AI Transparency Notice – GardenPilot

Version: ai-transparency-2026-09-v2
Effective: upon publication

GardenPilot uses artificial intelligence for significant parts of garden analysis, assistance, text and image generation and quality control.

This page supplements the notices shown directly in the relevant AI functions.

1. Interaction with AI

When you communicate with a GardenPilot assistant or chat function, the response may be generated wholly or predominantly by an AI system.

GardenPilot indicates in such functions that an interaction with an AI system is taking place, unless this is already obvious from the circumstances and context of use.

2. AI-generated garden visualisations

GardenPilot generates and/or modifies images using generative AI.

A GardenPilot visualisation is therefore not a photograph of a garden that has actually been built, but an artificially generated or AI-manipulated depiction of a possible design.

Where a visualisation is based on a real uploaded garden photograph, real parts of the source image may be combined with artificially generated or modified elements.

3. Transparency and technical provenance labels

GardenPilot provides disclosure of artificial generation or manipulation for AI-generated or AI-manipulated content in accordance with applicable transparency requirements, in particular the EU AI Act.

Depending on the model and technical generation route, output files may additionally contain technical provenance information, in particular:

  • machine-readable metadata,
  • Content Credentials / C2PA provenance information,
  • IPTC/XMP information,
  • invisible watermarks or provenance signals of a model provider, for example SynthID.

Such technical labels may originate from the relevant model provider or from the technical processing chain. Not every generation route uses or preserves the same labelling mechanisms.

4. Loss of technical metadata

Metadata or Content Credentials may be removed by third-party platforms, messaging services, social-media services, screenshots, file conversion or other image processing.

Such technical loss does not turn the underlying image into a genuine photograph.

5. Use by customers

Please do not deceptively present GardenPilot AI-generated visualisations to third parties as photographs of gardens that have actually been completed.

If you publish images or use them for your own business, journalistic or other purposes, additional statutory transparency or labelling obligations may apply.

6. Quality limitations

AI systems can make mistakes. Garden visualisations may in particular depict plants, materials, dimensions or construction details inaccurately.

The visualisations are design aids and are not building-law, structural or technical execution plans.

7. Contact

Questions about GardenPilot's use of AI and labelling may be sent to support@gardenpilot.de.

Verify the AI labeling of our images

Practical verification guide — not a contractual or legal text. It complements the transparency notice above (version 14.09.2026) by explaining how the labeling can be checked with freely available tools.

All garden visualizations created with GardenPilot are AI-generated. They are produced with image models by Google (Gemini) and — as a fallback path — by OpenAI. The sections below explain how our images are labeled and how anyone can verify that with freely available tools — no account needed, no access to our systems required.

Legal background: Article 50(2) of the EU AI Act (Regulation (EU) 2024/1689) requires AI-generated content to be marked in a machine-readable format and detectable as artificially generated.

How our images are labeled

Layer 1 — Signed C2PA manifest (Content Credentials)

The Google image models we use deliver their outputs with a cryptographically signed C2PA manifest; according to our measurements, the OpenAI fallback path likewise delivers such a manifest. It contains the machine-readable statement digitalSourceType = trainedAlgorithmicMedia (“created by an AI model”), a digital signature by the model provider (e.g. “Google LLC”) and a trusted timestamp. GardenPilot stores the model output byte-identically so this signature remains valid. Any subsequent modification of the file would break the signature and become visible during verification.

Layer 2 — Invisible watermark (SynthID)

Images from Google’s models additionally carry SynthID, a watermark embedded invisibly in the image pixels. The manifest from layer 1 documents this with the assertion “Applied imperceptible SynthID watermark”. Since GardenPilot stores the primary file byte-identically, the watermark is preserved there.

Layer 3 — Open metadata (IPTC/XMP) as fallback

Where an output carries no signed manifest — for example the downscaled WebP display derivative, which must be re-encoded for display — GardenPilot writes open IPTC/XMP standard metadata with the same value trainedAlgorithmicMedia, a description as an AI-generated garden visualization, and the model name.

In addition: visible notices

Wherever images leave the product context, we disclose the AI generation visibly — for example in shared project views (“The garden images are AI-generated visualizations”) and in the PDF project folder (notices about the AI origin of the visualizations).

How to verify an image — step by step

1. Use the original file

Verify the downloaded original file, not a screenshot. Screenshots and copies re-encoded by platforms no longer contain the metadata layers (see “Limits of verification”).

2. Check Content Credentials in the browser

Open contentcredentials.org/verify (a free verification tool by the Content Authenticity Initiative) and drag the image file into the window. For a GardenPilot original, the tool shows a validly signed manifest by the model provider (issuer e.g. “Google LLC”) and identifies the content as AI-generated. The exact presentation depends on the tool version.

3. Alternatively: command line

The open-source tool c2patool shows the same manifest locally: c2patool image.jpg. The output includes the signature status, the issuer and digitalSourceType.

4. Fallback layer: read the metadata

Files without a signed manifest (e.g. WebP display derivatives) can be checked with any metadata viewer, such as ExifTool: the DigitalSourceType field contains the IPTC value …/trainedAlgorithmicMedia.

A note on SynthID

For the invisible SynthID watermark, verification is offered by Google (for instance via the Gemini app); a generally accessible detector is currently not available. The publicly reproducible verification path for our images is therefore the C2PA manifest (steps 2–3).

Limits of verification — stated honestly

  • Many platforms and messengers re-encode images on upload and strip metadata in the process — including the C2PA manifest. A missing manifest therefore does not prove that an image is not AI-generated or not from GardenPilot. It only means this verification path is no longer available for the copy at hand.
  • Older images carry weaker labeling or none at all: depending on the generation path, images created before 24 July 2026 may carry no machine-readable labeling (retroactive labeling is not legally required), and images created between 24 and 28 July 2026 may carry only the open XMP labeling (layer 3).
  • For some detail images generated between 24 and 28 July 2026, verification may report an invalid signature. This was caused by an error in our processing that was fixed on 28 July 2026; the AI labeling itself was present throughout.
  • Screenshots contain none of the metadata layers. According to Google, the SynthID watermark in the pixels is more robust against common edits, but its verification lies with Google (see above).
  • Verification confirms the AI generation and the model provider. On its own it does not confirm that an image was created via GardenPilot — for individual cases, use the contact below.

Questions about verification

If you want to verify a specific image or have questions about the labeling, you can reach us at support@gardenpilot.de or via the contact form.

Verification guide last reviewed: 28 July 2026 · It is updated whenever the labeling changes. The transparency notice above carries its own version date (14.09.2026) — the two move independently.