Ask The PhD Community

HigherEd AI Daily

September 7 – Data Privacy, AGI Governance, and Domain-Specific AI for Higher Ed

Monday, September 7, 2026

This edition examines how contested data retention commitments from frontier AI providers, emerging calls for democratic oversight of increasingly capable AI systems, and the arrival of domain-specific models built for law and research together require higher education institutions to move from AI curiosity to campus-level strategy.

The Batch (deeplearning.ai) — Policy

What OpenAI and Anthropic’s Data Retention Policies Actually Promise, and What They Do Not

Both Anthropic and OpenAI updated their data handling commitments for business customers this week, but the details reveal a gap between marketing language and verifiable technical guarantees. Anthropic, which had required enterprise customers using Claude Fable 5 to allow 30-day data retention since June, announced a new program called Enterprise Frontier Safeguards (EFS) that will shift storage responsibility to the customer’s own servers or designated cloud providers; the 30-day retention window remains. OpenAI, meanwhile, reaffirmed zero data retention for business customers using its most capable models and previewed Private Safety Processing, a system designed to detect misuse across many requests without OpenAI staff reading the content.

A key tension runs through both announcements: neither company fully explains how automated safety systems can scan data that company employees supposedly cannot see. As The Batch notes, software the company writes must read and process data to scan it; "our employees can’t see it" is not the same as "our systems can’t see it." Neither company has published an independent technical audit of how this works in practice. OpenAI has promised a technical paper on its system by September; Anthropic has offered limited technical detail on EFS since its June announcement.

For institutions in regulated industries, including law schools, medical schools, hospitals affiliated with universities, and financial aid offices, the distinction between "we will not train on your data" and "we do not retain your data" carries significant legal and compliance weight. A prior New York Times copyright suit resulted in a court ordering OpenAI to preserve chat logs that would ordinarily have been deleted; zero-retention customers were reportedly unaffected, but the question of whether logs held on a customer’s servers could be reached in a suit against an AI company remains unresolved.

Why it matters for campuses

Universities deploying AI tools for research, advising, and administration need more than vendor promises; they need auditable technical guarantees that satisfy IRB requirements, HIPAA and FERPA obligations, and data governance policies. Procurement and general counsel offices should request documentation from AI vendors on exactly how safety monitoring works, what triggers human review, how long data persists under any scenario, and what legal processes could compel retention or disclosure. Institutional risk committees should flag that these policies are evolving; what is promised today may change as frontier labs revise their terms.

Read More

TLDR AI — Governance

OpenAI Publishes Progress on Recursive Self-Improvement and Calls for Democratic Governance of AGI

OpenAI released new details on its progress toward Artificial General Intelligence, specifically around Recursive Self-Improvement (RSI), a capability in which AI systems enhance their own performance autonomously. In an extended publication the company describes as a transparency effort, OpenAI argues that AGI must be democratically governed to benefit all of humanity; it states that an informed public debate about AI capabilities, risks, and safeguards is a prerequisite for responsible development. The company has committed to continuing public disclosures of its RSI progress and has called for other AI developers to be required by law to report their own progress on the same metrics.

This publication comes at a moment when the frontier AI development timeline has become a subject of serious academic and policy debate. Researchers across multiple disciplines, from computer science and philosophy to economics and international relations, are grappling with questions about what democratic governance of AI actually means in practice and which institutions are equipped to provide it. OpenAI’s framing notably elevates the public interest dimension of AI development; whether that commitment translates into substantive accountability mechanisms remains to be seen.

Why it matters for campuses

Faculty in political science, philosophy, law, public policy, and AI ethics have a rich and timely case study in OpenAI’s publication: a major private company invoking democratic values while self-reporting on a technology with no agreed governance framework. For institutions developing AI literacy curricula or research programs in AI governance, this moment illustrates the need for interdisciplinary capacity; the technical, legal, and political dimensions of AGI governance cannot be addressed by any single department. University policymakers should also consider how their institutions might contribute to or participate in the public deliberation OpenAI invokes.

Read More

The Batch (deeplearning.ai) — Research

Thomson Reuters Builds a Domain-Specific AI for Law, Finance, and News, and Releases a Free Version for Academic Use

Thomson Reuters announced Thomson, a proprietary large language model family purpose-built for legal, business, tax, finance, and news work. The large version, Thomson-1.0-Large, is built on a Qwen base model and trained on a curated 200-billion-token corpus drawn from Thomson Reuters’ extensive proprietary data holdings: case law, regulatory filings, news, contracts, and synthetic professional task examples. In independent benchmarks, it narrowly outperformed GPT-5.4 and Claude Sonnet 5 on completeness and factuality for tax and legal content. The model will be deployed first within CoCounsel Legal, Thomson Reuters’ AI-assisted legal research and drafting product, which claims over one million users.

Critically for higher education, Thomson Reuters has committed to releasing Thomson-1.0-Small, a 35-billion-parameter version built on Qwen 3.6-35B, as an open-weights model on Hugging Face under an academic and non-commercial license. The company invested roughly $40 million and three months of training to develop the family; it used a data engineering pipeline developed with partner DatologyAI to select a 200-billion-token training set from a 19-trillion-token candidate pool. The constitution guiding the model’s alignment, including a commitment to journalistic objectivity, is open source and adaptable for similar projects.

Thomson Reuters positions Thomson as a form of "corporate sovereign AI," meaning institutions can run it on their own hardware so that sensitive client or research data never leaves their systems. The company plans to allow customers to fine-tune Thomson further on their own proprietary data using the same mid-training methods, without requiring the cost of full pretraining.

Why it matters for campuses

Law schools, business schools, and journalism programs now have a concrete, freely available AI model built specifically for their professional domain; Thomson-1.0-Small will be available for academic researchers to study, fine-tune, and deploy without licensing fees. Beyond individual schools, the Thomson model illustrates a broader trend toward institutional ownership of domain-adapted AI: universities with significant proprietary research corpora, clinical records, or policy archives might find the mid-training methodology described in this paper directly applicable to building their own specialized tools. Research librarians, faculty who work with sensitive data, and technology transfer offices should take note of the data governance model Thomson describes as well as the open-source constitution, which may serve as a template for AI alignment work at the institutional level.

Read More

Tool of the Day

Spec-Driven Development with AI Agents (DeepLearning.AI and JetBrains)

This free short course, developed by DeepLearning.AI in partnership with JetBrains and taught by Paul Everitt, teaches a systematic method for directing AI coding agents: writing a project constitution that captures architecture and style guidelines, drafting a feature spec for each change, and running a plan-implement-verify loop. The course addresses a gap in most AI coding tutorials, which stop at "give the agent better context," by offering a repeatable professional workflow. It is directly relevant to faculty in computer science, data science, software engineering, and any discipline where students are learning to work alongside AI tools rather than simply prompting them.

Try it: Enroll free and use the plan-implement-verify loop to have an AI agent scaffold a simple course feedback form or grade calculator; write the feature spec yourself first, then compare how the agent’s output changes when it has a structured spec versus a freeform prompt.

Visit DeepLearning.AI

Have a great learning day!

Dr. Ali Green

Sources for This Edition

The Batch (deeplearning.ai) — September 4, 2026 issue: "Inside Key Changes in Data Policies, Ox Alpha Revealed, Taking Custom Models Beyond Fine-Tuning" (deeplearning.ai)
TLDR AI — September 7, 2026 issue: "OpenAI research automation, Tim Cook’s role, agent secrets" (tldrnewsletter.com)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green