Responsible AI

Responsibility is scored at every lock

Lock 1 scores meaningful (not gratuitous) AI use. Lock 2 dedicates ten points to responsible AI and privacy. Lock 3 evaluates what you actually shipped. A responsible-AI plan is not a compliance paragraph — it is a design artefact that names specific harms and shows specific countermeasures.

Name your harms specifically

Generic statements ("we will avoid bias") score nothing. Specific analysis does: who could be hurt if your model is wrong, in which direction, and how badly? A triage assistant that under-flags is different from one that over-flags. A scheme-eligibility tool that hallucinates a benefit sets a family up for rejection. Write the failure stories, then design against them.

• Wrong-output harms: what does a false positive and a false negative each cost, and to whom?
• Exclusion harms: who cannot use this — language, device, disability — and what happens to them?
• Data harms: what leaks if you are breached, and how bad is it?
• Misuse harms: who could weaponise this tool against its own users?

Privacy by design, practically

• Collect the minimum; prefer session-only and on-device processing where possible
• Encrypt personal data at rest; never log raw personal content
• Give users visibility and deletion rights over their own data
• For minors: guardian consent, and even stricter minimisation
• Know your third parties: every API that sees user data is part of your privacy surface

Test for bias with your real users in mind

Slice your evaluation: does accuracy hold across genders, languages, urban and rural inputs, low-end cameras? A small honestly-sliced evaluation beats a large aggregate number. Report the slices in your submission — including the ones that look bad, alongside what you changed.

Keep a human accountable

Mirror Foundry’s own design: AI recommends, humans decide. For any output that affects a person’s health, money, safety or legal standing, your product needs a human pathway — an escalation contact, an expert review step, a "get qualified help" route that is prominent, not buried. Demonstrating this pathway working is one of the strongest things a Lock 3 demo can show.