Tech Companies Allegedly Using Images from Dutch Creators to Train AI Without Permission
Recent research conducted by NOS suggests that AI models developed by companies like OpenAI and Midjourney may have been trained on images created by Dutch artists without proper authorization. This potential breach of copyright has been supported by four experts in the field.
When approached for a response, OpenAI declined to comment specifically on the allegations but pointed to their policy that allows image creators to file objections. Midjourney did not respond to inquiries.
Distinctive Styles
In the investigation, the AI models were tasked with generating images in the styles of notable Dutch artists such as Erwin Olaf, Eddy van Wessel, and Dick Bruna. The results displayed characteristics closely resembling the iconic works of these creators. For example, the models produced images of women in yellow dresses, reminiscent of Erwin Olaf's photography series. Similarly, black-and-white images of war-torn areas, a hallmark of van Wessel's work, were generated. The models even recreated imagery akin to Dick Bruna’s Miffy, though permission to include these images was not granted by the rights holders.
Interestingly, OpenAI's DALL-E model sometimes blocks the creation of images in the style of certain artists, citing a violation of their terms. However, bypassing these restrictions often proved possible by refreshing the session.
Expert Analysis
Four experts—Evangelos Kanoulas and Efstratios Gavves from the University of Amsterdam, Olya Kudina from TU Delft, and Noelle Cicilia from Brush AI—examined the AI-generated images. They all concurred that while it cannot be confirmed with absolute certainty, it is highly probable that the AI models were trained using Dutch works. According to Kanoulas, even with vague descriptions, the AI can associate specific styles with artists' names. Cicilia noted that generative AI models lack creativity, merely replicating the data they were trained on.
Legal Implications
Intellectual property lawyer Mirjam Elferink explained that copyright infringement is generally straightforward when protected works are used without permission. However, there's an exception for publicly accessible data, which can be utilized by scientific institutions and commercial entities under certain conditions—namely, unless explicit objections are made by rights holders, known as the "opt-out" option.
The works of Dick Bruna and Eddy van Wessel fall under this opt-out clause, potentially leading to copyright violations. However, this is not the case for Erwin Olaf, as Shirley den Hartog, who manages his estate, stated that Olaf had expressed a desire for his work to continue being appreciated worldwide after his passing.
Lack of Awareness and Legal Clarity
Hanneke Holthuis, a legal expert at Pictoright, highlighted that many artists are unaware of this complex copyright loophole. The possibility of AI models being trained with publicly available works was unforeseen, making it difficult for artists to preemptively protect their creations.
The recently enacted European AI law aligns with the opt-out system, requiring creators to declare their intention to opt out, while tech companies must be transparent about their training data to avoid copyright infringement. However, Holthuis pointed out a significant issue: once AI systems are trained, it is challenging to "unlearn" specific data, making it impossible to remove the influence of works not protected by opt-out before training.
Dutch artists can declare their opt-out preference at optoutnow.ai to prevent their work from being used without permission in the future.
Image AI generated for illustration and irony.
