The EU's new AI transparency law just kicked in. Here's why "you approve everything" was never just a slogan.

On 2 August 2026, Article 50 of the EU AI Act became enforceable: generative AI systems now have to make AI-generated content detectable, and AI-written text on matters of public interest has to say so. If a chunk of your outreach is written by a model and shipped without anyone looking at it, this is the week that started mattering.

This isn't legal advice, we're not your lawyer and every business should get its own read on this. But here's the shape of it, and why one part of it lines up with how Selda was already built.

What the law actually asks for

Article 50 covers four things: AI systems talking directly to people, AI-generated content in general, emotion recognition, and deepfakes or AI text on public-interest topics. The part that touches outreach is the content rule: output has to be marked and detectable as artificially generated. Providers get until 2 December 2026 for the technical marking. The disclosure obligation itself is live now.

The exemption that actually matters

The law carves out content that has "undergone a process of human review or editorial control", where an identifiable person holds responsibility for what gets published. The guidance is specific about what that means: real examination of the substance, by someone with the judgement to actually evaluate it, not a spell-check and a click. A rubber stamp doesn't count. Reading it, deciding it's actually right, and putting your name on it, does.

That's a fair line to draw. The thing regulators are actually worried about isn't "a model helped write this." It's content nobody stands behind, going out at a scale no person could have reviewed.

Why this isn't a scramble for Selda

Every message Selda drafts sits in test mode until you look at it. It goes out from your own inbox, under your own name, not Selda's. That's not a compliance patch added this week, it's the whole design: Selda proposes, you decide, nothing sends without you. The "identifiable person, real review" shape the law is asking for is the shape the product already had.

It's also, separately, just the thing that gets replies. A message nobody reviewed reads like nobody reviewed it. People can tell.

How to spot a message that wasn't reviewed

The tell is the same one that flags spam: it could have been sent to anyone. A few signs a message skipped the human part:

• It compliments something generic ("impressive company") instead of naming something specific and true.
• It restates the recipient's own job title back at them as if that were insight.
• The ask is vague, "let's connect" instead of one clear, small next step.
• It would read exactly the same with the name swapped for anyone else at the company.

A reviewed message does the opposite: one true, specific observation about that company, said the way a person would say it, with a clear ask. That's what a real look at the substance produces, whether the first draft came from a person or from Selda.

What to actually do about it

If any part of your outreach goes out with nobody reading it first, that's worth fixing regardless of this law, it was already the weaker version of outreach. Read what goes out under your name. Keep it specific to the person. Send it from your own address. That was always the safer bet, and now it's also the compliant one.