Ask The PhD Community

HigherEd AI Daily

September 5 – AI Data Governance Takes Center Stage

Friday, September 5, 2026

Today's edition covers three developments that intersect directly with how campuses procure, govern, and build with AI: competing data retention frameworks from the industry's two leading labs, a landmark acquisition that reshapes the open-source model ecosystem, and a domain-specific model release with an explicit academic access tier.

The Batch (deeplearning.ai) — GOVERNANCE

OpenAI and Anthropic Revise Data Retention Policies: Institutions Must Read the Fine Print

Anthropic and OpenAI each announced significant changes to how they handle business customer data this week, and higher education institutions currently evaluating or deploying either company's frontier models will need to study both carefully. Anthropic's existing policy has required enterprise customers using Claude Fable 5 to accept 30-day conversation retention since June. The company previewed a forthcoming program called Enterprise Frontier Safeguards (EFS), which will allow customers with zero data retention requirements to instead store that data on their own servers, on third-party cloud infrastructure such as Google Cloud, Microsoft Foundry, or Amazon Web Services, and to manage their own encryption keys and audit logs. Until EFS is available, eligible enterprise customers may use Fable 5 without Anthropic retaining any data.

OpenAI, meanwhile, reaffirmed that approved enterprise customers can access its most capable models with zero data retention: the company says it does not log prompts or responses for those customers. OpenAI also previewed Private Safety Processing (PSP), a system designed to detect misuse patterns across many requests without requiring any human to read individual conversations. Both companies describe essentially the same monitoring mechanism: automated software that scans for behavioral patterns over time. Neither has yet published a technical paper explaining how that system reads data the company claims its staff cannot access; OpenAI has promised documentation in September, and Anthropic's EFS technical details remain sparse.

The Batch analysis draws a distinction directly relevant to higher education: "we won't train on your data" and "we don't have your data" are materially different promises. Anthropic's June policy change forced exceptions to existing zero-retention contracts, causing many businesses and the ARC Prize Foundation to decline use of Fable 5 entirely rather than expose proprietary or sensitive data to uncertain retention conditions. The forthcoming EFS framework is a direct response to that resistance.

Why it matters for campuses

Universities operating in regulated fields, including law, medicine, student records governed by FERPA, and federally funded research, must treat these evolving policies as active compliance questions, not vendor marketing. Before deploying Fable 5 or comparable frontier models at scale, provosts, general counsel, and CIOs should verify which specific data retention standard applies to their contract, whether an independent audit of EFS or PSP design is available, and what protections govern conversation data if either company faces a legal hold or court order. Neither company has addressed whether data stored on a customer's servers could be reached in litigation targeting the AI vendor itself.

Read More
(subscription may be required)

TLDR AI — RESEARCH

Nvidia Acquires Hugging Face for $12.93 Billion: Open Research Infrastructure Enters New Ownership

Nvidia confirmed this week that it has completed the acquisition of Hugging Face, the platform hosting more than three million AI models and serving over 18 million developers worldwide, for $12.93 billion. CEO Jensen Huang stated publicly that the platform will remain open to all developers and that Nvidia hardware will not be a requirement for building or deploying on it. The acquisition represents one of the largest consolidations in the open-source AI ecosystem to date and marks Nvidia's most significant step beyond its role as a hardware supplier toward operating as a full-stack AI platform provider.

Hugging Face occupies a position in AI research comparable to what GitHub occupies in software development: it is the primary repository where academic researchers access pretrained models, share datasets, publish model cards, and distribute work to the broader community. The platform's commitment to open access and hardware-agnostic deployment has made it a foundational resource across university computer science, linguistics, social science, and biomedical research programs. Its role is also relevant to today's third story: Thomson-1.0-Small, the academic open-weights model from Thomson Reuters, is being distributed via Hugging Face.

Why it matters for campuses

Institutional research programs that depend on Hugging Face for model access, dataset sharing, and reproducible AI workflows are now operating on infrastructure owned by a major commercial hardware vendor. Jensen Huang's pledge to remain open and hardware-agnostic is a reassuring signal, but it does not constitute governance. University research computing offices and library systems should assess which models and datasets hosted on Hugging Face are integral to ongoing projects, evaluate whether institutional data sharing agreements need to be revisited under new ownership, and consider whether archival copies of critical models and datasets should be maintained independently to ensure long-term research continuity.

Read More

The Batch (deeplearning.ai) — TOOLS / RESEARCH

Thomson Reuters Launches a Domain-Specific LLM with an Open Academic License

Thomson Reuters announced "Thomson," a proprietary large language model family purpose-built for law, business, tax, finance, and news. The flagship model, Thomson-1.0-Large, is built on Qwen3.5-397B-A17B and retrained using a 200 billion token domain corpus curated from a 19 trillion token candidate pool, drawing in roughly equal parts from proprietary documents including case law, regulatory filings, and news content; synthetic professional task pairs; and general-capability material. In independent evaluations, Thomson-1.0-Large narrowly outperformed both GPT-5.4 and Claude Sonnet 5 on factuality and completeness in legal, tax, and news domains. It will be deployed initially inside CoCounsel Legal, Thomson Reuters' AI assistant for legal research and document drafting, with rollout to other products at unspecified future dates.

Of particular note for the academic community: Thomson Reuters is releasing Thomson-1.0-Small, a 35 billion parameter open-weights version, on Hugging Face under a license permitting academic and non-commercial use. The company is also open-sourcing its midtraining pipeline and the model's alignment constitution, which encodes principles including journalistic objectivity and is available for institutional adaptation. The company invested a reported $40 million over three months in training, using a method it calls Continual Learning, a mix of full-weight mid-training and fine-tuning. The company's research framing positions the approach as "corporate sovereign AI": institutions license and run the model on their own hardware, with sensitive data never leaving their environment.

Why it matters for campuses

Law schools, schools of business, journalism programs, and policy research centers have an openly licensed and domain-tested model in Thomson-1.0-Small to experiment with today. More broadly, the "sovereign AI" framing describes something universities have concrete reasons to pursue: models customized on institutional knowledge, hosted on-premises, and governed locally rather than subject to the shifting data retention and access conditions described in today's first story. The open midtraining pipeline makes this replicable for institutions with adequate research computing resources, and the open constitution provides a concrete template for thinking through AI alignment at an institutional level.

Read More
(subscription may be required)

Tool of the Day

MAI-Transcribe-2 (Microsoft)

MAI-Transcribe-2 is Microsoft's new speech recognition model that transcribes audio in 60 languages at a reported price of $0.10 per hour, with word-level timestamps, speaker diarization, and configurable transcription styles. Microsoft says the model outperforms Gemini 3.5 Transcribe, GPT-Transcribe, and Whisper V3-Large on accuracy and processing speed. For faculty conducting qualitative research, managing archives of recorded lectures, or teaching across multilingual student populations, this model offers a meaningful reduction in both transcription cost and turnaround time.

Try it: Pull an audio recording from a recent research interview, panel discussion, or guest lecture; run it through MAI-Transcribe-2 via the Microsoft AI API; and compare the speaker-labeled transcript against your current transcription workflow, paying particular attention to accuracy on technical vocabulary and any multilingual segments.

Visit MAI-Transcribe-2

Have a great learning day!

Dr. Ali Green

Sources for This Edition

The Batch by DeepLearning.AI (deeplearning.ai)
TLDR AI (tldrnewsletter.com)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green