AI Transparency

Most of this system is not a language model

Keel is a strategic intelligence system with a lot of moving parts. Much of it is conventional software that behaves the same way every time. Some of it is AI. The part that writes the prose you read is a language model, and that is the part that can be wrong in ways the rest cannot.

Treating “AI” as one undifferentiated thing is how people end up either trusting all of it or none of it. Neither is right. This page says which part is which.

Which Parts Are AI

• Collection and tracking: Not AI. Conventional software gathers sources and follows a development across editions. It behaves the same way every time, and its failure mode is a gap in coverage rather than an invention.
• Historical context: Not generative. Related past cases are drawn from our archive, never written into existence. A match can be weak or unhelpful, and we show you enough to judge that for yourself.
• Sorting and prioritising: Partly AI, working within fixed categories and consistent criteria rather than writing freely. The realistic error is something miscategorised or set aside, not something fabricated.
• The written analysis: This is the generative part, and the one to read critically. A language model drafts the strategic reading and writes the prose you see. We check generated passages against their sources where that can be done automatically, and flag what drifts - but that check does not cover everything, and it is not a guarantee.
• What you are shown: Selection and framing for your professional context. This affects what reaches you, not whether it is true. Detailed under Automated Processing and Profiling.
• Visuals and audio: Generated imagery and any synthetic narration we publish. Illustrative, never evidentiary - nothing here is a photograph of an event.

What This Is Not

Being direct about this is not a legal formality. It is the difference between using our work well and using it badly.

• The prose is not a verified record: Language models can be confidently wrong - misstating facts, inventing detail, misattributing sources, all while sounding certain. This applies to the written analysis, not to the retrieval and tracking underneath it. Verify anything you intend to act on.
• Foresight, not forecasting: Keel's job is to identify what is forming and what would confirm it - not to call outcomes. Where we do describe how something might develop, that is a possibility to watch, not a prediction, and we keep no scoreboard suggesting otherwise.
• Not exhaustive coverage: Broad, but partial, and skewed towards what gets published in the languages we cover. Absence from an edition is not evidence that nothing happened.
• Not trained on you: We do not train models on subscriber data and we do not sell your data to model providers. Personalisation happens when an edition is assembled, not inside a model’s weights.

Human Oversight

Editorial responsibility for what we publish rests with Keel Intelligence. Someone sets the editorial direction, reviews the assembled edition, and can alter or reject any part of it before it sends.

Whether that review happened is recorded for each edition, and the notice on the edition tells you which one you are holding. Where it was reviewed, the notice says so. Where it went out as assembled, the notice says that plainly rather than implying an oversight that did not happen.

Which Models

Chosen per task, not by brand.

• Run on our own hardware: A substantial share of the work runs on our own machines in Europe. That content never leaves them.
• Third-party models: Operated outside the EU, used where cultural precision or reasoning depth demands it. Engaged under data processing agreements with standard contractual clauses.
• Interactive features: Where we offer one, it will be an AI system and will say so before you use it, and its answers will carry the same limits as the written analysis. Nothing of that kind is live today.

Marking and Provenance

Alongside the visible notice, our published editions, feeds and generated media carry machine-readable markers identifying them as AI-generated. Text watermarking has no reliable technical standard today, and the AI Act acknowledges as much; we declare provenance in document metadata and feed fields, and embed it in images and audio at the point of generation.