EU AI Act
The EU AI Act, for small AI products
Most AI Act coverage is written for lawyers or for Big Tech. These pages are for the rest of us: solo founders and small teams who build on an LLM API and want to know what actually applies to them. We went through this for our own product — everything here comes from doing it, not from summarizing press releases.
Two things to know up front. First: if your product uses AI at all, some of this law already applies to you — the transparency obligations in Article 50 became enforceable on 2 August 2026, and the AI-literacy duty in Article 4 has applied since February 2025. Second: for most small products the obligations are genuinely manageable — disclosure copy, a marker on generated content, a bit of documentation. The scary parts of the Act (high-risk systems, conformity assessments, CE marking) target a narrow set of use cases most indie products never touch.
Self-check: which obligations apply to you?
Eight questions about your product — do you build on a model API, do users talk to it, does it generate content, do you publish the output — and you get the obligations that likely apply, in plain language.
Article 50 transparency obligations, explained
The part of the Act that applies to almost every AI product since 2 August 2026: telling users they interact with AI, marking generated content machine-readably, and disclosing AI text you publish. What each paragraph requires and what the exceptions actually cover.
How to mark AI content, modality by modality
Concrete marking options for text, images, audio and video: visible labels, email headers, C2PA Content Credentials, IPTC tags, watermarks and platform flags — ordered from ship-this-afternoon to needs-your-model-provider.
Provider or deployer? Roles when you build on an LLM API
Calling the Claude or GPT API does not make the model vendor responsible for your product. Why you become a downstream provider, what stays with the model provider, and how repurposing shifts responsibility under Article 25.
EU AI Act deadlines: what applies when
AI literacy since February 2025, transparency since August 2026, marking grandfathered to December 2026, high-risk postponed to December 2027 by the AI Omnibus. The dates that matter, with the ones people keep getting wrong.
Case study: how we made this product compliant
What we actually changed on deslopify.me for Article 50: disclosure copy at every AI touchpoint, a machine-readable marker on AI-assisted email, a watermarking feasibility assessment, and a compliance file. With the reasoning, not just the diff.
Compliance consulting
We help small AI products through the same process: classification, disclosure UX, marking feasibility, and the compliance file that documents it. Technical consulting from practitioner experience — not legal advice.
Why a slop-score site writes about the AI Act
Because we had to comply ourselves. deslopify.me scores text with an LLM judge, rewrites text with one, and drafts email with one — three AI systems under one small roof, each touched by Article 50. Working through classification, disclosure, and marking for our own product left us with exactly the map we wished had existed when we started. These pages are that map. Where a question is genuinely unsettled — and some are — we say so instead of pretending.
Not legal advice
These pages and the self-check share practitioner experience from making our own product compliant. They are technical orientation, not legal advice — for legal questions about your specific situation, talk to a lawyer.