You can get paid for what you post online, as long as it’s used to train the AI.
At least that’s what Dr. Margaret Mitchell, chief ethical scientist at U.S. open source AI company Hugging Face, says, calling on AI companies to use their technology to trace generated AI content back to its original creator.
“A lot of content is being taken away from creators, not just artists and writers, but also ordinary people who interact with what we write online, and they are not paid for their work,” she said at AI Everything in Cairo, Egypt.
“I envision a future where we can actually identify who was in the input space that made that possible. [for a model] To create that output and supplement them. ”
Generative AI relies more on some creators than others, and some AI-generated works often have clear and direct connections between input and output, such as a clear writing style or artist signature.
Recently, Japanese animator and director Hayao Miyazaki criticized the process for producing images that mimic his signature Studio Ghibli style.
But it’s not just musicians and artists whose work is used. After all, large-scale language models (LLMs) like ChatGPT and Google Gemini are trained on vast online databanks.
“We are all creators,” Mitchell said. BBC Science FocusTherefore, it states that reward models should consider contributions from all online users. For example, an experiment with poetry or a photo of a sunset I took on vacation five years ago.
Fortunately, there are potential models that can track output to input and reward creators based on the amount of input they provide, she added.
However, such a system does not currently exist, and current AI business models make it impossible to raise the funding to build one, Mitchell said, although he noted that some AI companies are beginning to investigate how it might work.
For example, in AI company Anthropic’s 2021 documents (recently unsealed by court order), CEO Dario Amodei proposed a “crazy idea” for such a model, theorizing that a payment distribution system could work similarly to monetization platform Patreon.
Mitchell said that even LLMs that are already built can apply existing technological approaches to make this possible. For example, clustering algorithms (machine learning techniques that group data points) can help track similarities and uncover author attribution.
To ensure privacy, this model requires online user consent, so users can choose to have their data associated with them (attributed and compensated if used) or remain anonymous.
“In order to innovate, we have to open the door to this kind of research,” Mitchell said. “Right now it’s not open at all.”
read more:
