AI Ethics & Responsible Use

Things we won't build

A few categories we decline regardless of budget:

• Systems designed to deceive end users into believing they are talking to a human when they are not.
• Surveillance or scoring of employees without their informed consent.
• Automated decision-making in domains where the cost of error to an individual is severe (immigration, credit, benefits) — we'll happily build decision-support tools that a human reviews, but not the decision itself.
• Anything in obviously legally-grey territory (deepfake identity work, scraping of paywalled or copyrighted content for redistribution).

Hallucination is an engineering problem, not a feature

Every LLM hallucinates sometimes. Treating that as “just how AI is” is the wrong response. Every system we build includes:

• Retrieval-augmented generation (RAG) so the model answers from grounded context, not its own memory.
• Schema-validated output: if the LLM returns something malformed, the system retries or escalates — it doesn't pretend the bad output is good.
• Confidence scoring on every AI decision, with a tunable threshold below which a human reviews.
• An audit log of every input, intermediate decision, and action taken.
• Reversible side effects where possible (draft, not send; suggest, not execute) for high-stakes workflows.

Transparency with end users

If an end user is interacting with an AI system, they should know it. We design chatbots to disclose they're AI, with an obvious escalation path to a human. We don't build “does it pass as human?” tests as a success metric.

Bias and fairness

Bias in LLMs and embedding models is real and well-documented. For systems making decisions about people, we:

• Document the model's known limitations in the project handover.
• Recommend human review in the loop for any decision affecting an individual.
• Test the system across the demographic range it will encounter where possible.
• Surface confidence and reasoning to the human reviewer, not just the conclusion.

Environmental footprint

AI inference uses energy. We optimise: smaller models for simpler tasks, caching of repeated queries, batch processing where possible, embeddings cached rather than re-computed. We don't pretend the footprint is zero, but we don't waste tokens either.

Worker impact

Automation replaces tasks, not always people. Where a project will materially change someone's job, we surface this in the discovery phase rather than ship and hope. Engagements aimed explicitly at headcount reduction are something we expect the client to be honest about with us and with their team.