ANNEX TO AI GUIDELINES AI Models and Systems
Information on the environmental, social, ethical, and quality standards of our AI service providers
Mistral
We have been working with our standard LLM provider, Mistral AI SAS, since 2024. Mistral AI develops various models for general-purpose queries (Mistral Vibe, also known as LeChat), as well as specialized models for software development (Devstral) and expert models (Mixtral). Mistral AI is headquartered in France and trains its models in France or Sweden, which significantly reduces their carbon footprint. Although U.S. companies (primarily Microsoft) have invested in Mistral AI, a takeover is unlikely, as compliance with EU regulations is the model provider’s stated USP. Mistral AI has not yet published any information regarding the working conditions of data workers. In addition to using it for chatbot subscriptions, we primarily use the smaller Open Rates models—which are comparatively very powerful—for energy-efficient text and image analysis tasks.
In addition, particularly to support our programming work, we have been using models from Google – which is part of Alphabet Inc., a publicly listed company based in the USA – since 2026. We use Google models in situations where the quality of code generation is more important than carbon footprint, digital sovereignty and usage costs. Compared to other AI providers, Google is relatively transparent about how its AI operates. The company has been publishing sustainability reports for 11 years, including details on the energy consumption of its main applications and its progress in optimising resource efficiency. Google operates its own data centres, predominantly powered by energy from renewable or low-carbon sources, and offsets 100 per cent of its remaining CO₂ emissions through climate protection projects. Unfortunately, Google pays less attention to the working conditions of data workers; a co-head of the Ethical AI research team was dismissed in 2020 after she raised concerns about discriminatory practices.
Anthropic
In addition, since 2025 we have continued to use models from Anthropic PBC, a (so far) public-interest organisation based in the USA. Anthropic models are used in situations where the quality of the knowledge base and language analysis are more important than carbon footprint, digital sovereignty and usage costs. Anthropic’s unique selling point (USP) is not only the high quality of its Frontier models but also its ethical standards regarding how and for what purposes AI is used. Less relevant and transparent, however, is the question of how the AI is operated. It is known that Anthropic does not operate its own data centres but instead uses those of Google, Amazon/AWS and xAI. These offset some of their emissions – Google 100 per cent, AWS 90 per cent. However, the xAI Colossus 1 data centre – which, according to media reports, is used by Anthropic for just under half of its required computing power – is operated without offsetting, using gas turbines and, in part, a local electricity mix, and is known for breaches of environmental regulations. Anthropic’s Frontier Models are among the largest in terms of parameter size. According to the company, they are comparatively energy-efficient in operation thanks to the ‘Mixture of Experts’ approach; however, Anthropic does not publish any data to substantiate this claim. Anthropic provides no information regarding the working conditions of its data workers.
Adobe
For vector graphics generation, we use Firefly, the generative AI suite from Adobe Inc., a publicly listed company headquartered in the USA. Firefly's key differentiator is its training data: unlike most competitors, Adobe trained its models only on licensed Adobe Stock content, openly licensed material, and public-domain content whose copyright has expired, which significantly reduces legal and ethical risk around the origin of training material. Adobe also leads the Content Authenticity Initiative and embeds Content Credentials (provenance metadata) into Firefly output, allowing recipients to see that and how AI was used. Firefly runs on AWS infrastructure. Adobe has been carbon-neutral across its global operations since 2020 and has set an SBTi-validated target of net-zero across its full value chain by 2050; however, its goal of 100% renewable electricity, originally due in 2025, remains unmet, with renewable coverage at roughly 74% as of the most recent report. Adobe has not published specific information on the working conditions of Adobe Stock contributors or other data workers involved in curating Firefly's training material.
Higgsfield
We use Higgsfield, a US-based company, to process images and generate videos. Higgsfield develops its own video and image models, but also routes workloads through third-party models such as those from OpenAI, Google, Alibaba, ByteDance and Kuaishou, whose environmental and social footprints are beyond Higgsfield’s control. The private company is funded by venture capitalists. With a view to responsible use, Higgsfield introduced a ‘Similarity Scoring’ tool in 2026 which, even before content is generated, flags potential matches with known individuals, celebrities or brand logos, thereby counteracting risks associated with deepfakes and the misuse of likenesses inherent in video generation. Higgsfield has not published any information on the energy supply or the carbon footprint of its computing infrastructure, nor on the origin of its training data or the working conditions of its data staff. A large part of its own computing infrastructure was recently migrated to the GMI Cloud for cost reasons; the GMI Cloud claims to adopt a ‘green-first’ approach.
Black Forest Labs
We use models from Black Forest Labs, a European AI company based in Freiburg, Germany, primarily for image generation. The company was founded in 2024 by former researchers at Stability AI (the developers of Stable Diffusion) and is developing the FLUX family of image and, increasingly, video models, which are released under a dual licensing model (open weights/commercial). Black Forest Labs is venture-capital funded, and its investors include Andreessen Horowitz, who are known for backing controversial companies; furthermore, Marc Andreessen and Ben Horowitz have been making substantial private donations to pro-Trump causes since 2024. However, no single investor holds a majority stake, and there are no signs of an imminent takeover. The company supports the C2PA standard for content attribution and tags generated images with metadata to indicate their AI origin (although this metadata can be removed). Black Forest Labs does not publish any information on the carbon footprint of the training or operation of its models, the origin of its training data, or the working conditions of the data workers involved. On BFL's trust website, however, it mentions Microsoft Azure as its cloud provider, which has been using 100% renewable energy in Europe since 2025 and is aiming for zero-waste certification by 2030.