In today’s AI ecosystem there are two general types of training data: organic and synthetic.

Organic data describes information generated by actual humans, whether that’s a piece of writing, a numerical dataset, a song, an image, or a video. Synthetic data is created by generative AI models using organic data as a base material.

Just as a photocopy of a photocopy can drift away from the original, when generative AI is trained on its own synthetic data, its output can also drift away from reality, growing further apart from the organic data that it was intended to imitate.

The use of synthetic data isn’t inherently good or bad. In medical research, for example, it’s a critically important tool that allows scientists to make new discoveries while protecting the privacy of individual patients.

At the Transparency Coalition, we are not calling for limitations on the creation or use of synthetic data. What’s needed is disclosure: Developers should be transparent in their use of synthetic data when using it to train an AI model.

Training data is the foundation of artificial intelligence. It’s what AI systems like ChatGPT use to provide answers to the prompts we provide. It’s what generative image systems like Midjourney and DALL-E use to conjure AI-created art.

With the emergence of generative AI, it now takes just a few button-clicks for anyone to create or manipulate data and convince others that fake content is real. 

Training data is the foundation of artificial intelligence. It’s what AI systems like ChatGPT use to provide answers to the prompts we provide. It’s what generative image systems like Midjourney and DALL-E use to conjure AI-created art.

Today most media/tech companies are coalescing around the standard created by the Coalition for Content Provenance and Authenticity (C2PA) .

Developers of AI systems should be required to provide documentation for all training data used in the development of an AI model.

Legislative policies requiring the disclosure of the use of AI are developing side-by-side along the emerging standards in AI provenance. They’re not perfectly in sync—and that’s okay.

With the emergence of artificial intelligence systems like ChatGPT and CoPilot, data privacy has emerged as one of the most urgent consumer protection issues of the 2020s.

At the Transparency Coalition we believe AI policy discussion and legislative action happen at many levels simultaneously. Our mission is to address known AI safety and privacy risks with practical solutions. We’re focused on bringing transparency to both AI inputs and AI outputs.

It’s not difficult to conjure up apocalyptic scenarios set in motion by the advancement of artificial intelligence. Humans have been entertained by techno-catastrophe since Mary Shelley published Frankenstein in 1818.

That’s not to say AI risks should be dismissed as fiction.

Most state data privacy laws in the U.S. operate on an opt-out basis, meaning that users are assumed to have consented to the use of cookies unless they actively decline, ie, opt out. 

Many of the problems caused by AI today are, at their heart, issues of transparency.

The developers of the most popular chatbots, like ChatGPT, refuse to divulge any information about the datasets on which their AI models trained. That denies corporate deployers and individual consumers the ability to judge the quality of the AI system.

AI computer systems are trained to think, learn, and make decisions independently. While common algorithms follow step-by-step instructions to solve specific tasks, AI systems can analyze data, recognize patterns, and improve their performance over time without explicit programming for every scenario.

At its heart, an AI system is a highly sophisticated computer program. That program, known as a model, requires enormous amounts of computing power and massive datasets. By ingesting the datasets, the model “learns” about the structure of language, for instance, or patterns derived from millions of images.

Sophisticated AI systems like ChatGPT and Copilot require two things: enormous amounts of computing power and massive datasets on which to train. Those assets have only become available in recent years, fueling the industry wide boom.