In July 2026, our transcription got slower. We want to explain why we thought that was worth it.

dembrane used to run two parallel processes to make sure everyone in the room got heard: AssemblyAI would create a verbatim transcript that Gemini would check. We're now doing our EU AI Act compliance run for next year, and AssemblyAI's terms require customers not to use the service in any way that could cause it, or a system it's integrated with, to be classified as 'high-risk' under the EU AI Act. Therefore we cannot include it in our pipeline anymore.

We were presented with a choice: follow the industry's trend of offloading the responsibility of compliance onto users, or try a different way of doing things: taking the time and effort to navigate these important compliance standards to make sure you can trust dembrane.

You might ask: why would I use dembrane in a "high risk" setting? Well, because "high risk" (as defined by Article 6 of the EU AI Act) includes education and vocational training, healthcare eligibility and emergency response, migration and other public issues and, most importantly, democratic processes. All things that are vital to active citizen participation, and that we want to actively support in our mission to make democracy more accessible.

We are building the next phase of dembrane around the EU AI Act, not just shallowly adopting it. That is why we are helping our customers with risk-management training on how to use our tool for all of their use cases while staying compliant. Two thirds of our high-risk clients (and counting!) have already received this training.

If you run participation, education, or care processes, "high risk" isn't a reason to avoid AI: it's a reason to be careful about which models to use and how.

The fine print problem

Do you know if you have to re-opt out of data training every time a vendor updates their terms of service? Neither did we. And that's the problem.

One other reason we no longer felt comfortable using AssemblyAI in our pipeline is their lack of transparency on whether we need to renegotiate custom agreements, and opt out of our (your) data being used, every time their terms of service are updated and we have to (re-)agree.

What this highlights is, unfortunately, an industry standard: patterns left behind the curtains on purpose. There is little transparency on what terms of service actually imply for users, and these legal terms change so often it is almost impossible to keep up.

Your conversations shouldn't be governed by fine print that shifts under your feet. When you run a citizens' assembly or a stakeholder session, you should not need a legal team on retainer to know where the audio goes.

What we're building

In practical terms: by the end of summer we aim to ship dembrane with an open source, EU sovereign transcription model, and eventually move our whole infrastructure to our own EU-hosted models. (PS: if you are an AI ops engineer and share our vision of a more democratic future, we are hiring!)