AI Transparency at Bricks

Bricks is signing the European Commission's Code of Practice on transparency of AI-generated content, published 10 June 2026 and entering reference application alongside Article 50 of the AI Act on 2 August 2026. The Code asks providers and deployers of generative AI to mark AI-touched content in two ways: a machine-readable mark that survives reposting and recompression, and a visible mark a person can recognise at a glance.

The icon you see in the footer of every Bricks page is the visible part of that commitment. The machine-readable part lives in the page's underlying metadata and in the platform's databases. The two halves are paired: the icon says what's true, the metadata makes that truth verifiable.

Our long-form position is published as a whitepaper: Transparency by Construction.

Every Bricks-built artefact — this page, the Studio dashboard, the marketing site, every app a tenant ships through Bricks — comes out of the same three-step chain:

1. A human conceives the intent — the goal, the constraint, the editorial frame. The human owns this step. The AI does not initiate.
2. AI drafts the result — the code, the schema, the copy, the layout. The AI works against the human's intent and stops where the intent runs out.
3. A human validates — reads the draft, accepts or rejects, edits where needed, and only then ships.

The construction chain above is one half of the answer. The other half is the set of guardrails that constrain how AI behaves once an artefact is in production. These are not aspirations. They are running code.

• Plans before changes. When Studio AI proposes a schema migration, a role change, an integration credential write, or any destructive operation, you see the plan first and approve it before it applies. The AI does not act unsupervised on production artefacts.
• Auto-snapshots before destructive operations. The platform takes a tenant-wide snapshot before any approved destructive operation under heavy-reasoning mode. If anything regresses, the snapshot restores cleanly. The AI cannot make a change that can't be undone.
• Sandbox-verify on heavy-reasoning operations. Multi-step operations that would otherwise need you in the loop run in an isolated sandbox first, where the AI verifies the operation works before applying it to your real environment. You see one approval, not twelve.
• No training on your data. Bricks resells inference from a curated catalog of AI providers under per-provider Data Processing Agreements that prohibit training on customer prompts and outputs. Your tenant data does not become anyone's training data.
• Curated provider catalog. You pick the AI provider, the hosting region, and the jurisdiction from a small audited list — not from an open BYOK field. The criteria for inclusion in that list are documented in our Curated Pluralism whitepaper.
• Sanitised AI output. AI responses are post-processed to strip references to external LLMs, API-key-handling instructions, and similar foot-guns before being shown to you. The AI cannot, by accident or design, route you toward a competing provider or leak a credential into your chat.
• Transparent billing. Every AI turn writes to your wallet ledger with the model used, the token counts, and the price. No subscription veil; no surprise bills; the cost is visible per message and per turn.
• You own the conversation history. Your Studio AI chat lives in your tenant schema. Export it, delete it, audit it — at any time, without contacting us.

Several things follow from this position that are worth saying explicitly rather than letting them be inferred.

• We do not generate synthetic faces, voices, or identity media. The deepfake provisions of Article 50(4) therefore do not apply to Bricks today. This is a scope choice, not a clever workaround. We do not intend to enter the synthetic-identity-media market.
• We do not personalise pricing, access, or content delivery based on profiling. The wallet ledger is identical across users. Role assignments are human-defined. The AI behaviour panel exposes model, jurisdiction, and cost discipline to the operator before any AI turn runs. None of these involve profiling.
• We do not manipulate AI prompts to nudge you toward upsells. Bricks's system prompts are auditable through the platform's source tree. They contain no commercial-bias instructions. We sell additional capacity through the wallet, not through coercion at the model layer.
• We do not run silent AI on your data outside the surfaces you can see. If a feature uses AI, the AI is visible — through the chat, through a labelled tool call, or through a marked workflow step. There is no background AI process touching your data.

The European Commission publishes three icons. Bricks uses all three, each one with a specific scope.

AI Modified — in footers

The artefact passed through the human-AI-human construction chain. You'll see it in the footer of every Bricks page and every tenant app built with Bricks.

AI Generated — inline next to specific content

The marked content was produced by AI and shipped without an intermediate human edit. You'll see this next to AI-generated images and ungated AI text — never as a blanket page mark.

AI — on interactive AI surfaces

You're about to talk to an AI. You'll see this on the Studio AI chat surface, accompanied by a textual disclosure at the start of every conversation.