AI transparency
How this platform was built
Projection-X was built with AI assistance. It says so here because it is true and because we think it is worth saying — there is no legal duty to. This page explains what the AI did, what a human decided, how it was checked, and what the check does not cover.
What was built with AI assistance
- The source code of the platform and the PX Hub — pages, data model, payment integration, licence logic — was largely written together with an AI coding assistant.
- So were parts of the documentation and the test catalogues that keep checking the code.
- Course content, effects, plugins and the shows you see come from real work at real events. They are not AI-generated.
What the platform does at runtime
- The platform runs no AI system. There is no chat, no assistant, no automatic text generation and no AI-assisted search.
- Measured on 18 Sept 2026 across dependencies, source code, outbound calls, environment variables and the content security policy: no AI provider in the production build, no AI key, no AI host.
- One exception we found ourselves and therefore name: when a product artefact is uploaded in the admin area, the platform asks VirusTotal. That service uses machine learning. It classifies a file — it talks to nobody and generates no content.
- If that changes, it will say so here before the feature goes live. A check in the development pipeline fires as soon as an AI service enters the production build.
How it was checked
- Internal white-box review by the operator, AI-assisted, on 12 Sept 2026, against 298 self-defined checkpoints drawn from OWASP ASVS, GDPR, TDDDG, German civil and consumer law, PAngV, UWG, DSA and BFSG. No review by an independent body, no certification.
- That produced 107 confirmed findings; a further 30 failed a counter-check and are therefore not recorded as defects.
- As of 3 Oct 2026: 87 of the 107 findings are closed, 20 are not — none of them in the highest severity band. For each open one, the file records the route by which the problem still occurs today. 6 findings sit with the operator because they need access, contracts or judgement calls that no code replaces; plus 10 operator tasks.
- 5398 automated checks run before every release; 5394 of them green, 4 left open as an explicit reminder. Each one covers every comparable place rather than one example — the most common defect in this review was a rule applied at one of several identical spots.
- Text contrast was measured against the live platform on 19 Sept 2026: 1164 text nodes across twelve pages, none below the WCAG 1.4.3 threshold.
What a human decides
- Prices, products, course content, legal texts and every publication are released by the operator. Editorial responsibility rests with him, not with a tool.
- Every change to the source code is readable in the change history, with its reasoning.
- The 107 remediation claims of 12 Sept 2026 were self-reported. On 18 Sept 2026 each one was reviewed in three rounds: refute it, decide it neutrally at the code, check it from two sides. 40 did not hold; they were worked through, and each fix was itself attacked three times. One more has been closed since; as of 3 Oct 2026: 87 closed, 20 not.
On the law — plainly
No rule requires labelling a website because it was developed with AI. Article 50 of the AI Act (EU) 2024/1689 attaches to two other things: operating an AI system, and publishing AI-generated content. Neither applies here.
The duty there also falls on the provider of the AI system — not on whoever uses an AI tool. The Commission's guidance expressly excludes source code from the marking requirement.
This page exists anyway. Anyone who wants to know how something they buy came about should be able to read it — even when nobody has to ask. The detailed assessment with citations, sources and the counter-check is in the repository at docs/KI-TRANSPARENZ.md.
What is open
- The legal assessment above is carefully sourced but is not a legal opinion. A lawyer review is still pending, together with 99 further legal questions from the 12 Sept 2026 review.
- No commissioned penetration test has taken place. The permitted scope is ready.
- Accessibility is checked statically and measured for contrast. A review with screen reader, keyboard and real users is pending.
- Whether any delivered images are AI-generated is currently being inventoried. The result will appear here.
This page describes the technical state at the date given. It is not an agreement on characteristics, not an assurance and not a guarantee within the meaning of § 443 BGB. The contract and the terms and conditions govern the scope of services. Changes are tracked here.
Questions?
If something here is unclear or does not match what you see: write to us. Answering questions about this page is part of the disclosure.
As of 3 October 2026 · Reviewed 12 September 2026 · Counter-reviewed 18 September 2026