Intel Focus publishes open-source investigations that people rely on to understand contested events. Our only real asset is whether our findings can be trusted. This page sets out, in public, where artificial intelligence is allowed anywhere near that work and where it is not.
It summarises for readers the commitments in our internal Artificial Intelligence Usage Policy, which binds every volunteer, contributor, associate, contractor and partner who researches, drafts, edits, translates, verifies, designs, codes or publishes in the name of Intel Focus, on any device, in any country.
Human judgement is the product. Everything else is a tool.
All judgement, all analysis, all reasoning about evidence and all research methodology at Intel Focus is human. No exception exists for speed, workload, deadline pressure or volunteer capacity.
A named, accountable person performs each of the following personally:
The reasoning behind the rule is straightforward. Generative models produce fluent text that is optimised for plausibility rather than for truth. They fabricate sources with the same confidence they cite real ones, they inherit the biases of their training data, and they cannot reason their way to distinguishing an authentic image from a synthetic one. An investigative finding produced that way is not a finding. It is an assertion with no chain of custody.
We accept full responsibility for every word we publish, including anything produced or shaped with the assistance of a tool. “The model said so” is never a defence, an explanation or a mitigation.
The following are prohibited absolutely. No approval route exists for any of them.
We do not treat AI as contraband. Used inside the boundary set by section 1, these tools save volunteer hours we do not have. The following are permitted:
Everyone working with us applies a single test before reaching for a tool: am I asking the machine to do the thinking, or to tidy up thinking I have already done? The first is prohibited. The second is encouraged.
Nothing that identifies, or is capable of identifying, a source, victim, witness, at-risk individual or member of our team is ever entered into an AI system. Not into an assessed tool, not in a private or temporary mode, and not in redacted form. Removing a name from a document does not make it safe, because context re-identifies people.
We also treat the shape of our own questions as a disclosure risk. A series of individually harmless prompts can, in combination, reveal an investigation’s target, timing and method to a third-country processor, so our researchers are required to consider the cumulative picture their prompts paint rather than each prompt on its own.
Audio containing an identifiable at-risk voice is never submitted to a hosted service. Where sensitive material has to be translated or transcribed, we use a model running entirely on our own hardware with no network egress, or a trusted human.
Intel Focus processes personal data as a controller under the UK GDPR and, for people in the EEA, under the EU GDPR. Any AI provider processing that data acts as a processor and requires a written processing agreement before unpublished material may be used with it.
Machine translation is permitted as a comprehension and triage aid. It is never the evidentiary basis for a claim. Where a translated passage carries analytical weight — a quotation, an admission, a name, a date, a threat or a claim of responsibility — a human speaker of the source language verifies it against the original before it reaches you.
Every machine-translated passage carried into an Intel Focus product is accompanied by a translation note stating the source and target language, the exact model and version used, a confidence level with a one-line justification, whether a human speaker of the source language has verified the passage and who did so, and the date of translation.
The confidence level describes our confidence in the translation, not in the underlying claim:
Transliterated names, place names, dates, ranks, unit designations and numerals are checked by a human in every case, at every confidence level. These are the elements models most often normalise into something plausible and wrong. Machine output that reads smoothly is not thereby accurate: fluency is the failure mode, not the assurance.
Every chart, map, diagram, graphic or other visualisation produced with AI assistance and published by Intel Focus carries a visible AI-assistance notice on the asset itself, not only in the surrounding text. The notice travels with the image when the image is screenshotted, cropped or reposted, which is how such assets actually circulate. Where the publishing channel supports it, we also embed content provenance metadata under the C2PA Content Credentials standard, in addition to the visible notice and never instead of it.
These are the wordings we use:
Correcting the grammar and spelling of a human-written draft is not disclosed, because it changes nothing about who did the thinking. Analytical text does not arise this way at all, because section 1 forbids it.
We disclose in every jurisdiction, to every reader. Article 50 of the EU AI Act, which applies from 2 August 2026, requires deployers to mark AI-generated or manipulated image, audio and video content and to label AI-generated text published to inform the public on matters of public interest. We treat that as a floor rather than a target, because the obligation we are actually meeting is one owed to you and not to a regulator.
No AI-assisted asset or passage reaches publication on the say-so of the person who produced it alone. A second named person checks it and records their approval.
That reviewer confirms that every figure in a visualisation traces to a checked source; that no analytical claim originates from a model; that the disclosure notice is present, correct and legible; that any translated passage carries its translation note; and that our rules on unpublished and identifying material were respected. The reviewer is never the person who created the asset. Sign-off is recorded with both names and the date, and the absence of a record is treated as the absence of sign-off.
We assess the legal regime a provider sits inside, the route the data takes, and whether a government can compel access without notice or remedy. We do not assess the nationality of a company, and we do not maintain a list of banned brands, because a criteria-based test survives changes in the market and a list of names does not.
A tool may be used for Intel Focus work only if every one of the following holds:
Applied to the market as it stands, this test excludes services whose governing law gives state authorities far-reaching access to data held within the sphere of influence of companies subject to it, whatever those services are worth technically. For an organisation whose work concerns state conduct, and whose contributors may be exposed to the states they research, that access risk has no acceptable level. Equally, a provider incorporated in Europe or North America is not automatically compliant. We check the criteria rather than the flag, and no service is approved because a colleague uses it.
Beyond the choice of tool, model training is switched off on every account used for our work, including personal accounts. We never use share-a-chat, public-link or publish-conversation features for any Intel Focus content, because those features generate publicly addressable URLs that search engines have repeatedly indexed. Persistent memory, chat history and third-party connectors to email, cloud storage, calendars and code repositories are disabled on any account that touches unpublished material.
Eventually something will go wrong. What matters is what happens next. Where an AI-assistance failure reaches a published product, the following applies regardless of fault:
We do not silently amend published material. A correction that is not visible to a reader who saw the original is not a correction.
Everyone who works with us completes AI literacy training before working on any product, and a refresher annually. It covers the failure modes of generative models, the rules on this page, and the internal policy in full.
If you believe something we have published carries a machine-generated claim, an unlabelled AI-assisted asset, or a translation that does not hold up, tell us at info@intelfocus.org. We would rather hear it from you than not hear it at all.
Intel Focus exists because someone has to do the patient, unglamorous work of establishing what actually happened. That work is human. It is slow because verification is slow, and it is trusted because it is slow. We will use machines to remove drudgery from that work, and we will not use machines to shortcut the part that makes the work worth publishing.
Where this is inconvenient, that inconvenience is the point.
This page sets out the public commitments in the Intel Focus Artificial Intelligence Usage Policy, version 1.0, approved by the Director. It is reviewed every six months and immediately on any material regulatory change. Where a partner, funder or publisher imposes stricter conditions on AI use than those described here, the stricter condition applies. Intel Focus Ltd is registered in England and Wales, company number 17112681.