Data Ownership

Module 1 discussed the methods used to train GenAI tools. Each GenAI model’s properties are determined by the data used to train them.

Many of the most sophisticated GenAI tools are trained on enormous datasets. This raises some important questions as to exactly whose data is used. Data ownership refers to a person or corporation having the rights and responsibility for their own data. Many of the authors/owners of the data on which these models were trained may be unaware that their data was used for training. Several have sued various GenAI creators over the use of their data.

Image-making GenAIs like DALL-E, Midjourney, and Stable Diffusion require massive amounts of image data for training. While the content generated by these AIs contains new elements, several users have noticed that some images very closely resemble other images found online.

For example, Getty Images has sued the makers of Stable Diffusion over some image output that appears to depict their property. In the example of a Getty-owned image from a soccer match below, the Getty watermark itself was recreated by the GenAI.

Example of a real picture compared to an AI-generated image
that turned out poorly.
Source: The Verge

Relatedly, some users have noticed that GenAI artwork often contains a scribble of a signature in the bottom corner. The AIs aren't likely explicitly programmed to do this, but they were trained on images that contained signatures, and so they have “learned” that signatures are a common feature of art.

Data Privacy

Beyond data ownership, data privacy is also an important consideration with GenAI. Data privacy refers to the individual’s capacity to control the extent, timing, and circumstances of how information specifically related to oneself (physically, behaviorally, or intellectually) is shared with others. Once private information gets into a GenAI system, unless explicitly stated otherwise in its privacy policy, it becomes part of that system moving forward.

Unless otherwise stated, even the requests we use to prompt GenAI for output get incorporated into GenAI. These requests are often used to fine-tune further or train the underlying model. If we were to write a prompt that itself contains some private content, then the next version of that GenAI tool could be trained based on it. Asking a question that might be stigmatizing can also become problematic in that the system will "learn" that information and draw on it to create future output.

Protecting our Data

In addition to our data, we need to safeguard others’ data. Confidential data about other students, friends, relations, or any research you do with humans for a class, work, or personal projects should not be entered into publicly available GenAI tools. Doing so violates FERPA and other regulations, and the information may well be extracted for purposes other than the original intent, including commercial purposes that some see as exploitative.

We advise reading the privacy policy of any GenAI systems before entering prompts.