Top AI Agents Fail 85% of Projects as Studies Warn of Cognitive Offloading

2 articles · Updated · The Conversation · Aug 19

Roughly 85% of projects handled by top-performing AI agents still fell short of commissioned-work standards in a July 2026 Centre of AI Safety study, underscoring risks in treating AI output as finished work.
Two in three people now use AI regularly, a 2025 University of Melbourne-KPMG survey of 48,000 respondents in 47 countries found, yet 61% had no AI training and 60% reported inappropriate or non-transparent workplace use.
That gap has already produced costly failures: South Africa withdrew its draft national AI policy over fake citations, and Deloitte partly refunded Australia after an AI-assisted report with fabricated sources cost taxpayers A$440,000.
The report argues the deeper damage is cognitive—speed-driven workplaces reward polished deliverables over the thinking process, weakening workers' ability to verify claims, judge quality and distinguish fact from falsehood.
Rather than ban AI, it urges longer timelines, human-led brainstorming and stricter critique of logic, references and methods so AI supports work without replacing critical thought.