Is your firm's published work being used to train AI?

What parts of an architecture firm's work are publicly reachable?#

The publicly reachable part of an architecture firm's work is whatever sits on an open web page or in a public file: project pages, portfolio PDFs, renderings, award and competition boards, articles in design publications and social media posts. Planning and zoning submissions posted in public meeting packets can also include drawings.

Photography adds a wrinkle. Architectural photographers often hold copyright in their images and license them to the firm, so the photographer may have as much interest in how those images are reused as the firm does.

What stays out of reach of scrapers?#

Internal project records stay out of reach of scrapers because they are not published on the open web. RFI logs, Bluebeam markup sessions, QA/QC comments, Deltek project budgets, staffing plans and internal email live behind logins on servers and cloud applications the firm controls.

That private layer is what makes consent-based licensing distinct. A crawler can collect a rendering, but it cannot see the review comments that shaped the design, the RFI that changed a detail or the staffing decision that kept a project on schedule.

The private layer can still leak in other ways, including a public share link or a misconfigured file server. Staff pasting client drawings into consumer AI tools, or software vendors whose terms allow training on customer content, are more realistic risks to internal records than scraping. An AI use policy, a review of vendor terms and periodic checks of sharing settings address these risks.

Can you tell whether your work was used in training?#

Usually a firm cannot confirm whether a specific rendering was used in training, because most model developers do not publish item-level lists of their training data. Research efforts have tried to document dataset sources: the Data Provenance Initiative, for example, audited 44 data collections spanning more than 1,800 fine-tuning text datasets and recorded their sources and licenses. Audits like that cover published research datasets, not every commercial model.

Some public image datasets can be searched by URL or caption, and some developers publish summaries of their data sources. Treat both as partial evidence. If the question matters for a dispute, counsel can advise on what can be learned and how to preserve it.

What controls can a firm put on its published work?#

A firm can put several controls on its published work, each of which signals a preference or adds friction rather than guaranteeing compliance. Used together, they make the firm's position clear and create a record that it asserted its rights.

Rules outside the United States can matter for firms whose work is published internationally. The European Union, for example, recognizes reservations of rights against text and data mining, and the way a reservation is expressed affects whether it counts. Assess any reliance on foreign rules with counsel.

• 1Crawler rules: add robots.txt directives for AI training crawlers that publish their names, and recheck them as new crawlers appear.
• 2Site terms: state that site content may not be used for AI training without permission.
• 3Image choices: publish lower-resolution images and avoid posting full drawing sets or detailed sheets.
• 4Content Credentials: attach provenance and a training preference to images where your tools support it.
• 5Publisher agreements: ask design publications how they handle AI crawlers and reuse of submitted material.
• 6Legal review: ask counsel about registering copyright in key works and about rules in other jurisdictions.

What can Content Credentials do for renderings and photos?#

Content Credentials can attach a signed provenance record to a rendering or photo and, through a separate assertion, state whether the firm allows AI training. The C2PA specification defines a Content Credential as a cryptographically bound structure that records an asset's provenance, including assertions about origin, modifications and use of AI.

The training preference is no longer part of the core C2PA standard. Version 2.0, released in January 2024, removed the training and data mining assertion, and C2PA's AI and ML guidance now points to the Creator Assertions Working Group's assertion, which can mark an asset as allowed, constrained or not allowed for AI use. Version 2.4, from April 2026, added an AI disclosure assertion.

The limits are practical. A credential states a preference; it does not stop a crawler that ignores it, and embedded metadata can be stripped when images are copied or re-uploaded. The C2PA explainer itself notes that Content Credentials show whether provenance data is intact and signed, not whether it is true.

How does scraping compare with consent-based licensing?#

Scraping and consent-based licensing differ on every point that matters to a firm: who decides, what is included, whether client rights are checked and whether the firm is compensated. The comparison explains why a firm can object to unpermitted scraping and still consider licensing its internal records.

Illustrative: an architecture studio audits its public footprint#

Illustrative: a fictional architecture studio known for cultural buildings reviews its website after a client asks whether project images are being used to train AI. The marketing lead finds old competition boards with detailed plans, drawing excerpts in qualification PDFs and photographers' images posted at full resolution.

The studio removes the detailed plans, republishes images at lower resolution, adds AI crawler rules and site terms, and attaches Content Credentials with a training preference to new renderings. It writes to the client explaining each step and its limits.

Separately, the principals discuss whether the studio's internal review records could be licensed on its own terms. They start with a metadata-only fit check, because that layer was never public and remains entirely under the studio's control.

How SourceX approaches published and private work#

SourceX works only with consent-based licensing of internal operating records; it does not collect public portfolios, and its own rights in a deidentified dataset are set out in the signed supplier agreement. The fit check gathers metadata about systems and record families, and nothing is shared during that initial assessment.

When a firm decides to proceed, it chooses the records and approves each stage of the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery. The SourceX Evidence Packet then documents the provenance and licensing rights of what was included, the permitted use, the privacy record and the release authorization, which is precisely the record that scraping never produces.