The Story of a Tarot Deck to be Public Domain Very Soon

I’ve been doing tarot for the lst 20-ish years, and recently I came across a tarot deck that is about to enter the public domain. This got me thinking about how images, like those in the tarot deck, might be used to train AI models, and whether there’s a way to determine if a specific image has been included in a model’s training data.

The idea is simple: recreate the about to become public deck using AI, and then compare the generated images with the original ones to see if the model has been influenced by the specific images from the deck. If the generated images closely resemble the originals, it could indicate that the model has been fed by those specific images.

Now, my inquest is to reuse those to be released images for a maybe only personal deck, or maybe doing physical copies of the deck for resale. The ethical and legal implications of this are still something I’m trying to navigate, especially considering the images will soon be in the public domain. But if I’m giving trouble to my as regarding permissions and copyrights, it might be worth waiting until the images are officially in the public domain to avoid any potential issues. Still, the prototype phase of this experiment feels like a gray area, and I’m curious to see how it unfolds once the images are publicly available.

The thing is that images can be identified by unique patterns, features, or metadata that distinguish them from other images. AI models often learn these distinctive characteristics during training, which means that if a model has been fed a specific image, it might generate outputs that closely resemble the original image’s unique elements. By analyzing the similarities between generated images and the original ones, it may be possible to infer whether the model has been influenced by the specific images in question.

And this can be easily tested by generating multiple images using the AI model and comparing them with the original images from the tarot deck. By examining the degree of similarity, one can make an informed guess about whether the model has been influenced by the specific images in question.

Of course, this method is not foolproof. AI models can sometimes produce outputs that coincidentally resemble certain images without having been explicitly trained on them. Therefore, while this approach can provide valuable insights, it should be complemented with other investigative techniques and a healthy dose of skepticism. Ultimately, the goal is to better understand the relationship between training data and model outputs, especially as it pertains to images that are on the verge of entering the public domain.

The Experiment

The experiment involves several steps to test whether an AI model has been influenced by the specific images from the tarot deck:

1. Select the Image(s): Choose a set of images from the tarot deck that are about to enter the public domain.
2. Digitally Trace the Image(s): Create digital tracings of the selected tarot images to capture their unique features and patterns. This step helps in generating AI images that closely resemble the originals for comparison purposes. The tracings are done in the color and style of the original images to maintain fidelity. If a red sphere is in the image, it should be traced in red as well.
3. Write Metadata of the Image(s): Record relevant metadata for the selected tarot images, such as creation date, artist, and distinctive features; and the latter is as much as possible detailed and specific, including color schemes, shapes, and any unique elements present in the image. This information can help in the comparison process and provide additional context for assessing the influence of the images on the AI model.
4. Generate AI Images: Use an AI model to generate images based on the digital tracings and metadata of the selected tarot images. This step aims to produce AI-generated images that closely resemble the originals for comparison purposes.
5. Compare Images: Analyze the generated images and compare them with the original tarot images, focusing on unique patterns, features, and overall resemblance.
6. Assess Similarity: Determine the degree of similarity between the generated images and the originals. High similarity may suggest that the model has been influenced by the specific images.
7. Document Findings: Record observations and any patterns noticed during the comparison process to draw conclusions about the model’s training data influence.

By following these steps, one can systematically investigate whether an AI model has been fed by specific images, while keeping in mind the limitations and potential for coincidental similarities. It ain’t foolproof, but it provides a structured approach to understanding the influence of specific images on AI models. Without being too explicit on the placement of the elements within the images, this method allows for a careful and nuanced analysis of potential training data influence.

Conclusion

In conclusion, while it is challenging to definitively determine whether an AI model has been trained on specific images, the outlined experiment provides a structured approach to investigate potential influences. By carefully selecting images, creating detailed digital tracings, recording comprehensive metadata, generating AI images, and systematically comparing them, one can gain valuable insights into the relationship between training data and model outputs. However, it is essential to remain cautious and consider the possibility of coincidental similarities, as no method can guarantee absolute certainty, but only with a high probability.