New Certification Ensures Fair Training of Generative AI Systems

New Certification Determines Fair Training of Generative AI Systems

A new certification has been introduced to assess whether generative AI systems have been trained in a fair manner. The certification, provided by Fairly Trained, a non-profit organization, aims to determine whether generative AI systems have been trained using consented data.

Background

As debates continue between creators and generative AI companies regarding the appropriate use of data for AI system training, Fairly Trained has emerged to bridge the gap. The company believes that consumers should be aware of which companies prioritize creators’ consent in their training data.

Fairly Trained was founded by CEO Ed-Newton Rex, who previously worked as the Vice President of Audio at Stability AI. Rex’s concerns about the use of copyrighted data for generative AI system training led to the establishment of Fairly Trained.

Certification Process

Companies seeking certification must meet specific requirements. The training data used must be provided to the model developers in accordance with contracts with rights holders. The data should either be available under open licenses or owned entirely by the model developers.

Obtaining a license from the rights-granting institution is considered consent for the purpose of certification. The same protocol applies to models used to generate synthetic data for AI system training.

Companies must have a robust data auditing process and maintain records of the training data used for each model to complete the certification process.

Certification and Fees

Fairly Trained currently offers a single certification called L Certification or License Model Certification. Providers of generative AI systems that use “consented” data for training can apply for this certification.

Companies can initiate the application process by filling out a simple online form. Fairly Trained reviews the submitted documents and may request additional information. Upon successful completion of the submission, companies are required to pay an annual certification fee ranging from $500 to $6,000 based on their revenue.

It is important to note that if a company changes its data practices in a way that violates Fairly Trained’s rules or categories, the certification may be canceled without a refund.

Current Status

Fairly Trained has already certified eight startups in the generative AI industry. The aim of these certifications is to promote transparency and accountability in the training of generative AI systems by prioritizing consent and fair use of data.

Sources: Computerworld, InfoWorld