Researchers Call for Safe Harbor for the Evaluation of AI Tools

Artificial intelligence stakeholders are calling for safe harbor legal and technical protections that will allow them access to conduct “good-faith” evaluations of various AI products and services without fear of reprisal. More than 300 researchers, academics, creatives, journalists and legal professionals had as of last week signed an open letter calling on companies including Meta Platforms, OpenAI and Google to allow access for safety testing and red teaming of systems they say are shrouded in opaque rules and secrecy despite the fact that millions of consumers are already using them.

“We propose that AI companies make simple policy changes to protect good faith research on their models, and promote safety, security, and trustworthiness of AI systems,” the signatories wrote in the open letter posted to the MIT website.

“Independent evaluation is necessary for public awareness, transparency, and accountability of high impact generative AI systems,” begins the first section, followed by the allegation that “currently, AI companies’ policies can chill independent evaluation.”

An academic paper released concurrent with the letter notes that “the terms of service and enforcement strategies used by prominent AI companies to deter model misuse have disincentives on good faith safety evaluations. This causes some researchers to fear that conducting such research or releasing their findings will result in account suspensions or legal reprisal.”

VentureBeat reports that “in a recent effort to dismiss parts of The New York Times’ lawsuit, OpenAI characterized the Times’ evaluation of ChatGPT as ‘hacking.’”

“What OpenAI bizarrely mischaracterizes as ‘hacking’ is simply using OpenAI’s products to look for evidence that they stole and reproduced the Times’s copyrighted works,” NYT’s lead counsel told VentureBeat.

A big part of the interest in safe harbor protection seems motivated by a desire to get under the hood to see how generative AI foundation models are trained, a subject on which most of the big AI companies are extremely secretive. Anthropic, Inflection and Midjourney were also mentioned by VentureBeat in the context of restrictive terms of service.

A blog post summarizing the point of view of the paper and letter describes independent risk assessment as “an essential mechanism for providing accountability,” and says the group’s focus is “on post-release evaluation of models (or APIs) by external researchers beyond the model developer.”

“The effort lands as AI companies are growing aggressive at shutting outside auditors out of their systems,” according to The Washington Post.

No Comments Yet

You can be the first to comment!

Leave a comment

You must be logged in to post a comment.