Ask The PhD Community

HigherEd AI Daily

June 7 – Policy, Copyright, and the New Research Productivity Standard

Sunday, June 7, 2026

Three signals converged this week for higher education leaders: a new federal AI executive order shaping how frontier models are governed, research showing that fine-tuned writing tools can surface copyrighted text verbatim, and data confirming that AI is fundamentally reshaping research productivity at institutional scale.

The Batch (deeplearning.ai) — POLICY

White House Issues Executive Order on Frontier AI Models

The White House issued an executive order this week providing new governance guidance for companies that build frontier AI models. The order promotes continued AI development while establishing a framework to address national security concerns; specifically, the risk that advanced models could be used to automate the detection of software vulnerabilities at scale. The order was shaped significantly by findings related to Anthropic's Mythos model, which demonstrated enhanced capabilities in identifying code vulnerabilities across complex software systems.

The order asks frontier labs to voluntarily share their models with the federal government and collaborate on cybersecurity efforts. Notably, it does not impose hard liability requirements on model builders for third-party misuse, a provision that AI development advocates had warned would stifle innovation. Andrew Ng, writing in The Batch, described the result as "a reasonable compromise between encouraging AI development and protecting security," while also cautioning against ongoing regulatory overreach driven by speculative risk scenarios.

The order comes amid broader concerns about industrial-scale model distillation, a practice through which the outputs of frontier models are systematically used to train competing systems at low cost, sometimes through unauthorized API access. The White House acknowledged this as an adversarial threat and committed to working with the private sector to build defenses against it.

Why it matters for campuses

Federal AI governance has direct implications for higher education institutions that procure AI tools, manage research computing infrastructure, and train the next generation of AI researchers. Campus technology officers and academic leaders should monitor how frontier labs respond to the voluntary sharing framework and whether this reshapes the availability and pricing of the AI services that universities depend on for research and instruction. Institutions with federal research contracts should pay particular attention to how new cybersecurity requirements filter down to their AI procurement processes.

Read More

The Batch (deeplearning.ai) — RESEARCH

Fine-Tuning for Creative Writing Tasks Causes LLMs to Reproduce Copyrighted Books Verbatim

A new study from researchers at Stony Brook University, Carnegie Mellon University, and Columbia Law School has found that fine-tuning large language models on a specific creative writing task causes them to reproduce substantial portions of copyrighted books, even when those books were not part of the fine-tuning data. The task involved expanding plot summaries into paragraphs matching the style and content of existing novels. Fine-tuned versions of GPT-4o, Gemini 2.5 Pro, and DeepSeek-V3.1 reproduced up to 90 percent of source text verbatim in some cases.

The finding reveals a significant gap in how AI alignment training handles copyright protection. System prompts and human-preference fine-tuning can suppress verbatim output under typical use conditions, but they do not erase memorized text from the model's weights. Fine-tuning that rewards verbatim generation effectively unlocks that memorized content, counteracting safety guardrails. In tests, models generated verbatim spans of up to 440 words from books they had encoded during pretraining; the effect generalized well beyond the specific titles used in fine-tuning.

The researchers measured what they call "book memorization coverage" across dozens of contemporary novels. Fine-tuned on novels by one author and tested on books by dozens of others, all three models produced high rates of verbatim reproduction. The study was conducted in collaboration with Columbia Law School, signaling direct legal relevance; it has not yet been independently replicated.

Why it matters for campuses

This research raises concrete questions for institutions that have deployed AI writing tools for student use, for faculty who use fine-tuned models in creative and scholarly work, and for academic libraries navigating copyright compliance in an AI-assisted environment. Instructional designers and academic integrity officers should monitor developments in this area closely; fine-tuned writing tools may present copyright exposure that standard AI use policies have not yet addressed. General-purpose AI writing tools evaluated as low-risk may behave differently once fine-tuned, a distinction that campus technology assessments may not currently capture.

Read More

TLDR AI — TOOLS

Anthropic Reports 80 Percent of Production Code Is Now AI-Authored

Anthropic reported this week that 80 percent of its new production code is now authored by Claude, its own AI system, resulting in engineers shipping eight times more code than in prior periods. The figure marks a measurable threshold in the shift from AI as a coding assistant to AI as a primary contributor to production software. Anthropic has described this internally as part of a broader strategy of "recursive self-improvement," in which AI systems play a direct role in building their own successors.

The development mirrors a pattern emerging across leading AI organizations, where the practical distinction between AI-assisted and AI-authored output is rapidly narrowing. Anthropic's research arm has published a formal articulation of recursive self-improvement as a near-term phenomenon, noting that the approach already enables autonomous agents to accelerate productivity substantially in research and coding workflows. Internal benchmarks show the system allowing typical engineers to ship eight times more code per period than they did previously.

Why it matters for campuses

Computer science and data science departments face a direct curriculum question: what does software engineering education look like when 80 percent of production code at a leading AI lab is not written by human engineers? Beyond CS, research computing teams at universities should assess how AI-assisted coding could accelerate computational research across disciplines. The figure also raises questions about intellectual contribution standards in computational scholarship; how institutions, disciplinary associations, and academic publishers define and evaluate authorship in code-heavy research is a governance issue that warrants institutional attention now.

Read More

Tool of the Day

QVAC SDK

QVAC is an open-source SDK that runs large language models, speech recognition, image analysis, and translation entirely on a local device, with no cloud connection, no API keys, and no subscription required. After a recent update integrating TurboQuant, QVAC now delivers up to five times more context capacity from the same hardware. For faculty and researchers working under FERPA, HIPAA, or institutional data governance policies, it offers a practical path to using capable AI tools without routing any content through external servers.

Try it: Install QVAC via npm, load a local LLM, and use it to draft written feedback on a set of anonymized student submissions — all student data stays on your machine throughout.

Visit QVAC SDK

Have a great learning day!

Dr. Ali Green

Sources for This Edition

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

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green