Running a Tintin model locally on your laptop might just transform copyright

[Georgia Jenkins]

Home / AI model marketplaces / Artificial Intelligence / copyright exceptions and limitations / copyright law / cultural heritage / dialogue / digital platforms / expressive / Georgia Jenkins / open source / UGC / Running a Tintin model locally on your laptop might just transform copyright

Running a Tintin model locally on your laptop might just transform copyright

 Georgia Jenkins Monday, June 01, 2026 - AI model marketplaces, Artificial Intelligence, copyright exceptions and limitations, copyright law, cultural heritage, dialogue, digital platforms, expressive, Georgia Jenkins, open source, UGC

When the Kat model runs locally Photo by Rita Daisy on Unsplash

Quietly, we are witnessing a new phase of artificial intelligence (AI) user-generated content (UGC). Users no longer need to access AI models through cloud services like ChatGPT, Claude, Gemini or Stability AI; instead, they can run open source models locally on their own devices. This story has developed dramatically over the last few months as new versions of Chinese open-weight models from Qwen, DeepSeek, GLM and MiniMax have disrupted open source AI model marketplaces (e.g. HuggingFace) and ‘surpassed US models in total downloads’. Alongside the publication of model weights and huge efficiency gains, these models are more effectively designed for local use.

Now, users can ‘download, run, study and modify them’ according to their own rules.

 

Platform UGC vs ‘AI model UGC’

This is a significant shift from traditional UGC where platforms were used to disseminate content and dialogue. The acronym we all want to forget in the EU, OCSSPs (online content-sharing service providers), pursuant to Article 17 of the DSM Directive, was an (unhelpful) way of identifying and recognising the value that platforms play (and extract) by moving beyond a mere intermediary role for hosting (and incentivising) potentially infringing copyright works. Fraught with uncertainty over user rights safeguards (e.g. COMMUNIA’s post-DSM Copyright Report here), this Kat suspects that we may encounter familiar territory when applying these principles to open source AI model marketplaces, and perhaps beyond.

HuggingFace is illustrative. It is an AI model marketplace based on open source software principles (e.g. openness and collaboration) where users share training datasets, models, and tools for AI projects with space for communication and dialogue. Often these communities converge on niche forms of UGC where users interact with the intent to share knowledge, resources and feedback. This is possible despite the resource intensive nature of fine-tuning large image generation because model marketplaces offer more efficient tools, primarily Low-Rank Adaptation (LoRA) models that can train foundation models (e.g. Stable Diffusion). When combined with a specific dataset, these tools alter foundational models to generate specific content using far less compute. LoRA has already been used to fine-tune Stable Diffusion on niche parody-esque examples (e.g. Pokémon).

During this Kat’s recent trip to Belgium for the KU Leuven CiTiP’s IP and AI Conference, she highlighted HuggingFace spaces dedicated to Tintin where fine-tuned versions of Stable Diffusion XL can generate content in the style of Tintin. These spaces then list Qwen models helpful for using this content to create sequence frames and dialogue. Though some noted the Hergé Estate’s characteristic knee-jerk reaction to any creative reuse of Tintin (e.g. here, here, and more recently here in relation to the Angoulême International Comics Festival - IPKat here), others have termed these spaces as ‘AI Generated Content (AIGC) social platforms’. As users ‘showcase their creations, participate in discussions, and receive feedback, […] creating a sense of community’, this takes form as creative infrastructure, moving beyond the traditional copyright lens of dissemination on platforms and even, the simple provision of tools. One would hope that copyright can adapt to this changing function without overly restricting freedom of expression and creativity more broadly.

 

Open weight models 2.0 run locally challenge traditional communication concepts

So, what has changed in the last few months? While users already download models and run them locally, new versions of Chinese open weight models are a masterclass of algorithmic efficiency. DeepSeek made headlines last year when it was reported that they spent $6million to train their V3 model compared to OpenAI’s $100million through architecture optimization and by activating a small portion of the model’s parameters for each task. For users running these models locally, they no longer need to pay per request, nor is their data shared with AI companies, and by extension, rightsholders through Article 53(c) of the AI Act. It also likely removes guardrails (e.g. refusals to generate, evading responses and harmful keyword filtering) which reduce rightsholders’ enforcement window against the reproduction and communication to the public of allegedly infringing AI-generated content.

This enforcement gap makes the holistic interpretation discussion during the C-250/25 Like Company (IPKat here) CJEU hearing even more consequential. Such an approach assumes there is a reproduction during input stage (e.g. data collection and preparation, and training AI models) when determining infringement of output. Google referred to copyright’s fundamental characterisation as a bundle of rights alongside international tort law principles, to rebut this interpretation. It could be said that assuming infringement at input stage risks a much-needed assessment of whether acts related to training AI models are captured by copyright. Further, it could result in EU copyright law stretching far beyond its borders to jurisdictions where training likely occurred. Similar arguments have been made about Article 2 of the AI Act and were avoided by the High Court of England and Wales in Getty Images v Stability AI (IPKat here) in relation to Getty's primary claim of infringement.

Running open source models locally also means that users can adjust settings like model temperature which ‘controls the randomness of an LLM’s outputs’ through APIs. Again, during the Like Company hearing, Google argued, inter alia, that the probabilistic character of the system and the presence of hallucinations, meant that there could be no reproduction. Somewhat akin to a ‘non-copyright’ use argument, randomness placed outputs outside the copyright system entirely. In contrast, when users run models locally, they can deliberately increase the model’s temperature to generate potentially more parodic or transformative output.

 
Comment

The spectrum of engagement with models clearly challenges the copyright system. A system, in this Kat’s view, which, until recently, was oriented around individual acts of copying, with platforms cast as new points of interference to bridge technological enforcement gaps. However, UGC occurring on AI model marketplaces demands a shift in rhetoric and approach. Their ability to redistribute creative agency and control over the tools of cultural production should prompt us to reflect on how copyright law should respond to creativity that occurs through shared infrastructures.

Cases like Pelham I (IPkat here), Pelham II (IPKat here) and Mio/konektra (IPKat here) illustrate the growing tendency to view the boundaries of copyright through the lens of communication. Infringement depends on recognisable elements, for both phonograms and works of applied art, and creative-based exceptions centre upon dialogue. It could be said that the future trajectory of copyright depends on the dialogue between relevant interests, and to the extent that they do not, perhaps copyright’s individually focused boundaries should shrink. Instead, the collective value of creativity that benefits non-expressive uses (training AI models or downstream monetisation of generated AI content) should be recognised and remunerated through an area of law more suited to valuing collective expressive, namely cultural heritage.

At the very beginning of this Kat’s short trip to Belgium, her all-time favourite Belgian took her to ‘Tintin’s Imaginary Museum’ in Spa, an exhibition that brings to life the first major exhibition of Hergé’s work (in 19179) where Hergé had designed the poster. Amongst the mythical Tintin objects like the moon rocket, this exhibition included quotes from Hergé’s that hint how he may have viewed his own dialogue as an artist:

Should others take over Tintin, they might do it better than me, or maybe not. One thing is certain, they would do it differently and, at the same time, it wouldn’t be Tintin anymore.

This Kat ponders whether copyright can accept these differences without eroding the very dialogue that grounds it – thoughts that she notes in her forthcoming book Creative Reuse.

Do you want to reuse the IPKat content? Please refer to our 'Policies' section. If you have any queries or requests for permission, please get in touch with the IPKat team.