A.I. Goes to War + Is ‘A.I. Brain Fry’ Real? + How Grammarly Stole Casey’s Identity

What It Covers

Kevin Roose and Casey Newton examine three converging AI stories: Claude's deployment inside classified U.S. military systems during the Iran conflict, BCG researcher Julie Bedard's findings on "AI brain fry" affecting 14% of heavy AI users, and Grammarly's unauthorized use of real journalists' identities to sell a fabricated expert-review feature.

Key Questions Answered

• •AI in Military Targeting: Claude, integrated into Palantir's Maven Smart System since 2024, has already suggested hundreds of missile targets, issued precise location coordinates, and compressed weeks-long battle planning into real-time operations. It remains the only AI model deployed inside classified U.S. military systems. OpenAI and Google's Gemini are cleared for non-classified Pentagon use, meaning Claude's monopoly on classified systems is temporary.
• •"Human in the Loop" Is Eroding: Military officials publicly maintain humans retain final firing authority, but analysts warn that when AI handles target selection, timing, location coordinates, and post-strike analysis, the human role reduces to pressing a button. The elementary school strike in Iran that killed over 175 people, mostly children, prompted immediate questions about whether AI misidentified the target.
• •AI Brain Fry vs. Burnout: Julie Bedard's BCG study of 1,488 workers found 14% of heavy AI users experience "AI brain fry," defined as cognitive strain from excessive oversight of AI tools beyond one's processing capacity. Critically, brain fry showed no correlation with burnout — they are distinct conditions. Brain fry is cognitive overload; burnout is emotional exhaustion. Treating them as the same leads to wrong interventions.
• •The Three-Tool Cliff: Workers using up to three AI tools simultaneously report productivity gains, but crossing to four or more tools triggers a measurable reversal — increased stress, cognitive overload, and diminished output quality. The mechanism is classic multitasking failure compounded by output governance: more tools generate more deliverables requiring human review, creating a compounding oversight burden that exceeds cognitive capacity.
• •Marketing Workers Hit Hardest: Marketing managers report the highest brain fry rates across all professions surveyed, while lawyers and compliance professionals report the least. BCG's prior skill-disruption modeling found marketing manager roles were 90% disrupted by AI from a skills perspective — the highest of any tracked profession. Jobs with undefined quality thresholds, like image creation and campaign generation, create endless iteration loops that drive cognitive fatigue.
• •Grammarly's Identity Exploitation Model: Grammarly's "Expert Review" feature displayed real journalists' names — including Casey Newton, Cara Swisher, and Timnit Gebru — as editorial advisors without consent, compensation, or actual involvement. The underlying model appeared to be a non-frontier version generating generic, low-quality advice. Investigative reporter Julia Angwin filed a class action complaint against Grammarly's parent company. Within days of Newton's reporting, Grammarly disabled the feature entirely.