Ask The PhD Community

HigherEd AI Daily

September 6 – Data Policies, Open Ecosystems, and Domain AI

Sunday, September 6, 2026

Three developments this week demand close attention from campus leaders: the diverging approaches of OpenAI and Anthropic to enterprise data retention, a major acquisition reshaping the open AI model ecosystem, and the arrival of a domain-specific model purpose-built for legal and financial knowledge work.

The Batch (deeplearning.ai) — GOVERNANCE

Anthropic and OpenAI Draw Different Lines on Enterprise Data Retention

This week brought two distinct responses to a question every institution using AI tools should be asking: where does your data go, and for how long does it stay? Since June, Anthropic has required business customers using its Claude Fable 5 model to allow the company to retain conversation data for 30 days. That policy alienated enterprise customers with zero data retention (ZDR) contracts and caused several organizations, including the ARC Prize Foundation, to decline using the new models rather than expose private information.

Anthropic has since announced a forthcoming program called Enterprise Frontier Safeguards (EFS). Under EFS, businesses with ZDR policies would retain data for the same 30 days but store it on their own servers or those of approved cloud providers such as Google Cloud, Microsoft Foundry, or Amazon Web Services. Customers would manage their own encryption keys and audit logs. Until EFS launches this fall, eligible enterprise customers can use Fable 5 and Fable 5.1 without Anthropic retaining any data. OpenAI, by contrast, reaffirmed zero data retention for approved business customers and previewed Private Safety Processing (PSP), a system designed to detect misuse patterns across requests without OpenAI reading the underlying prompts. Both companies describe automated monitoring but neither has published independent audits or full technical specifications.

The Batch's analysis makes a critical distinction that institutions should internalize: "we will not train on your data" and "we do not have your data" are meaningfully different promises, carrying different implications for legal discovery, security breach exposure, and regulatory compliance.

Why it matters for campuses

Higher education institutions operate under FERPA, IRB protocols for research data, HIPAA on medical campuses, and state privacy statutes. For campus leaders evaluating AI vendors, the difference between these data policies has material consequences for compliance, legal discovery risk, and student trust. Institutions currently running AI pilots or expanding deployments should review their data processing agreements with both vendors before committing to broader rollouts, particularly in contexts involving student records, unpublished research, or clinical information.

Read More

TLDR AI — RESEARCH

Nvidia Acquires Hugging Face for $12.93 Billion, Reshaping the Open AI Model Ecosystem

Nvidia confirmed this week that it has acquired Hugging Face for $12.93 billion. Hugging Face is the central hub of the open-source AI model ecosystem, hosting more than three million models and serving upwards of 18 million developers worldwide. The acquisition makes Nvidia the owner of the platform that most academic AI researchers use to access, share, and fine-tune open-weight models.

Jensen Huang stated that the Hugging Face platform will remain open and that Nvidia compute will not be required to build or deploy on it. The assurance is notable, but the change in ownership still introduces new considerations for institutions and researchers who have treated Hugging Face as a neutral infrastructure layer. The platform has served as the primary distribution channel for foundational research models, fine-tuned academic variants, and datasets that power a wide range of scholarly inquiry.

Why it matters for campuses

A substantial portion of AI research at universities, including work on bias, fairness, domain adaptation, and low-resource languages, depends on open-weight models hosted on Hugging Face. Faculty, graduate students, and research computing teams that access these models for teaching and scholarship now operate within an infrastructure owned by a hardware company with significant commercial interests in AI deployment. Campus research officers and library professionals should monitor how access terms, governance policies, and model availability evolve under new ownership. The transition also signals a broader pattern worth discussing in technology policy courses: the open AI infrastructure layer is consolidating rapidly, and academic communities have limited formal representation in those decisions.

Read More

The Batch (deeplearning.ai) — TOOLS

Thomson Reuters Launches Domain-Specific AI for Law, Finance, and News, With Open Weights for Academic Use

Thomson Reuters released Thomson, a family of domain-specific large language models built for law, finance, tax, and news. The flagship model, Thomson-1.0-Large, is built on a Qwen3.5 base and fine-tuned using a 200-billion-token dataset curated from proprietary legal filings, case law, contracts, regulatory documents, and news content. The company reports that Thomson-1.0-Large outperforms GPT-5.4 and Claude Sonnet 5 on completeness and factuality for tax, legal, and news tasks. Thomson Reuters invested $40 million in training over three months and partnered with DatologyAI for dataset curation. A smaller version, Thomson-1.0-Small at 35 billion parameters, will be released as open weights on Hugging Face for academic and non-commercial use.

The model's alignment constitution, which encodes journalistic objectivity and professional domain standards, is also being released as open source and made available for adaptation. Thomson Reuters frames its approach as a form of corporate sovereign AI: rather than yielding sensitive data to a larger AI company, institutions license the model and run it on their own hardware, ensuring that sensitive client or research data never leaves controlled environments. The company says future versions may incorporate further proprietary data and a revised base model.

Why it matters for campuses

For law schools, business schools, journalism programs, and faculty conducting research in these disciplines, Thomson-1.0-Small offers a free, academically licensed model that may perform meaningfully better than general-purpose alternatives on field-specific tasks. Institutions considering AI tools for student legal research, regulatory analysis, or newsroom training applications have a new foundation to evaluate. The case also reinforces a point worth raising in AI literacy and curriculum discussions: model selection and training data curation matter considerably, and a smaller model trained on the right data can outperform a much larger general-purpose one on the tasks that define professional knowledge work.

Read More

Tool of the Day

Microsoft MAI-Transcribe-2

MAI-Transcribe-2 is Microsoft's new speech recognition model, priced at ten cents per hour of audio. It supports 60 languages and includes speaker diarization, configurable transcription styles, and word-level timestamps. According to Microsoft, it outperforms competing transcription services from OpenAI, Google, and ElevenLabs on both accuracy and speed, making it a compelling option for educators who need to transcribe lectures, interviews, or recorded discussions at scale and at low cost.

Try it: Upload a recorded lecture or set of research interviews and use MAI-Transcribe-2 to generate a word-level, speaker-labeled transcript; then bring that transcript into a text analysis tool or share it with students as an accessible companion to your audio or video content.

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)
NVIDIA Blog (blogs.nvidia.com)
Microsoft AI (microsoft.ai)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green