Responsible AI

1What we won't build

We decline work on:

• Systems intended to deceive people about whether they're dealing with a machine.
• Biometric surveillance or emotion inference used to score, rank, or police individuals.
• Autonomous systems that take consequential decisions about a person's liberty, livelihood, or health without a human who can see the reasoning and overrule it.
• Synthetic media of real people made without their consent.
• Anything designed to mislead, to suppress participation, or to target a group on a protected characteristic.
• Weapons and systems whose purpose is to cause physical harm.

We've turned down work on these grounds. We'll keep doing it.

2Data

Client data trains client models. Full stop. It is never pooled, never used to improve a model for another client, and never sent to a third-party API without written instruction. Where we can, we train and serve inside the client's own infrastructure so the question doesn't arise.

We ask about provenance before we ingest anything: where the data came from, what consent covered it, and whether we're allowed to use it for this. If a client can't answer, we don't train on it.

3Evaluation

We build the eval set before we build the model, and it includes the cases we expect to fail — not just the ones we expect to pass. Where the model's output affects people, we measure performance across subgroups, not just in aggregate, and we report the gaps even when they're inconvenient.

Every model ships with an honest statement of what it can't do, where it degrades, and the conditions under which it should not be relied on.

4Humans in the loop

Consequential decisions get a human. That means a confidence threshold below which the system escalates rather than decides, an interface that shows the person why the model said what it said, a route to override that doesn't require an engineer, and a log of what was overridden — because that log is the best training data you will ever get.

5Traceability

Every inference in a production system we build is logged with its inputs, its output, its confidence, and the model version that produced it. When someone asks "why did it do that, in March", the answer exists.

6Teaching

We teach AI in 50+ institutions, and every programme covers failure modes, bias, evaluation, and the limits of these systems — not as a closing lecture, but as part of how you build. An engineer who has never seen their model fail on a subgroup hasn't finished learning to build it.

7Getting it wrong

We will get things wrong. When we do, we want to hear about it. If you believe a system we built is causing harm, write to ‹responsible-ai@eurekamoment.co›. It reaches an engineer, not a queue, and we will look into it.