Microsoft's 18 guidelines for human–AI interaction
TL;DR
AI features usually fail because users don't know what they can do, when to trust them, or how to fix them when wrong. Microsoft Research's 18 guidelines are the shortest checklist: set expectations, easy dismissal and correction, explanations, cautious updates.
AI features usually fail not because the model is bad, but because users don't know what it can do, when to trust it, or how to fix it when it's wrong. In 2019, Microsoft Research (Amershi et al.) published 18 guidelines for human–AI interaction, validated across many real products. Years later they remain the best design checklist for chatbots, copilots and agents.
18 guidelines in 4 phases
| Phase | Guidelines |
|---|---|
| Initially | 1. Make clear what the system can do · 2. Make clear how well it can do it |
| During interaction | 3. Time services based on context · 4. Show contextually relevant information · 5. Match relevant social norms · 6. Mitigate social biases |
| When wrong | 7. Support efficient invocation · 8. Support efficient dismissal · 9. Support efficient correction · 10. Scope services when in doubt · 11. Make clear why the system did what it did |
| Over time | 12. Remember recent interactions · 13. Learn from user behaviour · 14. Update and adapt cautiously · 15. Encourage granular feedback · 16. Convey the consequences of user actions · 17. Provide global controls · 18. Notify users about changes |
Applying them to today's LLM products
- Guidelines 1–2 (set expectations): the welcome screen states which documents the chatbot answers from and what it can't do; add a note that "it can be wrong, check the sources".
- Guidelines 8–9 (dismiss, correct): copilot suggestions must be dismissable with one key and editable in place, not forcing a restart.
- Guideline 10 (scope down when unsure): instead of inventing an answer, ask a clarifying question or offer a few options. This is anti-hallucination design at the UX layer.
- Guideline 11 (explain): show citations, the tools called and the data used.
- Guidelines 14 and 18 (change): switching models or prompts can change behaviour; tell users rather than doing it silently.
- Guidelines 16–17 (consequences, control): for agents that take real actions (sending email, deleting data), preview the consequences and provide a master switch.
Complemented by Google's People + AI Guidebook
Google's PAIR guidebook points the same way: decide whether AI is really needed, calibrate user trust (neither over- nor under-trusting), design how feedback is collected, and handle errors gracefully. The two complement each other: Microsoft gives a compact checklist, Google gives process and examples.
Risks of ignoring them
- Over-trust: users who don't know the limits will use wrong output in important decisions.
- Abandonment: after a few uncorrectable mistakes, users switch the AI off entirely.
- Lost trust after updates: sudden behaviour changes after a model switch make users think the product broke.
How to apply it
- Dev: Every AI feature needs an easy way to dismiss and correct output; log corrections as improvement data.
- Tech Lead: Put the 18 guidelines into the design review checklist for every AI feature.
- PM: Write a "feature limitations" section in the spec, then surface it in the UI, not just in internal docs.
Sources & references
- Guidelines for Human-AI Interaction — Microsoft Research
- People + AI Guidebook — Google PAIR