Skip to main content

AI Features and the EU AI Act

Last Updated: 25 August 2026

RetroTechCollector uses AI in three places. This page explains what each one does, what it is allowed to decide, what data reaches the model provider, and how our obligations under Regulation (EU) 2024/1689 (the EU AI Act) apply.

We would rather be specific than reassuring. Where a limitation exists, it is stated.

The three features​

1. Photo identification​

Point your camera at an item and we suggest what it is.

Your photo is downscaled on our servers and sent to Google Gemini on Vertex AI. The model returns a structured description — a likely name, item type, model number, plausible lookalikes and search terms. It does not choose a catalogue entry. It cannot: the model has no access to our catalogue, so letting it name a row would invite invented matches. We take its description, match it against our own catalogue with our own scorer, and show you ranked candidates to confirm or reject.

Nothing is added to your collection until you pick one.

2. Condition grading​

Photograph a flaw and we suggest a cosmetic condition, a functional status and a list of defects.

This is advisory only. The suggestion pre-fills the grade fields on your listing. You review it, change it, or ignore it. The listing carries the grade you publish, not the model's.

3. Description drafting​

We can draft a listing description from information you have already entered — title, category, condition, functional status, grade and any flaw notes. This one is text-only; no image is sent. The draft is a starting point you edit before publishing.

Our position under the EU AI Act​

These are transparency-tier systems, and here is why​

The Act reserves its heaviest obligations for high-risk systems — those listed in Annex III, such as systems used for employment decisions, access to essential services, creditworthiness, law enforcement or biometric identification. None of our features fall in those categories.

More importantly, the argument does not rest on category alone. What makes a system high-risk in practice is that it decides something consequential about a person. Ours decide nothing:

  • Identification produces candidates you choose between. If you reject them all, nothing happens.
  • Condition grading pre-fills a field you own and can overwrite. What a buyer sees is what the seller published.
  • Description drafting produces text you edit before it exists publicly.

In each case a person is in the loop by design, not by policy — the feature does not have a path that reaches another user without a human decision in between. That is why we treat these as subject to the Act's transparency obligations (Article 50) rather than the high-risk regime.

What transparency means here​

  • You always know when AI is involved. These features are invoked deliberately, are labelled where they appear, and are never applied silently to your content.
  • Output is identified as AI-generated or AI-assisted at the point you are asked to accept it.
  • You can avoid AI entirely. Every one of these features has a manual path: add items by form, barcode or CSV import; set the condition yourself; write your own description. Nothing in the product requires you to use AI, and no feature is withheld if you do not.

What we do not do​

Stated plainly, because these are the practices the Act restricts most tightly:

  • No biometric identification or categorisation. We do not attempt to identify people, infer characteristics from faces, or process biometric data. These features look at retro hardware.
  • No emotion recognition.
  • No social scoring.
  • No subliminal or manipulative techniques.
  • No automated decision-making with legal or similarly significant effects. Nothing about your account, your access, your money or your standing is decided by a model. This is also our position under Article 22 UK GDPR — see GDPR.

Accuracy, and who is responsible​

These models are useful and they are wrong sometimes. Identification confuses visually similar hardware, particularly regional variants and later revisions. Condition grading sees one photograph and cannot assess anything the photograph does not show — it has no way to know whether something powers on.

So: you remain responsible for what you publish. If you accept a suggested identification, condition or description and list an item on that basis, the listing is your representation to the buyer, not ours. Check it. This matters commercially as well as legally — a misdescribed item is a dispute, and disputes are adjudicated against what the listing said.

Data sent to the model provider​

IdentificationCondition gradingDescription drafting
Image sentYes, downscaledYes, downscaledNo
Text sentSearch context—Title, category, condition, functional status, grade, flaw notes
Account identifiers sentNoNoNo

The provider is Google Cloud Vertex AI. Requests are authenticated with our own service credentials; your account identity is not transmitted as part of the prompt.

What we keep​

We store the outcome, not the input. For an identification we record the status, what the model said it was, its confidence, how many catalogue candidates we showed you, and which model version ran (for cost and quality monitoring). For a condition grade we record the status, the suggested condition and functional status, the confidence and the number of defects found.

We do not retain the photograph you submitted for AI processing. Photos you choose to attach to an item or a listing are stored as your content in the ordinary way, and are covered by the retention and deletion rules in GDPR.

Usage limits​

AI features are metered per calendar month and reset at the start of the month (UTC). The allowances by plan are listed on Plans and billing. Reaching a limit disables the AI shortcut for the rest of the month; it never blocks the manual path.

When the Act applies​

The AI Act entered into force on 1 August 2024 and applies in stages. The prohibitions on unacceptable-risk practices and the AI-literacy provisions applied from 2 February 2025. Obligations for general-purpose AI models applied from 2 August 2025. The bulk of the remaining obligations, including the transparency duties relevant to this page, apply from 2 August 2026, with certain high-risk categories phased to 2 August 2027.

We are treating the transparency obligations as live now rather than waiting.

Changes and questions​

If we add an AI feature, change what a model is permitted to decide, or change what data it receives, we will update this page and say what changed. If a future feature would cross into automated decision-making about people, we will say so here explicitly before it ships.

Questions about any of this: [email protected].