and biased decisions are enough (Course 29).

Common layers:

• Training data filters and policies
• Post-training (instruction tuning, preference methods)
• Product rules and refusals
• Monitoring and abuse reporting
• Human oversight for high-stakes actions

No layer is perfect. Defense in depth is the point.

A model refuses to provide bomb-making instructions but still helps with chemistry homework. Boundaries are designed—and sometimes bypassed (literacy without recipes: Course 29).

Read one lab’s published usage policy or model card and list three allowed uses and three disallowed uses in your own words.

⚠️ Myth: Safety filters mean the model is “censored into stupidity.”

✓ Reality: Products trade off capability, liability, and harm reduction; critique is fair, denial of tradeoffs is not.

⚠️ Myth: Alignment is only about sci-fi takeover.

✓ Reality: Much safety work is present-day misuse and reliability.

• NIST AI RMF: https://www.nist.gov/itl/ai-risk-management-framework ↗
• ANN live: https://www.ainerdnetwork.com/learn/alignment-and-safety ↗
• Lab model/system cards (cite the specific card when quoting a model)