AI Daily Digest - 2026-03-25

Table of contents

GitHub Expands Copilot Training on User Interactions, With Opt-Out for Individual Paid Users

Category: Enterprise AI Source: GitHub Blog

GitHub said that starting April 24 it will use interaction data from Copilot Free, Pro, and Pro+ users to train and improve its AI models unless those users opt out in settings. The change covers prompts, outputs, code snippets, nearby context, comments, file names, repository structure, navigation patterns, and feedback signals. Copilot Business and Copilot Enterprise accounts are excluded from the new policy.

The announcement matters because it widens the flow of real-world developer behavior into model improvement at a moment when coding assistants are competing on accuracy, context awareness, and trust. GitHub says training on Microsoft employee interaction data has already improved acceptance rates across multiple languages, and it is now applying that logic to a broader user base. For developers, the practical takeaway is that private repo content “at rest” is not used, but content processed during active Copilot use can be included for training unless the user disables the setting.

This is likely to sharpen the privacy debate around developer tools. Many engineers will see the policy as consistent with how mainstream AI products are evolving, while others will view the default as too aggressive given the sensitivity of code, comments, and workflow metadata. It also reinforces a market split: consumer and prosumer AI tooling is drifting toward opt-out data collection, while enterprise plans are increasingly sold on contractual isolation and governance.

Judge Reopens Key Copyright Claims in Meta AI Case

Category: Policy & Regulation Source: CourtListener

A federal judge in Kadrey v. Meta Platforms, Inc. granted plaintiffs leave to file yet another amended complaint, allowing them to add a contributory infringement claim tied to Meta’s alleged use of BitTorrent while downloading copyrighted works for AI-related purposes. The court also denied class discovery for now, saying that step can wait unless the plaintiffs survive summary judgment on the remaining distribution-related claims.

The order is notable less for a final ruling on AI training legality than for what it revives procedurally. Judge Vince Chhabria said Meta previously won summary judgment on the named plaintiffs’ training claim because plaintiffs failed to present meaningful evidence on market harm, not because the court definitively blessed AI training as fair use. That distinction keeps pressure on one of the industry’s central legal defenses: the idea that copying books and other works to train generative models is broadly protected.

For AI companies, the immediate implication is that copyright litigation over training data remains unsettled and highly fact-specific. For rights holders, the ruling keeps alive additional routes for challenging model development practices, especially where acquisition methods and file-sharing mechanics may create separate liability questions. The case continues to be one of the most closely watched signals for how U.S. courts may distinguish training, distribution, and contributory infringement in the generative AI era.

Google Positions TurboQuant as a New Compression Play for Efficient AI

Category: Research Source: Google Research

Google Research published details on TurboQuant, a new effort aimed at pushing extreme model compression without giving up practical usefulness. While the post is light on the broader competitive framing, the headline message is clear: model efficiency is becoming as important as raw scale, especially as teams try to deploy capable systems on tighter hardware, lower power budgets, and cheaper inference stacks.

Compression work matters now because the economics of AI are shifting from headline model launches to operational realities. If methods like TurboQuant can reduce memory and compute requirements enough to preserve quality at lower cost, they become relevant far beyond research labs. Enterprises want smaller serving footprints, edge deployments need tighter models, and open-source builders increasingly care about what runs well outside hyperscaler infrastructure.

The larger implication is strategic. Efficient AI is no longer a side quest; it is central to who can ship, where models can run, and how widely advanced systems can be adopted. TurboQuant lands on the same day as several stories about local inference, small-model serving, and consumer-grade hardware, reinforcing that efficiency is becoming one of the field’s defining battlegrounds. ⚡

New U.S. Jobs Risk Index Maps AI Exposure Down to Cities, States, and Salaries

Category: Skills and Workforce Source: Digital Planet at Tufts University

Digital Planet at Tufts University published what it describes as the first American AI Jobs Risk Index, estimating that 2.7 million to 19.5 million U.S. jobs could be displaced by AI over the next two to five years, with a midpoint of roughly 9.3 million. The report argues that the most exposed roles are not only routine jobs but many high-income, knowledge-intensive occupations, including writers, programmers, and web designers.

One of the report’s most important contributions is geographic specificity. It identifies “Wired Belts” such as Silicon Valley, Boston, Seattle, Washington, and university-centered metros as especially exposed, with San Jose and Washington, D.C. ranking high by proportional risk. The report also estimates that annual wages tied to at-risk jobs range from $200 billion to $1.5 trillion, showing that AI disruption could hit income concentration as much as employment totals.

For business leaders and policymakers, the value here is not just the large national number but the map of where transition pain may surface first. The report also highlights a policy mismatch: the regions building AI infrastructure and talent ecosystems are often the same regions most vulnerable to labor displacement from that technology. Even if the estimates prove high, the study adds more urgency to retraining, local transition support, and more realistic workforce planning around cognitive automation.

NVIDIA Says Power-Flexible AI Factories Could Ease Grid Strain Instead of Worsening It

Category: Infrastructure and Hardware Source: NVIDIA Blog

NVIDIA highlighted a white paper from Emerald AI, EPRI, National Grid, and Nebius showing how AI clusters can temporarily reduce their power draw during periods of grid stress. In a U.K. demonstration, a 96-GPU Blackwell Ultra cluster adjusted consumption in response to simulated disruptions and even a recreated “TV pickup” surge, the familiar spike in electricity demand when millions of people boil kettles during halftime.

That framing is strategically important because AI data centers are increasingly criticized as inflexible, power-hungry loads that worsen grid constraints and raise costs for everyone else. The pitch from Emerald AI and NVIDIA is that future AI factories can act more like responsive industrial assets, throttling lower-priority workloads while protecting high-priority jobs and helping operators manage peak demand without immediately overbuilding infrastructure.

If that approach works outside demonstrations, it could influence permitting, utility negotiations, and public acceptance of large AI campuses. It also lands in sharp contrast with calls elsewhere in the U.S. to pause new data center construction altogether. The next big question is whether power-flexible scheduling can move from pilot-stage orchestration into enforceable, real-world commitments that regulators and communities trust.

Quick Updates

• Sanders and Ocasio-Cortez propose a pause on new AI data centers: Progressive lawmakers introduced a bill to temporarily halt new AI data center construction while federal safeguards are developed, citing electricity, water, climate, and community impacts. Source: The Guardian
• Prediction markets see Anthropic leading today but a far messier year-end race: PredictMarketCap’s roundup says Claude is heavily favored in the near term, while year-end odds tighten and traders price GPT-5.5 as likely by midyear. Source: PredictMarketCap
• NVIDIA argues the future belongs to both open and proprietary models: Recapping GTC discussions, NVIDIA says enterprise AI will rely on model orchestration, specialist systems, and a mix of open and closed approaches rather than a single winner-take-all model stack. Source: NVIDIA Blog
• AI is helping detect landslide and avalanche risk earlier: BBC Future reports that researchers are using radar satellite data and AI models to flag unstable slopes in Nepal, Britain, and alpine regions, improving early warning and evacuation planning. Source: BBC Future
• Health NZ warns staff against using ChatGPT for clinical notes: New Zealand health officials told staff not to use free AI tools like ChatGPT, Claude, or Gemini for clinical documentation, citing privacy, security, and accountability concerns. Source: RNZ
• Chicago is getting an AI-first elementary school with no traditional teachers: Alpha School plans a high-tuition campus where AI handles core instruction and adults act as guides, a model drawing both political support and skepticism over evidence, surveillance, and student development. Source: Block Club Chicago
• EchoMark pitches AI steganography to trace video game leaks: At GDC, the company showed tooling that subtly alters text and screen output to create recipient-specific fingerprints that AI can later identify in leaked materials. Source: PC Guide
• Lemonade 10.0.1 makes AMD Ryzen AI NPUs easier to use on Linux: The new release simplifies setup, improves hardware detection, and reportedly delivers a modest throughput gain for Linux users working with Ryzen AI hardware. Source: Phoronix
• A developer blog captures the small-model deployment gap: One builder describes fine-tuning a narrow Gemma-based model for Bento voice commands, only to find that hosting and distribution economics remain harder than training itself. Source: Field Notes
• ATLAS claims strong coding results from a consumer GPU setup: An open-source pipeline shared on GitHub says it can push a 14B model running on a roughly $500 GPU past Claude Sonnet on a coding benchmark through multi-attempt generation and selection. Source: GitHub
• A researcher put a salvaged Tesla Model 3 computer on a desk and booted it: The hardware hacking writeup shows the car OS running outside the vehicle, with accessible internal services and a path for deeper automotive security research. Source: David Schütz
• The White House showcased a humanoid robot during an AI education event: CNN reports that First Lady Melania Trump introduced Figure’s “Figure 3” robot while discussing AI’s future role in classrooms and child safety. Source: CNN
• An arXiv-linked debate questions whether LLM scaling alone leads to AGI: A Reddit discussion amplified a recent paper arguing that real intelligence requires continuous autonomous learning from experience, not just better next-token prediction. Source: arXiv
• A Reddit post proposes a formal model for how online arguments escalate: The IDDS 2.1 framework maps state transitions from disagreement to dogpiles and introduces concepts like identity activation and moral protective framing. Source: Reddit
• A CACM blog warns that model collapse is already visible on the open web: The argument is that AI-generated content is increasingly contaminating training corpora, making data provenance a more urgent infrastructure problem. Source: Communications of the ACM