Hard Fork

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

March 13, 2026

Key Takeaways Copied to clipboard!

  • AI tools are being actively deployed by the U.S. and Israeli militaries in the conflict with Iran for intelligence processing and mission planning, marking a turning point in military AI application. 
  • Researchers have identified a distinct psychological phenomenon called 'AI brain fry,' characterized by mental fatigue from excessive oversight of AI tools beyond cognitive capacity, which is separate from general burnout. 
  • Grammarly faced backlash and ultimately disabled its 'Expert Review' feature after being exposed for falsely attributing writing insights to unconsulted and uncompensated experts, including journalists like Casey Newton, highlighting major ethical issues in AI identity appropriation. 
  • The unauthorized use of personal identity by AI companies like Grammarly in features like 'expert review' is driven by the potential for profit, leading to reputational risk for the individuals whose identities are co-opted. 
  • A potentially 'good version' of AI style-matching exists if companies strike revenue-sharing deals with creators and guide users toward actual, verifiable expertise (like suggesting specific authors or articles) rather than relying on LLM hallucinations. 
  • Following public exposure, Grammarly disabled its 'expert review' feature, suggesting that direct criticism can successfully halt problematic AI implementations that leverage personal reputations without consent. 

Segments

AI in Iran War Deployment
Copied to clipboard!
(00:02:38)
  • Key Takeaway: AI is being used by U.S. and Israeli militaries for intelligence processing, target identification, and mission planning, moving beyond theoretical application.
  • Summary: AI tools are processing vast amounts of data from drones and sensors to create real-time dashboards for military operations. Current use focuses on intelligence, mission planning, and logistics, stopping short of fully autonomous lethal decisions. This deployment mirrors surveillance capabilities that could eventually be turned against domestic populations.
Claude’s Role in Military Operations
Copied to clipboard!
(00:11:25)
  • Key Takeaway: Anthropic’s Claude model is currently the only AI deployed inside classified U.S. military systems, integrated into Palantir’s Maven Smart System to suggest targets and prioritize operations.
  • Summary: Claude is reportedly essential to U.S. operations in Iran, turning weeks-long battle planning into real-time execution by suggesting hundreds of targets. The model used is largely the same as the consumer version, potentially with minor fine-tuning for classified environments. This reliance has led the Pentagon to formally declare Anthropic a supply chain risk.
AI Infrastructure as War Targets
Copied to clipboard!
(00:15:24)
  • Key Takeaway: Data centers and fiber optic cables supporting AI infrastructure in the Middle East are becoming legitimate military targets, as evidenced by Iranian strikes on Amazon AWS facilities.
  • Summary: Iran reportedly struck Amazon data centers in the UAE and damaged one in Bahrain, claiming the action targeted infrastructure supporting enemy military activities. This highlights the vulnerability of major cloud infrastructure built in geopolitically unstable regions. Disruptions to undersea fiber optic cables in the Strait of Hormuz also pose a significant, unmitigated risk to global internet traffic.
Defining AI Brain Fry
Copied to clipboard!
(00:22:30)
  • Key Takeaway: AI brain fry is a newly studied condition defined as mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity, distinct from general burnout.
  • Summary: Surveyed workers reported symptoms like having ‘12 browser tabs open in the head’ due to the mental effort required to manage AI tools rather than perform core tasks. This cognitive strain correlates with the intensification of work and increased multitasking, but not with traditional emotional burnout metrics. Marketing managers reported experiencing this phenomenon most frequently.
Grammarly’s Identity Theft Feature
Copied to clipboard!
(00:47:50)
  • Key Takeaway: Grammarly’s ‘Expert Review’ feature falsely implied insights came from unconsulted, uncompensated experts like Casey Newton and Kara Swisher, offering generic advice to paying subscribers.
  • Summary: The feature used names of prominent figures, including critics of AI like Timnit Gebru, without permission or compensation, leading to a class-action complaint. The advice generated was generic ‘word salad,’ suggesting the product was substandard despite its high subscription cost. Grammarly disabled the feature after being publicly confronted, acknowledging the misrepresentation.
AI Identity Theft Fallout
Copied to clipboard!
(01:01:42)
  • Key Takeaway: Unauthorized use of identity for profit will continue unless companies face consequences, as demonstrated by Grammarly’s initial action.
  • Summary: The practice of AI companies profiting from uncompensated identity use is expected to persist because money is involved and companies will engage in it if they can avoid repercussions. The speaker expressed satisfaction that Grammarly retracted the feature after being exposed. This incident highlights the ongoing risk of identity exploitation in the pursuit of AI monetization.
Designing Ethical AI Features
Copied to clipboard!
(01:02:02)
  • Key Takeaway: Ethical AI features could involve revenue-sharing models where creators are compensated for the use of their style or expertise in training or suggestion models.
  • Summary: A hypothetical ‘good version’ of the feature would involve striking deals with creators, such as paying Casey Newton 10 cents every time his style is used to edit emails. This framework supports sharing revenue based on creative work. Such tools could guide writers toward actual experts, for example, suggesting Catherine Boo’s narrative structure and linking to subscription content for relevant passages.
Grammarly’s Flawed Execution
Copied to clipboard!
(01:03:39)
  • Key Takeaway: Grammarly’s implementation risked damaging experts’ reputations by attributing poorly generated, generic content to them based on minimal source material like speaker bios.
  • Summary: The speaker was offended by being used without consent, worrying that poor AI output mimicking his style would lead to him receiving blame for bad writing. Another expert, Matt Honan, discovered his supposed expertise was sourced only from an old event speaker bio, illustrating a lack of thoughtful execution by Grammarly. This demonstrates that the AI was laundering reputation based on flimsy, outdated source data.
Podcast Promotion and Credits
Copied to clipboard!
(01:05:35)
  • Key Takeaway: The episode production team and contact information for listener feedback were detailed following the main discussion.
  • Summary: Hard Fork is produced by Whitney Jones and Rachel Cohn, and edited by Viern Pavich. Listeners can email feedback to [email protected]. The segment concluded with the announcement that Grammarly had disabled the expert review feature.