AI Daily Digest - 2026-02-15

Table of contents

Pentagon Threatens to Cut Anthropic Over AI Use Restrictions

Category: Regulation and Policy Source: Axios

The Pentagon is considering ending its partnership with Anthropic after the AI safety company refused to grant the military unrestricted access to its Claude models, according to a senior administration official. The standoff centers on Anthropic’s insistence on maintaining hard limits around fully autonomous weapons and mass domestic surveillance—restrictions the Department of Defense says are unworkable.

Tensions escalated following reports that Claude was used during the recent U.S. military operation to capture Venezuela’s Nicolás Maduro. An Anthropic executive allegedly contacted Palantir to inquire whether Claude had been deployed in the raid, which involved kinetic fire and casualties. However, Anthropic flatly denied the characterization, stating it has “not discussed the use of Claude for specific operations” and that its conversations with the Pentagon focus solely on “Usage Policy questions” around autonomous weapons and surveillance—not current military operations.

The dispute highlights a broader culture clash between AI safety advocates and defense officials. The Pentagon is pushing four leading AI labs—Anthropic, OpenAI, Google, and xAI—to allow military use of their models for “all lawful purposes,” including sensitive weapons development and intelligence operations. While OpenAI, Google, and xAI have shown flexibility, Anthropic remains the holdout. The senior official described Anthropic as the most “ideological” of the labs when it comes to perceived AI risks, though they conceded replacing Claude would be difficult because “the other model companies are just behind” in specialized government applications. Anthropic signed a contract worth up to $200 million with the Pentagon last summer and was the first to deploy models on classified networks.

Frontier AI Models Fail 96% of Real-World Freelance Jobs

Category: Research and Innovation Source: Revolution in AI

A new benchmark testing AI’s ability to complete real paid work reveals a stark gap between hype and reality: even the best frontier models fail over 96% of professional freelance tasks. The Remote Labor Index (RLI), developed by researchers using 240 actual Upwork jobs averaging $630 in value, found that Claude Opus 4.5—the top performer—succeeded in just 3.75% of assignments. GPT-5.3 Codex managed 2.80%, while Gemini 2.0 Ultra achieved only 1.25%.

The failures weren’t due to wrong answers but practical delivery issues that would be unacceptable in paid work: corrupt or unusable files, incomplete deliverables, quality falling short of professional standards, and internal inconsistencies like architectural renderings that don’t match floor plans. The benchmark deliberately included complex client-style work—video production, CAD design, game development, architecture—where “almost correct” still counts as failure. Models performed better on tasks resembling first-draft work like creative ideation, simple advertising, and report writing, but collapsed when asked to deliver polished, professional-grade outputs with proper file formats and asset management.

The findings suggest that while AI can accelerate certain workflow stages, the “last mile” of professional delivery remains firmly in human hands. Most common AI benchmarks test isolated skills like writing paragraphs or solving logic puzzles, but the RLI measures end-to-end execution against real client expectations. The gap reveals why many CEOs struggle to see ROI from AI deployments despite breathless headlines about workforce replacement: the technology excels at generating drafts but fails at the quality control, consistency, and technical polish that clients actually pay for. The researchers maintain an updated leaderboard at RemoteLaborAI, with full methodology available in their published paper.

NPR Host Sues Google Over Alleged AI Voice Cloning

Category: Security and Privacy Source: The Washington Post

David Greene, longtime NPR host and former “Morning Edition” co-anchor, is suing Google after discovering that the company’s NotebookLM AI tool appears to use a synthetic voice strikingly similar to his own—without his permission or compensation. Greene learned about the similarity when a former colleague emailed in fall 2024 asking if he’d licensed his voice to Google. In his lawsuit, Greene alleges the tech giant violated his rights by building a product that replicated decades of vocal training and professional development.

NotebookLM is Google’s popular AI tool that generates on-demand podcasts from user-uploaded documents, complete with synthetic hosts that discuss the content. The case highlights growing legal and ethical questions around AI voice cloning, particularly when companies train models on publicly available audio without explicit consent from the individuals whose voices are replicated. Greene spent his career cultivating a distinctive broadcast voice recognizable to millions of public radio listeners—an asset he argues was appropriated without authorization.

The lawsuit adds to mounting concerns about AI companies’ training practices, particularly regarding voice, image, and likeness rights. Unlike text-based training data disputes, voice cloning raises additional questions about performers’ rights and the value of distinctive vocal characteristics developed through years of professional work. Google has not yet publicly responded to the specific allegations. The case could set important precedents for how AI companies must handle voice data and whether public figures retain control over synthetic reproductions of their distinctive vocal traits.

Anthropic CEO Doubles Down on China AI Chip Export Restrictions

Category: Regulation and Policy Source: Times of India

Anthropic CEO Dario Amodei has intensified his opposition to selling advanced AI chips to China, directly contradicting Nvidia CEO Jensen Huang’s ongoing lobbying efforts. In a wide-ranging conversation on the Dwarkesh Patel podcast, Amodei framed AI chips as a national security chokepoint rather than a trade commodity, warning that equal access to powerful AI systems could create a standoff “far more dangerous than nuclear deterrence.”

Amodei’s position—which he first articulated at Davos by comparing chip sales to “selling nuclear weapons to North Korea”—argues that while AI’s economic benefits should be shared globally, critical compute power must remain out of authoritarian hands. He reframed the export debate around political freedom rather than quarterly revenue, suggesting that economic growth from AI will arrive “faster than we can take it,” but the hard part will be protecting democratic values and preventing authoritarian governments from weaponizing AI against their own citizens.

The dispute has become increasingly personal despite Nvidia pouring over $10 billion into Anthropic. Huang has accused Amodei of pushing for more AI regulation primarily because it benefits Anthropic’s competitive position, and has criticized his warnings about AI eliminating half of entry-level jobs as self-serving fear-mongering. The Nvidia-Anthropic rivalry is emerging as the defining policy feud in AI, with Amodei betting that limiting China’s access to frontier compute is essential for global stability, while Huang sees it as protectionism that will ultimately fail and harm American competitiveness.

Open Source Medical AI Combines Knowledge Graphs with Compact Models

Category: Healthcare and Life Sciences Source: Reddit

Researchers have released Open Book Medical AI, a hybrid architecture that combines a compact 3GB language model with a deterministic medical knowledge graph containing 5,000 nodes and 25,000 edges spanning diseases, symptoms, treatments, risk factors, diagnostic tools, body parts, and cellular structures. The system addresses a critical challenge in healthcare AI: most medical models rely on large, opaque language models that are difficult to audit and expensive to deploy in HIPAA-compliant environments.

The 3GB footprint is the key innovation, enabling on-premise deployment in hospitals and labs without patient data ever leaving the facility—a compliance requirement that most 80GB-class frontier models cannot meet. The architecture separates concerns: the knowledge graph provides deterministic medical reasoning and constrains what relationships and treatments the system can suggest, while the compact LLM handles natural language explanation and contextualization. Every response can be traced back to specific nodes and edges in the graph, creating a structured audit trail showing the path from symptom to disease to diagnostic tool to treatment.

This design offers several advantages over pure LLM approaches: dramatically lower hallucination rates since medical claims must exist within the structured ontology, verifiable answers via RAG-based audit layers, complete control over the medical knowledge base including update governance, and the ability to run inference locally without cloud dependency. The model is currently available for public testing on Hugging Face Spaces, where it demonstrates clinical case analysis and diagnostic reasoning capabilities. The hybrid approach suggests that in high-stakes domains like healthcare, controllability and traceability may matter more than parameter count.

Quick Updates

• New SCAM Benchmark Exposes AI Agent Security Flaws: 1Password released SCAM (Security Comprehension Awareness Measure), an open-source benchmark testing whether AI agents can complete workplace tasks without falling for phishing, social engineering, and credential theft. Claude Opus 4.6 achieved only 35% safety score in a recorded scenario involving a fake SharePoint credential harvesting page. Source: 1Password SCAM

• Axios CTO: AI Lets One Engineer Complete 3-Week Project in 37 Minutes: Dan Cox, Axios’ chief technology officer, reports that AI-based “agent teams” enabled an engineer to replicate a three-week project in 37 minutes. The company reduced its tech team from 63 to 43 people while doubling output, with plans to double again. Cox notes the bottleneck is no longer writing code but “how fast humans can absorb change.” Source: Axios

• Senate Democrats Introduce Bill to Ban Grocery Store Surveillance Pricing: Senators Ben Ray Luján and Jeff Merkley introduced the Stop Price Gouging in Grocery Stores Act, which would ban electronic shelf labels in large stores and require disclosure of facial recognition use. The legislation targets “surveillance pricing” where AI algorithms adjust prices based on factors like customer demographics, time of day, or individual shopping history. Source: Gizmodo

• Nvidia CEO Jensen Huang Cancels India AI Summit Appearance: Nvidia confirmed CEO Jensen Huang will not attend next week’s India AI Impact Summit, where he was expected to be a major draw. Prime Minister Narendra Modi will inaugurate the event Thursday. No reason was given for the cancellation. Source: Reuters via Reddit

• Study Explores “AI Psychosis” and Distributed Delusions: New research published in Philosophy & Technology examines how routine reliance on generative AI can lead to “distributed delusions” where users hallucinate with AI rather than being hallucinated at. The paper argues chatbots’ conversational style allows them to function as both cognitive artifact and quasi-Other, making them particularly seductive vehicles for sustaining and elaborating delusional thinking. Source: Springer

• Open Source Stability Kernel Released for Agentic AI Systems: Developer noblebrendon-cloud released coherence-stability-kernel v1.0, a deterministic runtime stability framework for agentic AI that monitors five risk signals and enforces regime-based operational limits. The MIT-licensed system uses a ~3GB footprint and provides structured audit layers to constrain AI agent behavior without relying on large opaque models. Source: GitHub