Microsoft is exploring a way to credit contributors to AI training data

Microsoft is launching a research project to estimate the impact of specific training examples in text, images and other types of media that produce generator AI models.

It Per a job list Dating with December that was recently rebuilt on LinkedIn.

According to the search list of a research intern, the project will try to prove that the models can be trained in such a way that the effects of specific data – such as photos and books – can be “efficiently and effectively guessing” on their results.

“Current neural network architecture is opaque in terms of source supply for their generation and there is […] This is good to change, “the list reads.”[One is,] In the future we will want the future, in the future, what we want is to encourage, recognize and potentially pay for people who contribute specific valuable data to unexpected models. “

AI -driven text, codes, images, videos and songs are in the center of generators Several IP suit Against AI companies. Often, these companies train their models on a lot of data from public websites, some of which are copyrighted. Many agencies argue Fair -use Their data-scraping and training practices gives s. However, Creatives – from artist to programmer to author – basically do not agree.

Microsoft himself faces at least two legal challenges from copyright holders.

New York Times Case against the Tech Giant And some of its associates, OPNA, accused the two companies deploy two companies in their articles and violated the copyright of the Times. Several software developers Microsoft has also filed a case against the firm’s Githab Copilot AI Coding Assistant to the farm was trained illegally using their protected works.

Microsoft’s new research attempt, which describes the list as “training-time introduction”, ” Report Jaron is involved in Lanier, Skilled technicians and inter -disciplinary scientists At Microsoft Research. April 2023 is a OP-Ed of New YorkLania wrote about the concept of “data dignity”, which means “Digital Staff” to be associated with “people who want to be familiar to create it”.

“When a large model provides a valuable output, a large model will be looking for the most unique and influential contributors to a data-status method.” “For example, if you ask a model for an animated movie in the oil-painting world of cats in an adventure, some key oil painters, cat portraits, voice actors and writers-or their resources-it can be counted unique to create new masterpieces.

In the meantime, several companies are trying to do it, nothing. The AI ​​model developer, brewer, which has recently raised 40 million Million as Venture Capital, has claimed that “programmatically” data owners have been compensated according to their “overall impact”. Adobe and shutterstock also pay regular payments to contributors, although the exact payment amounts begin to become opaque.

A few large labs have established programs for paying separate contributors beyond the licensing agreements with platforms and data brokers. They have instead provided the way to “opt out” the training of copyright holders. However, some of these opt-out processes are extremely rigid, and not only trained in future models.

Of course, the project of Microsoft could be a bit higher than the proof of the concept. There is a precedent for this. Back MayOpeni says it is developing similar technology that trains the creators that they want to include their tasks – or allow to exclude from training data. However about a year later, the equipment has not yet seen daylight and it is often often Did not see internally as priorityThe

Microsoft could also try ”Wash“Here – or regulator and/or court decisions disrupts its AI business.

However, the company is investigating ways to trace training, which is significant in the light of other AI labs recently published on fair use. A number of tops including Google and OpenAI have been published The policy documents proposed The Trump administration weakens copyright protection as it is related to AI development. Open is there Clearly called the US government Model training is to do fair use coding, which argues that developers will relieve the burden of burden.

Microsoft did not immediately respond to any request for comment.