Guidelines by field
AI guidelines for software and AI developers what the field's own bodies ask
The people who build AI have the oldest codes and the newest laws. Professional codes ask engineers to avoid harm; company safety frameworks and two US state laws focus on rare, catastrophic risks. Most of these frameworks are written by the companies themselves.
Last reviewed 27 October 2026. Every claim links to its source.
At a glance
Avoid harm, be honest and fair
Respect creators' work, privacy and confidentiality.
Build ethics in from the start
Turn stakeholders' values into system requirements.
Set thresholds in advance
Define what safeguards a model needs before it crosses a dangerous capability level.
Be willing to pause
Delay deployment, or even development, if safeguards are not ready.
Publish, report, protect whistleblowers
California requires serious incidents to be reported within 15 days.
Who has spoken
- Binding law, regulation, court ruling or enforceable professional rule
- Voluntary guidance, principles or a code an organization chooses to follow
Voluntary Jun 2018
Association for Computing Machinery
Code of Ethics and Professional Conduct
Contribute to human well-being, avoid harm, be honest and fair, respect privacy, and evaluate systems thoroughly. Binding on ACM members as a condition of membership.
Voluntary Sep 2021
IEEE
IEEE 7000-2021
A model process for addressing ethical concerns during system design.
Voluntary 2023 to 2026
Twelve AI companies
Frontier safety frameworks
Published policies on dangerous capabilities, safeguards and when to pause, compared by the research group METR. The three best known were all rewritten in 2025 and 2026.
Binding In force Jan 2026
State of California
SB 53, Transparency in Frontier AI Act
The largest developers must publish safety frameworks, report critical incidents within 15 days (24 hours if danger is imminent) and protect whistleblowers, with penalties up to $1 million per violation.
Binding Enacted Dec 2025
State of New York
RAISE Act
A second state law aimed at frontier developers, amended in March 2026.
Voluntary Jul 2025
European Commission
General-Purpose AI Code of Practice
Signed by most major developers as a way to show they meet the AI Act's duties for general-purpose models.
Amplify existing pathways to severe harm.
Binding and voluntary
Binding
California's SB 53 and New York's RAISE Act for the largest developers, and the EU AI Act's duties for general-purpose models since August 2025.
Voluntary
Company frameworks (self-written and self-enforced, though California now requires the largest developers to publish and follow one), the ACM Code, IEEE standards and the EU Code of Practice.
What changed in 2025 and 2026
- All three leading company frameworks were rewritten. One company replaced its earlier if-then commitments with what it calls more achievable ones, plus regular risk reports. Anthropic makes Claude, the assistant that helped build this site.
- The first US laws aimed at frontier developers took effect in California and were enacted in New York.
Where people still disagree
Developer frameworks focus on rare, severe risks such as biological weapons, cyberattacks and loss of control, while almost every other field focuses on everyday errors, privacy and fairness. Both are real, and they call for different safeguards.
Questions to ask yourself
- Have I written down, before I build, what my system must never do and how I would know if it did?
- Who outside my team gets to check my risk assessment?
- If something goes badly wrong after release, who do I tell, and how fast?
How this lines up with my everyday guidelines
| My everyday guidelines | The field's nearest rule |
|---|---|
4. Use for good Never use AI to deceive, manipulate or harm others. | Contribute to human well-being and avoid harm (ACM Code). |
6. Stay curious and critical Question outputs. AI can be confident and wrong. | Evaluate systems thoroughly before and after release (ACM; METR). |
Not part of my everyday guidelines, which are for users rather than builders. | Set capability thresholds in advance and be ready to pause (company frameworks; METR). |