Where we draw the line on AI

We have written an AI ethics policy. That sentence could mean almost nothing, so this is what ours actually says, and why we bothered.

AI is now part of how the work gets done here. The useful question is not whether to use it but how, in a way that makes the work better rather than cheaper and worse. These tools are powerful enough to deskill a team, displace people, and harm the users of whatever gets built. They are also powerful enough to make the work more interesting and more durable. Which of those you get is a choice, so we made ours explicit.

The lines we won’t cross

Some of the policy is a plain list of things we will not build, whatever the commercial pressure. No AI companions that pretend to be a real person. No generating images or video of identifiable people without their consent. No surveillance systems that monitor workers rather than help them. No systems that make consequential decisions about people, in employment, credit, healthcare, justice, or access to services, without a human meaningfully in the loop. No weapons. Nothing trained on material whose rights were never obtained.

Writing these down is easy when no one is asking you to cross them. The point of writing them down now is that the answer is ready before someone does.

How we actually work with it

The rest is about the daily habits, and it comes down to judgement staying with people. A person owns every output, however it was produced, which means we do not ship code we do not understand to save time. We choose tools that extend our people’s judgement rather than replace it, and we hold what we build for clients to the same line. We tell clients where we use AI in their work, and where we won’t. And we keep a record of how AI-assisted code was produced, so when a client or a regulator asks the hard question, the answer exists.

The bits we won’t pretend about

An honest policy has to cover the things the industry would rather gloss over. The value in these models is mostly in their training data, and a lot of that data was taken without a licence. That is a real problem, not a technicality, and it is not one any single company can fully solve. There is an environmental cost too, small per query but real in aggregate, which we think calls for being deliberate about where AI genuinely earns its use rather than for pretending it is free or pretending it is forbidden.

None of this is a marketing position. It is closer to a working document we can be held to, by our clients, by the people who work here, and by anyone who asks. The full policy is here, and we mean it to be read.