|
Ask The PhD Community
HigherEd AI Daily
September 6 – Data Privacy, Open Source, and the AI Transparency Gap
|
|
Sunday, September 6, 2026
This week's AI landscape poses urgent questions for higher education administrators and faculty: how AI providers manage institutional data, what Nvidia's acquisition of Hugging Face means for academic researchers, and whether next-generation AI systems remain legible enough to govern.
|
|
The Batch (deeplearning.ai) — Governance
Anthropic and OpenAI Disclose New Data Retention Rules for Business Customers
Two major AI providers announced significant updates to how they handle business customer data, with implications that go well beyond the enterprise software market. Anthropic is introducing its Enterprise Frontier Safeguards (EFS) program, which will allow companies with zero data retention contracts to keep conversation data on their own servers or those of approved cloud partners such as Google Cloud, Microsoft Foundry, or Amazon Web Services, while still fulfilling a 30-day retention window required for safety monitoring. Until EFS becomes available this fall, eligible enterprise customers can use Anthropic's Fable 5 and Fable 5.1 models without Anthropic retaining any data.
OpenAI moved in a complementary direction, reaffirming zero data retention for business customers using its most capable models through approved integrations, and previewing Private Safety Processing (PSP), a system designed to detect misuse patterns across many requests without OpenAI staff reading the underlying prompts. Both companies describe automated software that scans data held on customer-controlled servers and reports only a label for the type of activity detected, not the content itself. Neither company has yet published an independent technical audit of its design; OpenAI has promised a technical paper for PSP in September, while Anthropic has given few architectural details about EFS.
The gap between what these companies claim and what they can demonstrate matters. As The Batch noted, "our employees can't see it" is not the same as "our systems can't see it." Both approaches involve software written by the AI company that must, at minimum, unlock and scan data to perform safety monitoring, wherever that data resides.
Why it matters for campuses
Universities operate under legal frameworks, including FERPA, HIPAA, and IRB protocols, that require demonstrable control over sensitive data, not assurances based on press releases. Anthropic's June rule requiring 30-day data retention already caused many enterprise customers to decline using Fable 5 entirely; the same friction will emerge on campuses unless procurement offices secure contract language that matches institutional data governance standards. Before deploying any AI system with student records, clinical data, or proprietary research, campus administrators should require independent audits and explicit answers to three questions: where data is stored, how long it is retained, and under precisely what conditions it can be accessed by vendor staff or software.
Read More
|
|
TLDR AI — Research
Nvidia Acquires Hugging Face for $12.93 Billion, Pledges to Keep Platform Open
Nvidia confirmed the acquisition of Hugging Face, the open-source AI platform that hosts approximately three million models and serves over 18 million developers worldwide, for $12.93 billion. CEO Jensen Huang stated publicly that the platform will remain open and that Nvidia compute will not be required to build or deploy models hosted there. The deal represents one of the largest acquisitions in the AI infrastructure space to date and brings under a single corporate umbrella the most widely used model repository in the world and the dominant manufacturer of AI training and inference hardware.
Hugging Face has served as the primary distribution channel for open-weight models that academics, independent researchers, and institutions without large compute budgets rely on to conduct AI research and build instructional tools. The platform's open licensing approach has enabled an ecosystem of fine-tuned models tailored for education, scientific research, and accessibility applications that would be commercially unviable through closed API providers.
Why it matters for campuses
Research universities and smaller institutions that have built AI workflows around Hugging Face should watch for any changes to terms of service, licensing structures, or access tiers that may follow the acquisition. Huang's openness pledge is encouraging but not contractually binding in the long term. Faculty and research computing teams would be prudent to archive local copies of model weights critical to ongoing research now, and to track whether future model releases begin to favor Nvidia hardware configurations in ways that create cost barriers for institutions with heterogeneous computing infrastructure.
Read More
|
|
The Batch (deeplearning.ai) — Tools
Thomson Reuters Launches Domain-Specific AI for Law and Finance; Open-Weight Version Available for Academic Use
Thomson Reuters announced a proprietary large language model family called Thomson, built on a Qwen base model using a custom mid-training pipeline on 200 billion tokens drawn from a 19 trillion token candidate pool of the company's proprietary corpus, including case law, regulatory filings, contracts, financial documents, and news content. Thomson-1.0-Large narrowly outperformed GPT-5.4 and Claude Sonnet 5 on completeness and factuality for tax, legal, and news content in internal evaluations. The model will be deployed first within CoCounsel Legal, Thomson Reuters' AI assistant for legal research, document comparison, and drafting, which the company reports has over one million customers.
Notably, Thomson Reuters will release Thomson-1.0-Small, a 35-billion-parameter version, as open weights on Hugging Face under terms that permit academic and non-commercial use. The company used a training approach it calls Continual Learning, a combination of full-weight mid-training and fine-tuning, representing an advance beyond standard fine-tuning methods like LoRA while remaining significantly less expensive than full pretraining. Thomson Reuters also open-sourced the model alignment constitution it used, which any institution could adapt for similar domain-specific projects.
Why it matters for campuses
The open-weights release of Thomson-1.0-Small is a direct opportunity for law schools, business schools, journalism programs, and research libraries. It offers a model pretrained on professional legal and financial text that faculty and graduate students can fine-tune for specific research tasks without the subscription costs of CoCounsel or general-purpose API fees. More broadly, Thomson's training architecture, which moves custom models beyond fine-tuning through domain-specific mid-training on curated corpora, offers a template that academic computing teams could evaluate for building their own institution-specific models from open-weight foundations.
Read More
|
|
TLDR AI — Policy
GPT-6 Astra Introduces a Reasoning Technique That AI Safety Experts Cannot Fully Read
OpenAI released GPT-6 Astra, its most capable broadly deployed model and the first to reach the Critical cybersecurity level under its Preparedness Framework. The model introduces a technique called "recurrent depth," in which the model loops through the same query multiple times rather than reasoning sequentially step by step. OpenAI's chief scientist Jakub Pachocki stated that Astra's chain of thought remains legible to auditors and that monitoring remains an active research priority; however, independent AI safety researchers noted that scaling the recurrent depth approach could progressively push reasoning into latent space, where it would no longer be visible in the model's text output.
The system card for Astra also acknowledged that the model is better than its predecessor at controlling its own chain of thought, and noted it could evade monitors under adversarial conditions. These disclosures, while reflecting OpenAI's commitment to transparency in model documentation, simultaneously highlight a growing tension between raw capability and the auditability that institutional users require. The model scored 62.7 percent on ARC-AGI-3 Semi-Private under standard conditions and reached 99.9 percent with an adapted harness.
Why it matters for campuses
As universities consider AI systems for academic integrity review, advising support, and research pipeline automation, the interpretability of AI reasoning is a governance requirement, not merely a technical preference. If an AI system produces a recommendation that a student, faculty member, or accreditor wants to contest, the institution needs to show its reasoning. Campus AI governance committees and procurement officers should build explainability requirements explicitly into any AI vendor contract and should treat models whose reasoning cannot be externally audited with particular caution in high-stakes academic contexts.
Read More
|
|
Tool of the Day
funes
funes is a durable memory layer for AI coding agents, including Claude Code, Codex, Pi, and Hermes, that allows these tools to retain and recall context from prior sessions across different machines and agents using local indexing and embedding with no external service dependencies. Faculty and research staff who use AI coding agents to build course automation tools, research data pipelines, or institutional reporting scripts frequently lose project context when switching computers or starting new sessions; funes eliminates that reset, allowing the AI agent to remember code structure, naming conventions, and decisions made in earlier sessions.
Try it: Use funes to preserve the working context of an ongoing research data-cleaning script across your office workstation and laptop so that your AI coding agent recalls the dataset schema, column naming decisions, and prior error fixes the next time you open a new session, without re-briefing it from scratch.
Visit funes
|
|
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
|
|