Hidden Licensing Risks in Training AI Models with Outputs from Open Models
I. Introduction
In recent years, businesses have increasingly deployed AI models within their own environments in response to growing concerns over data security and privacy.[1] In particular, large language models (“LLMs”) operated within a company’s own infrastructure (commonly referred to as “local LLMs”) have emerged as a practical option for implementing generative AI in environments that handle highly sensitive information, such as those in the financial sector.
As the use of local LLMs continues to expand, businesses may increasingly seek to improve model performance by using the outputs of existing AI models as training data for other models.
In practice, however, when introducing an AI model, businesses do not always evaluate the model’s licensing terms with the future use case of training another AI model using its outputs specifically in mind. Nevertheless, if it is later discovered that such use violates the applicable license terms, the consequences may vary depending on both the terms governing the AI model that generated the training data and the manner in which the subsequently trained model is deployed. In some cases, the business may be unable to continue using the relevant training data and may be required to retrain the model from scratch. In others, achieving compliance may require amendments to the business’s own terms of use and renewed user consent after the relevant service has already been launched. Such remedial measures can impose significant operational and legal burdens.
This article examines the licensing terms applicable to several major open models and considers the obligations that may arise where their outputs are used to train another AI model. It then discusses practical measures that businesses should consider during both the implementation and operational phases to appropriately manage AI model outputs and mitigate the associated legal risks.
To continue reading, please download the full article in PDF format below.
[1] Examples include the following:
- “MUFG deploys real-time anonymization across unstructured data,” Asian Banking & Finance, February 2026
- “BNP Paribas provides its businesses with an LLM as a Service platform to accelerate the industrialization of generative AI use cases.” BNP Paribas Press Release, June 5, 2025
- “HSBC and Mistral AI join forces to accelerate AI adoption across global bank,” HSBC Media Release, Dec 1, 2025
- “The biggest AI shift is taking place in your employees’ bags,” Fast Company, May 5, 2026
https://www.fastcompany.com/91536554/the-biggest-ai-shift-is-taking-place-in-your-employees-bags