Common data licensing mistakes in generative AI training agreements

The UK High Court rejected Getty Images’ primary copyright claims against Stability AI because the training of Stable Diffusion occurred outside the United Kingdom. Getty argued that Stable Diffusion contains copies of its images, which makes its availability in the UK an instance of importing infringing copies. However, the Court held that AI model weights do not constitute a copy of images under the Copyright, Designs and Patents Act 1988. The weights consist of statistically trained parameters rather than stored reconstructions or copies of photographs. This decision shows that training an AI model on protected images outside the UK limits the ability of rights-holders to pursue primary copyright infringement claims in UK courts. The Court also found limited and historic trade mark infringement because earlier versions of Stable Diffusion generated images with Getty or iStock watermarks. This finding applies if such outputs are generated in the course of trade and cause consumer confusion.

The legal liability of using open source models stems from the fact that 76% of analyzed models contain training datasets with licenses that are incompatible with the model’s own declared license in the ecosystem. An assessment of 30 open source AI models from Hugging Face and 230 associated training datasets showed that 76% of models have at least one training dataset with a license that is incompatible with or considerably more restrictive than the model’s own declared license. One in four training datasets either has no recorded license or contains no concrete information regarding its origin. These models have served over 381 million downloads under these conditions.

License risks exist at three distinct levels. Level 1 involves models trained on unlicensed or unknown datasets, which accounts for 25% of all training datasets examined. In 15% of cases, the dataset has no license declaration, meaning all rights are reserved. Level 2 involves datasets with Creative Commons noncommercial restrictions or research-only limitations. While 6.5% of datasets have explicit noncommercial restrictions, these datasets end up in models released under licenses like Apache 2.0 that allow commercial use. Level 3 involves copyleft licenses like GPL or Creative Commons ShareAlike. If a court determines that a model is a derivative work, a company might be legally required to release its proprietary model as open source.

The ownership of an AI system involves three separate legal questions: the model weights, the training data, and the outputs. For developers, an offer letter does not constitute a legal IP assignment for model weights or checkpoints. A contractor agreement without an explicit work-for-hire provision transfers no rights to the company. Most AI startups fine-tune foundation models from providers such as OpenAI or Google, which means the provider retains ownership of the base model weights. In these cases, the developer only owns the fine-tuning layer and the custom weights derived from that process.

Training data carries different risks depending on its source. Web-scraped data carries high risk because scraping copyrighted content may constitute infringement. Licensed training data is cleaner if the license covers AI training and sublicensing. Using customer data to train models without authorization triggers contract breaches and privacy violations. Synthetic data generated by third-party AI systems is subject to the terms of service of the source provider. OpenAI’s terms, for example, have restricted using ChatGPT outputs to train competing models.

Runway provides different rights based on the subscription tier. Paid subscribers own their video outputs and can use them commercially. Free plan users are limited to personal, non-commercial use. Runway uses inputs and outputs to train its models by default for all non-enterprise tiers. Enterprise customers negotiate separate terms that exclude their data from training.

Runway prohibits the impersonation of individuals without consent and the generation of real people’s faces or voices without permission. The terms also ban content depicting minors under 18 or child safety violations. You should pull up the model card and look at the training datasets to see how they handle these restrictions.

The US court in the Northern District of California denied Stability AI’s motion to dismiss Lanlam Act claims. Getty Images alleged that Stability AI’s outputs produce distorted versions of the Getty or iStock watermarks. The court found that reports of users encountering AI generated images bearing distorted Getty watermarks were sufficient to support a reasonable inference of likely consumer confusion. These outputs could mislead consumers about the image source or affiliation if they are used in the course of trade. A similar dispute involves the Fastcase v Alexi case in the District of Columbia. Fastcase sued Alexi in November 2025 because Alexi used a legal research database to train a commercial generative AI product. The 2021 agreement between the parties restricted use to internal research purposes and prohibited commercial use or publication.

The machine unlearning problem creates a gap between what an AI system learns and what an organization can verifiably demonstrate regarding the rights used for that learning. Current unlearning techniques do not reliably remove learned patterns. They only suppress the surface expression of knowledge while leaving underlying representations intact. The EU AI Act, which entered phased enforcement in August 2025, requires providers of general-purpose models to publish training data summaries and demonstrate copyright compliance. Noncompliance can result in fines of up to €15 million or 3% of global annual revenue. California’s AI Training Data Transparency Act took effect on January 1, 2026, requiring developers to disclose training data sources. Can an organization ever truly prove that a model has forgotten a specific piece of licensed data?

AI contracts require a different structure than traditional SaaS agreements. Traditional agreements assume deterministic behavior, but AI outputs are probabilistic. Enterprises must negotiate the narrowness of data use clauses to ensure that confidential information does not support vendor product development. Contracts should also define output ownership, as purely AI-generated content often lacks human authorship. Performance obligations should include hallucination rates, factual accuracy, acceptance and rejection rates, fairness indicators, false positives, false negatives, latency, and drift thresholds.

The verdict on current AI licensing is that the risk resides in the data, not the model license. Organizations must build a Software Bill of Materials for every model in production. This document must list the training datasets, their license status, and the date of verification. If you are acquiring a company that uses AI models, you must run an AI model audit before closing. This audit involves checking training data lineage rather than just open source packages. Organizations must budget for remediation costs, such as retraining or replacing a model, if they discover incompatible licensing.