AI Processing

Last updated 2026-07-21. Under legal review.

This page describes the AI processing Frontfiles runs on uploaded content, as it is actually implemented and enabled today. It is part of how the service works and is referenced by our Terms. Where a point is still with counsel, we say so on this page rather than papering over it.

What AI does on Frontfiles

When a creator uploads content, Frontfiles runs an ingest-time pipeline that proposes structured metadata about that creator’s own upload:

• Image description, caption, and headline suggestions (Google Vertex AI, Gemini).
• Structured image detections such as labels, text in the image, and landmarks within regional capability (Google Cloud Vision).
• Grouping suggestions (“these uploads look like one story”), computed on our own infrastructure over embeddings. To name each group, the captions of its files, and the upload brief if one was written, are sent to Google Vertex AI.
• Search-ranking embeddings (Google Vertex AI text-embedding-004) so public search can rank published content by semantic similarity to a query.

Nothing AI-written publishes itself

Every AI proposal is a draft suggestion that lands in a review surface. An asset cannot be published until its creator has reviewed and approved its details. The publish API enforces this on the server: an unreviewed asset is rejected with an explicit review-required response. The creator’s approved version is the published record.

One exception, and it is most of what is on the site today. The archive we imported on 4 August 2026 was published in bulk as part of the import, without passing through that review step: as of 2 September 2026 that covers 9,393 of the 9,499 published files. The rule above has held for every file published since. If one of the imported files is yours and its details are wrong, you can edit or unpublish it from your account, and we will correct it on request.

Creator controls

A creator can turn off AI suggestions for their account entirely; the pipeline then returns nothing for their uploads, and from then on their uploads are not sent for AI analysis, with one exception: safety scanning of files uploaded to a newsroom continues regardless of this setting. If we cannot read the setting, the upload pipeline does not send their uploads. The same setting stops profile suggestions drafted from a document or text they supply, and a search summary that includes one of their files or stories is written from the results’ basic details, without AI. Proposals can be dismissed, overridden, or accepted item by item, and those actions are recorded.

Profile research

Separately from the upload pipeline, profile research drafts a creator’s profile from their published work. It has been paused since 3 October 2026: it cannot be requested, and none of the services named here receives anything from us. When it runs and you request it, we send your name as a search query to Serper, and Firecrawl reads any personal site address you give us and some of the articles that search finds under your name. Your name is also looked up in Wikidata’s public query service and Bluesky’s public account search and, when the results mention an award, prize or fellowship, on the public websites of the Pulitzer Prizes, the International Women’s Media Foundation and the International Center for Journalists. We use the results to draft profile text (a biography, coverage areas, outlets you have been published in) as a proposal.

This runs only when you ask for it, and no more than once every 30 days. Like every other proposal, nothing it drafts appears on your profile until you have reviewed and accepted it. The proposal and the actions taken on it are recorded. These services receive your name and web addresses to read; they do not receive your files.

Where processing happens

Each creator’s content is processed by Google Cloud inside a fixed geographic boundary: content from EU creators is processed within the European Union, and content from US creators within the United States. When a creator’s region cannot be determined, the system defaults to the European Union.

Frontfiles uses the Vertex AI path, to which Google’s commitment that customer data is not used to train Google’s models applies. Entity-name resolution (Knowledge Graph) sends only public entity-name strings produced by in-region detection to a global endpoint; it never sends content bytes or personal data.

Under review: the precise Google Cloud terms citation for the no-training commitment.

People shown in uploaded content

Uploaded editorial imagery may depict identifiable people. The current pipeline produces descriptive metadata such as captions, labels, detected text, and landmarks. It performs no face recognition and no identity matching, and the cross-border web-identity lookup that could surface where an image appears online is switched off (see the next section).

Under review: the lawful basis for processing imagery that depicts third parties, including the journalistic exemption analysis.

Explicitly switched off today

• Tier-C cross-border grounding (web, landmark, and logo lookups at a global Google endpoint, which would send image bytes outside the creator’s region) is disabled: the feature flag is off, and all previously granted opt-ins have been revoked with audited revocation records. If this is ever enabled, it will be per-creator opt-in behind a renewed disclosure.
• SynthID and AI-content detection are not part of the pipeline; no public API is available to us.
• C2PA manifest signing on downloads is implemented but disabled.
• No generative AI feature writes user-facing content anywhere outside the proposal and review pipelines described above, and in both of them a person reviews and accepts the draft before anything is published.

Cost, logging, and auditability

Every AI call is audit-logged with provider, purpose, cost accounting, and cache hits, behind circuit breakers and per-creator cost caps. AI never makes publish or visibility decisions; the only state it touches is the proposals table the creator reviews.

Questions

Under review: the dedicated contact channel for AI processing questions, and the EU AI Act screening for this use case.

Until that is settled, questions about AI processing reach us through the contact page and we will route them.