How we use artificial intelligence

AI tools are now ordinary in everyday work, and they bring real gains in efficiency, drafting and insight alongside real questions about confidentiality, accuracy, bias and environmental cost. Saying nothing has become its own kind of statement. So this page says what we do.

It is drawn from SOIF’s internal AI policy, adopted in December 2025 and next reviewed in June 2026. That document governs how our team works; this one tells you what it means for what you are reading.

• Be intentional. AI is a tool, not a replacement for thinking. We use it where it adds real value and not because it is there.
• Be transparent. AI-generated or AI-assisted content is disclosed — in public work, in formal documents, and to each other.
• Protect confidences. Client-identifiable information, and anything not already public, is never entered into an unapproved tool.
• Verify everything. Outputs are checked, facts validated, references confirmed one by one. We assume any output contains errors until shown otherwise.
• Account for bias. These systems reproduce the biases of what they were trained on. Outputs are checked to make sure a diverse range of perspectives survives, and we go to our own global networks rather than to a model when we want to know what people think.
• Represent people responsibly. AI-generated content must not misrepresent, marginalise or exclude any community.
• Count the environmental cost. These tools consume real energy. We do not use them for tasks a simpler one would answer.

Our team may use AI tools to support our work. Unless otherwise agreed with a client, that use is limited and controlled: it may enhance drafting, suggest improvements to an argument, check for errors, or support ideation.

We do not use AI to generate first drafts of client deliverables or proposals. Our work is led by humans, we review any content for accuracy, and our team retains full accountability for the advice and outputs we provide.

For data analysis, visualisation and research we may use specialist tools that incorporate machine learning and natural language processing.

We maintain strict safeguards against sharing confidential client or personal data with external AI services, and we review our use of them for legal and ethical compliance.

Nothing on this site is generated by AI when you load it. There is no model between you and the book. The words, the artwork and the structure are served as they were written and made.

Nothing you write here is sent to an AI system. Not your email, not your notes on a passage, not anything you post in Play. No reader data of any kind reaches a model.

One editorial task did use AI. The book links its key concepts to a glossary, and finding every place a concept appears across twenty-one conversations is the kind of work that is both mechanical and easy to do badly. We used a language model, offline and against the book text only, to suggest those links. Every suggestion is marked as a suggestion and confirmed by an editor before it appears — the model proposes, a person decides.

AI was used in writing this book, as an aid to writing rather than as its author. Catarina has written her own account of that, at length and without softening it, and it is better read there than summarised here: About the author → How the book was written.

The short version, in her words: she does not naturally think in linear arguments. The thinking happened in conversations, systems maps and practice; AI helped convert that oral material into a draft she could then shape, challenge, reorganise and complete. “It did not replace the thinking. It helped me translate the thinking into writing.” Multiple people edited and sharpened the manuscript, and an advisory board gave continual feedback on intentionality and ethics.

She is also direct about why that is uncomfortable — that large language models are rooted in the past by design, trained on what has already accumulated power and visibility; that their datasets carry the exclusions of the societies that produced them and reproduce those dynamics at scale; and that the dominant business model is extractive, in creative labour, working conditions and environmental cost. Her judgement was that the risk of not writing — of not making a primer that can travel — outweighed it, and that being explicit about the decision is part of mitigating it.

The conditions she set for herself, which we hold this site to as well:

• AI can surface, organise, synthesise and challenge ideas, but it cannot and should not generate them independently.
• Be explicit about attribution and provenance wherever possible.
• Treat AI as something requiring active stewardship rather than passive acceptance.
• Use AI in service of power-brokering rather than power concentration.

One thing worth naming, because a careful reader will notice it. SOIF’s rule above is that AI does not write first drafts of clientdeliverables. This book is not client work: it is the author’s own, written under the conditions she has set out, and she is accountable for every word of it. The two are not in conflict, but they are different commitments and it would be evasive to let one stand in for the other.

Being clear about the risks is part of using these tools honestly, so here are the ones we hold ourselves against.

• Confidentiality. Inputs to a consumer AI tool may be retained or used for training. We disable that where a tool is used at all, and keep confidential material out of them entirely.
• Fabrication. These systems state invented facts, sources and quotes with complete confidence. Every reference is verified by hand.
• Bias. Outputs can reinforce stereotypes and quietly drop perspectives. We do not use AI to stand in for asking people.
• Trust. Using AI in published work without saying so damages the relationship it was meant to serve. Hence this page.
• Environmental cost. High energy use for marginal gain is not a good trade, and we treat it as a reason not to.

If something here is unclear, or you want to know more about how a particular part of this book was made, write to info@soif.org.uk. What we do with your data is set out separately in the privacy notice.