Copyright, Stock Licences and Training Data Provenance
Copyright subsists in literary, dramatic, musical, and artistic works, as well as in sound recordings, films, and typographical arrangements. When you use any of these works in marketing, you must have a licence from the copyright holder or be in scope of an exception to copyright. This applies regardless of whether a human or an AI system created the work.
Generative AI models are trained on vast datasets of images, text, audio, and video. Many of these works are copyrighted. If an AI system generates output that is derived from copyrighted training data, the advertiser (the person using the AI system) may bear secondary liability for copyright infringement. This is a developing area of law, but companies have already faced legal action.
The legal status of generative AI training and copyright is still contested and evolving. Some jurisdictions (the UK, the US) recognise a fair use or permitted use exception for training machine learning systems on copyrighted works. The EU Copyright Directive permits text and data mining for research organisations under Article 3, but commercial mining faces restrictions. However, a court has not yet definitively ruled on whether training a commercial generative AI model on copyrighted works without licence is lawful. Until that question is settled, assume that using output from an AI trained on copyrighted data carries legal risk.
For marketing teams, this means: investigate the provenance of any generative tool you use. If you use an off-the-shelf image generation service (DALL-E, Midjourney, Stable Diffusion), ask the vendor what training data was used and whether the vendor carries liability for copyright infringement. If the vendor says it trains on unlicensed scraped internet data, you inherit that risk. Some vendors now offer indemnification (a guarantee to defend you if sued), which is a signal that they take the copyright risk seriously.
Stock image licences are a safer source of images for marketing. Stock agencies (Getty Images, Shutterstock, Adobe Stock) licence images from photographers who have agreed to licensing. Using a stock image in your marketing is permitted under the licence agreement. Always check the licence terms to ensure the use you intend (paid advertising, web use, time limits, geographic restrictions) is permitted. Most stock licenses prohibit using images for certain purposes without additional fees.
If you hire a photographer or designer to create marketing assets, ensure a work-made-for-hire agreement is in place, which specifies that the copyright in the work belongs to your company, not the creator. Without such an agreement, the creator retains copyright and you have only a limited licence to use the work.
For AI-generated images, the copyright status is uncertain and varies by jurisdiction. Some jurisdictions hold that copyright does not subsist in images generated solely by AI systems because copyright requires human authorship. Others hold that if a human prompted the AI and selected the output, copyright may vest in the human. To be cautious, treat AI-generated images as uncopyrighted (owned by no one) and use them without claiming copyright in them. If you have any doubt, use stock images or hire a creator.
Training data provenance matters for reputational risk as well. If an AI model was trained on datasets known to contain biased, toxic, or non-consensual imagery, using output from that model in marketing can attract criticism and regulatory scrutiny.