How I build AI systems responsibly
The AI agents I build and run deal with real people’s information. At work they help with hiring intake, onboarding, compliance records and training enrolments. Outside work I run public projects that test how far AI models can be trusted. Below are the commitments I hold myself to, based on the ACS Code of Professional Ethics (2023), the Australian Privacy Principles and the Maiam nayri Wingara principles.
Consent
When someone gives an organisation their information, they give it for a purpose, and that doesn’t cover handing it to an AI system. Before an agent reads anyone’s data I check that the use matches the purpose it was collected for. If it doesn’t, the agent goes without the data until the person has been told and the organisation has signed off. Consent over Aboriginal and Torres Strait Islander data is collective and can be withdrawn, so fields like Indigenous status are kept out of every agent’s context unless a community has agreed to their use.
Data minimisation and retention
Each agent sees only the fields its task needs, and once the task is finished it keeps nothing. Its logs record what it did without copying anyone’s personal details. Personal data stays in the systems that already hold it under the organisation’s privacy obligations, and I don’t copy it into new tools or test sets, which is why my public evaluation work uses only synthetic or public data.
Human oversight
Agents can prepare, sort and draft, but a person makes every decision about an individual: who’s shortlisted, whether a record is compliant, whether something goes out. My own research is the reason I hold to this. I tested three AI models as judges against my own labels, and they agreed with me only slightly more than chance would, agreed with each other far more, and one of them favoured its own answers. A system can be consistent and still get it wrong every time, and when the subject is a person, that’s not acceptable.
Bias
An agent that assesses people doesn’t go live until I’ve tested it against human judgement on real cases, run it again with the inputs reversed, and checked for length and self-preference effects. I also write down the things it isn’t allowed to decide.
Security
Anything an agent reads can contain instructions, so I handle model output the way I’d handle any other untrusted input. Code checks it, and it never builds a request by itself. Credentials are kept in a managed store and never in files or prompts, and every change to my public repositories gets scanned for leaked secrets.
Speaking up
The ACS Code says I shouldn’t keep quiet about unprofessional conduct. If an automated system I work on harms people or breaks these commitments, I’ll take it to my manager first, in writing.
I’ll revisit this statement once a year, and any time I build a system that makes or shapes a decision about a person. The research behind the human oversight section is Who Judges the Judge.