Deloitte’s $440k AI Slip Shows Oversight Hasn’t Kept Pace

About a month ago, observers first raised concerns that a report Deloitte had produced for the Australian federal government included material that could not be traced to real sources. Multiple instances of fictitious citations and misattributed quotes were flagged: references to papers and judicial rulings that simply do not exist, and a quote supposedly from Justice Davies that appears nowhere in the relevant court judgment. That suspicion spurred further scrutiny, and Deloitte just admitted it had used generative AI in producing the report, prompting it to partially refund the government and issue corrections.

That incident is not isolated. It exemplifies a deeper and growing problem in how AI is being used in professional contexts: the normalization of “AI slop”, that is, machine-generated content that is polished on the surface but flawed underneath, and which slips through because it appears credible. The Deloitte case should serve as a warning: capability without oversight is a liability.

In the Deloitte report’s original version, academics noted that several of the cited works by scholars at the University of Sydney and other institutions do not exist, and that the legal quotes were distorted or fabricated. Deloitte defended its findings but acknowledged the need to correct references. When the revised version was posted, it included disclosure of AI assistance via Azure OpenAI GPT-4o. The firm maintained the substance of the report but conceded that footnotes and citations had errors.

Meanwhile, in a different neighborhood of AI use, OpenAI’s Sora 2 model has reignited debates about how generative media enters public discourse. The tool allows users to generate short videos from prompts, potentially including uploaded faces and voices. Some of its early outputs were mistaken by viewers for real footage, prompting criticism over labeling, intellectual property, and the boundary between synthetic and authentic content. Vinod Khosla, an early investor in OpenAI, dismissed the critics of Sora 2 as “tunnel vision creatives” and urged that viewers themselves should assess the output.