Ask The PhD Community

HigherEd AI Daily

July 12 – AI Governance Takes the Wheel

Sunday, July 12, 2026

This week's AI headlines circle back to a single question for higher education: who gets to decide what these systems are allowed to do, and how much say do campuses have in that decision.

The Batch — GOVERNANCE

Anthropic Restores Claude Fable 5 After a Government-Ordered Shutdown

Anthropic's advanced Claude Fable 5 and Mythos 5 models went dark for nearly three weeks after the U.S. Department of Commerce imposed export controls on June 12, following a report that Amazon researchers had found a way to prompt Fable 5 into producing cyberattack-related code. With no reliable way to verify user nationality in real time, Anthropic suspended access for everyone rather than risk violating the order.

Access was restored July 1 with new safeguards; a cybersecurity classifier now blocks the reported bypass technique in more than 99% of cases, routing flagged requests to the smaller Claude Opus 4.8 instead. Anthropic is also working with Amazon, Microsoft, Google, and other partners on a shared industry framework for scoring the severity of AI jailbreaks, and has committed to giving federal evaluators earlier access to future frontier models before release.

Why it matters for campuses

This episode is a preview of what campus IT and legal offices should expect going forward: national security concerns can pull a widely adopted AI tool offline with almost no warning, disrupting research and coursework built around it. Institutions relying on a single frontier model for grading pipelines, research computing, or student-facing tools should have a fallback plan, and administrators tracking AI policy should watch the emerging jailbreak-severity framework as a possible template for campus AI risk classification.

Read More

The Batch — RESEARCH

A Brain-Reading System Now Turns Thought Into Text, No Surgery Required

Researchers at Meta, the Basque Center on Cognition, Brain and Language, and several French research institutions unveiled Brain2Qwerty v2, a system that decodes full sentences from non-invasive magnetoencephalography (MEG) brain scans. Trained on roughly 22,000 sentences from nine volunteers, the pipeline pairs a signal-decoding model with a fine-tuned language model to reconstruct what a person intended to type. Word accuracy reached 61% on average and 78% for the best-performing participant, a sharp improvement over the prior version.

The system targets people who lose the ability to speak or type after a stroke, accident, or neurological disorder, and it works without the brain implants that current neuroprosthetic approaches require. Meta and its research partners are releasing the training code, and a $5 million fund is being set aside to seed more open datasets in this area.

Why it matters for campuses

This is squarely relevant to disability services offices, neuroscience and rehabilitation programs, and any institution weighing how assistive AI reshapes accessibility obligations. It also doubles as a useful teaching example of AI research that is genuinely early stage; word-error rates remain too high for daily use, a good reminder for students learning to distinguish promising research from deployable technology.

Read More

TLDR AI — POLICY

OpenAI's Copyright Defense Runs Into a Credibility Problem

News organizations suing OpenAI over the use of copyrighted material in training data allege the company misled the court for roughly two years about whether it could search its own ChatGPT logs. According to the filing, a former OpenAI privacy engineer's account indicates the company had already conducted such searches before the litigation began, despite representing to the court that doing so was not feasible. The discovery dispute could expose OpenAI to sanctions.

The case sits within a broader wave of copyright litigation against AI developers, but the alleged discovery misconduct adds a new dimension; courts are now weighing not just whether training on copyrighted work is lawful, but whether AI companies are dealing honestly with the legal process itself.

Why it matters for campuses

Faculty in law, library science, journalism, and media studies have a live case study in front of them, and general counsel offices should note that AI vendor representations about data retention and searchability are now facing direct legal scrutiny. Any campus with active AI vendor contracts may want to revisit what those agreements actually promise about data handling.

Read More

TLDR AI — TOOLS

GPT-5.6 Lands With Three Tiers, Reshaping the Cost-Capability Tradeoff

OpenAI began rolling out its GPT-5.6 family on July 9: Sol, the flagship model built for coding, cybersecurity, and scientific reasoning; Terra, a mid-tier model priced at half of GPT-5.5 with comparable performance; and Luna, a low-cost model for everyday tasks. Sol ships with what OpenAI calls its most robust safety stack yet, including layered misuse classifiers and roughly 700,000 GPU-hours of automated red-teaming aimed at finding so-called universal jailbreaks.

Because the three tiers carry meaningfully different pricing, from $1 to $30 per million tokens depending on the tier and direction, the release effectively hands institutions a menu of capability-versus-cost tradeoffs rather than a single upgrade.

Why it matters for campuses

Procurement and IT decisions about which AI tools to license for coursework or research computing increasingly hinge on tier selection, not just vendor selection. Faculty who built assignments or research workflows around a specific model's behavior should expect that behavior to shift as institutions decide which tier to adopt.

Read More

Tool of the Day

ChatGPT Work

ChatGPT Work, powered by GPT-5.6, is OpenAI's new agent mode for longer, multi-step projects. Give it a goal and it drafts a plan, gathers context from approved files and connected tools, works through the steps, and pauses for review or approval before finishing. OpenAI's own education newsletter highlights concrete campus use cases: refreshing a course against accessibility guidance and learner feedback, assembling a program review or accreditation packet, building a source-backed research brief, and coordinating operational work across teams.

Try it: Give ChatGPT Work a single, well-scoped task you already understand well, such as comparing your current syllabus against updated accessibility guidance and learner feedback, and review the plan it proposes before approving any step.

Visit ChatGPT Work

Have a great learning day!

Dr. Ali Green

Sources for This Edition

The Batch (deeplearning.ai)
TLDR AI (tldrnewsletter.com)
Anthropic (anthropic.com)
OpenAI (openai.com)
Meta AI Research (facebookresearch.github.io)
Ars Technica (arstechnica.com)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green