Tracking what data in used in a machine learning data set can be a daunting task, but as regulations and public scrutiny intensifies, lawyers say it could be a useful tool to stay compliant.

Last week Facebook Inc. joined other developers in the quest to better trace the data used in data sets. In a research paper titled "Radioactive Data: Tracing Through Training" Facebook announced a new method to trace images used in data set for training software. 

To be sure, Facebook isn't the first to announce a method that provides transparency into data sets. In its report, Facebook highlighted numerous data tracking mechanisms including watermarking, differential privacy and membership inference. 

Lawyers contacted by Legaltech News said confirming the usage of specific information in a data set throughout the development of a software could be necessary as regulatory and public pressures grow over data privacy.

Such information can be be leveraged as evidence that an entity isn't compliant with corporate or regulatory privacy policies.

Just this month Facebook agreed to a $550 million settlement over Biometric Information Privacy Act (BIPA) violations and Google was served a similar class action lawsuit over alleged violations of the Illinois law. 

However, Georgetown University Law Center professor Anupam Chander noted that figuring out what data is used in a machine learning data set is likely used for "constrained circumstances" to ensure a company's data isn't being used without its permission, not to provide transparency to data subjects.

Chander cited the recent news of facial recognition app Clearview scraping billions of images from Facebook, YouTube and Venmo for its law enforcement clientele as usage companies would want to prevent. 

"You see the Clearview data set and Facebook has objected to its use of its data and so this is another way to demonstrate that Clearview or some third-party vendor used Facebook's images that has been manipulated to make these types of results."

While data tracking methods may help companies follow their data's usage, Chander said Facebook's method may not be helpful for spotting biased data.

"You need to be able to change the underlining data [in order to identify it] without changing the outcome, that's the promise of [Facebook's] paper. … It may not be so easy to change the underlining data without affecting the outcome in substantive ways when it comes to decisions about credit or employment," he said.

Still, as companies tangle with understanding how their data is being used and potential public backlash, Riesen noted that providing transparency into data sets may potentially leak software insights to competitors.

"This could expose information to a competitor about your proprietary machine learning and AI algorithms that you intended to be a competitive advantage or trade secret. This could lead to competitors studying how your algorithms are processing certain data," he said.

However, only the data sets used for the algorithm could be exposed, and not the processes. What's more, when weighing regulatory compliance and public relations, some companies may prefer to make their data sets more traceable when faced with varying privacy regulations, Riesen said.