The AI label is now part of the product: what August’s transparency rules mean for generative AI startups

For the past few years, the generative AI playbook for startups has been simple: build fast, integrate seamlessly, and let the AI feel like magic. As of 2 August 2026, the trick needs a label – and the label is no longer something Legal quietly adds at the end. It is something Product has to ship.

Article 50 of the EU AI Act is now in application, and it brings new obligations for anyone building generative AI. Providers of systems that interact directly with people must generally disclose that users are dealing with AI, unless that is already obvious. Providers of systems generating synthetic text, audio, image, or video must make that content detectable as artificial. Deployers face a further set of duties: disclosing deepfakes, flagging emotion-recognition and biometric-categorisation systems, and labelling certain AI-generated content published on matters of public interest.

The instinct for most founders will be to forward this to external counsel and keep shipping. That misses what has actually changed. This is not a new line item for the compliance folder. It reaches into interface design, content pipelines, metadata and, increasingly, whether an enterprise buyer trusts what you are selling.

Transparency has become a product requirement

Startups have historically treated regulation as friction: something that slows onboarding or adds unwelcome text to a clean interface. Article 50 breaks that model, because most of what it demands cannot simply sit inside a privacy notice. It has to live in the product itself. This is the product-layer version of an argument I made while writing about governance-as-code: rules that exist only on paper cannot keep pace with a system that ships weekly. Article 50 effectively agrees. Compliance here is architecture, not paperwork.

Three problems, one architecture

The obligations are split into three distinct engineering and design problems. The first is the chatbot: if you are building a sales assistant, support agent or digital adviser, disguising it as human is no longer a defensible product strategy. The disclosure does not need a flashing banner – an introductory line, a persistent badge or a clear identifier will usually do – but it has to sit inside the interaction itself, not buried in terms and conditions.

The second is harder. Synthetic content – text, audio, image or video – now needs a machine-readable marking that survives export and compression. This is not a UI fix. Depending on the product, it may mean watermarking, provenance metadata or adopting a standard such as C2PA, and it belongs in the same conversation as latency and model cost, not bolted on after launch.

The third sits with whoever publishes the output. Deployers, not just providers, carry disclosure duties for deepfakes and for AI-generated text on public-interest matters that has not had human editorial review. A startup that builds the generation technology and also publishes with it can find itself wearing both hats, which means the sensible product decision is to make the customer’s compliance easier too, not only your own.

For UK startups, Brexit is not a transparency strategy

British founders may want to file this under “Brussels problem.” That holds until the first customer in Paris. The AI Act reaches providers outside the EU wherever their systems are placed on the EU market, or their output is used within it – incorporation in London does not create a regulatory boundary around a SaaS product.

There is a domestic reason too. The UK has not copied the EU AI Act, but existing data protection law already imposes transparency and fairness duties wherever personal data is involved, and the Data (Use and Access) Act 2025 has reshaped, not removed, the accountability framework around automated decision-making. The ICO is actively building out further AI and automated-decision-making guidance, with transparency and public trust as explicit themes. The route is different. The destination is not.

See Also

What early-stage founders should know about voice tech before launching

The compliance feature is worth mentioning in your next sales call

This is where founders should stop treating transparency as overhead. Enterprise buyers increasingly want evidence that a vendor can explain what its system is doing and manage the regulatory exposure downstream – that is a procurement advantage, not a cost centre. By the end of July, around 190 organisations had already signed the European Commission’s voluntary Code of Practice on Transparency of AI-Generated Content, ahead of enforcement powers switching on. This is no longer theoretical; an implementation ecosystem already exists.

Three shifts should follow:

• Put the disclosure question into product discovery, not launch week
• Build for the customer’s compliance problem, not only your own. Native labelling tools beat a PDF explaining what they should do after export
• Stop marketing compliance apologetically. In enterprise AI, demonstrable transparency is what shortens the distance between an interesting demo and an approved vendor

The end of stealth AI

There is a limited grace period for the marking obligation on generative systems already on the market – providers have until 2 December 2026. But extra time to implement is not a reason to postpone the architecture. The startups that handle this well will not be the ones bolting an “AI-generated” badge on the night before launch. They will be the ones who built the label in from the start, because they understood the interface, the metadata and the export pipeline were never really separate from the regulation in the first place.