Responsible AI
First written: 5 July 2026 · Last updated: 10 October 2026
AI writes much of our code. People stay responsible for all of it.
The short version
- A named person is accountable for every tool.
- AI-written code is tested and reviewed by people before it reaches a participant.
- We say where and how AI was used.
- AI results are shown as likelihoods, never as facts.
How we use AI
Members build their tools with AI coding assistants, mainly Claude Code by Anthropic. AI helps write code, tests and documentation. It does not decide what a tool should measure, whether it is ready, or how its results should be read. People do.
People stay accountable
Every tool has a named member who is responsible for it. AI suggestions are reviewed before they are accepted, and the responsible member must be able to explain what the code does.
Verify, don't trust
Code that runs is not the same as code that is right. AI-written code is tested against known answers and edge cases, and reviewed by people, before it is used with real participants or real data.
Open about AI use
Each tool's documentation says where AI was used to build it and which models were involved, so other researchers can judge the tool and reproduce its results.
Recorded versions
When an AI model is part of how a tool produces results, for example a model that classifies text, the tool records which model and version produced each result. If the model changes, results from before and after can be told apart.
Participants' data stays protected
We do not paste participants' personal data into AI chat tools. Where a tool sends data to an AI service as part of how it works, people are told before it happens, and only what is needed is sent.
Nothing synthetic passed off as real
AI can generate convincing fake data. We never present generated content as real research data, real participants or real results. Demos and examples use clearly labelled sample data.
Likelihoods, not verdicts
When a tool uses AI to make a judgement, such as whether a post is likely AI-written, it presents the result as a likelihood, not a fact, and leaves the decision to the person reading it.
Watching for bias
AI systems can work better for some groups than others. We test tools on varied inputs and tell researchers about any gaps we find, so they can account for them.
Questions
Questions about how we use AI are welcome at hello@jointhinkspark.com.