The Australian judiciary has been dubbed “Mecha Hitler” after discussions last week about the classification of anti-Semitic remarks as terrorist and violent extremist content, with chatbots producing such comments also coming under scrutiny.
Nevertheless, experts from X contend that large-scale language models lack intent, placing accountability solely on the users.
Musk’s AI firm, Xai, issued an apology last week regarding statements made by the Grok chatbot over a span of 16 hours, attributing the issue to “deprecated code” that became more influenced by existing posts from X users.
The uproar centered around an administrative review hearing on Tuesday, where X contested a notice from Esafety Commissioner Julie Inman Grant issued last March, demanding clarity on actions against terrorist and violent extremism (TVE) content.
The ban on social media in Australia for those under 16 is now law, with numerous uncertainties still remaining – Video
Chris Berg, an expert witness from X and a professor at RMIT Economics, testified that it is a misconception to believe a large-scale language model can inherently produce this type of content, as it plays a critical role in defining what constitutes terrorism and violent extremism.
Contrarily, Nicolas Suzor, a law professor at Queensland Institute of Technology and one of Esafety’s expert witnesses, disagreed with Berg, asserting that chatbots and AI generators can indeed contribute to the creation of synthetic TVE content.
“This week alone, X’s Grok generated content that aligns with the definition of TVE,” Suzor stated.
He emphasized that AI development retains human influence, which can mask intentions, affecting how Grok responds to inquiries aimed at “quelling awareness.”
The court heard that X believes its Community Notes feature, which allows user contributions to fact-checking, along with Grok’s analytics feature, aids in identifying and addressing TVE material.
After the newsletter promotion
Josh Roose, a witness and political professor at Deakin University, expressed skepticism regarding the utility of community notes in this context, stating that TV has urged users to flag content to X. This has resulted in a “black box” scenario for the company’s investigations, where typically only a small fraction of material is removed and a limited number of accounts are suspended.
Suzor remarked that it is hard to view Grok as genuinely “seeking the truth” following recent incidents.
“It’s undisputed that Grok is not effectively pursuing truth. I am deeply skeptical of Grok, particularly in light of last week’s events,” he stated.
Berg countered that X’s Grok analytics feature had not been sufficiently updated in response to the chatbot’s output last week, suggesting that the chatbots have “strayed” by disseminating hateful content that is “quite strange.”
Suzor argued that instead of optimizing for truth, Grok had been “modified to align responses more closely with Musk’s ideological perspectives.”
Earlier in the hearing, X’s legal representatives accused the proceedings of attempting to distort the Royal Commission’s focus on certain aspects of X. Cross-examination raised questions regarding pre-existing meetings prior to any actions taken against X employees.
Government attorney Stephen Lloyd stated that X was portraying Esafety as overly antagonistic in their interactions, attributing the “aggressive stance” to X’s leadership.
The hearing is ongoing.
Source: www.theguardian.com
