AI Pullback Has Officially Started (www.planetearthandbeyond.co)

The AI “pullback” is underway: multiple studies and high-profile failures show generative models are falling short of hype and eroding confidence across industry and academia. An MIT analysis found 95% of AI pilots didn’t boost profit or productivity; METR found coding tools can slow developers; Gartner reports office AI agents fail to complete tasks ~70% of the time. Corporate adoption among large firms slipped (14% → 12%) while cancellations of AI projects jumped from 17% to 42%. Wiley’s ExplanAItions study shows researcher use rose but confidence fell—fewer now say AI exceeds humans and worry about hallucinations rose from 51% to 64%. Even Deloitte had to refund a government report after AI-generated errors rendered it unusable. Technically, the core problem is hallucination and the human-overhead required to catch it. Melbourne researchers argue AI helps “low-skill” tasks (notes, simple customer service) but harms high-skill, accuracy-critical work because oversight negates productivity gains—and low-skill workers often can’t identify model errors. Academia’s feedback loop was compromised when both authors and peer reviewers used AI, producing fake papers and prompting journals to ban AI in review. For the AI/ML community this signals urgent priorities: rigorous evaluation beyond surface metrics, robust hallucination mitigation, better uncertainty calibration, human-in-the-loop design, and domain-specific verification before widescale deployment. The era of unchecked AI optimism is ending; engineering rigor and conservative deployment will dictate whether the field recovers credibility.