Ask The PhD Community

HigherEd AI Daily

May 31 – Europe Recalibrates, and So Must We

Sunday, May 31, 2026

This week’s AI landscape is shaped by a pivotal shift in global governance, as the European Union steps back from its most aggressive AI Act provisions, while new research surfaces on identifying AI-generated text and the growing role of autonomous agents in how information moves across the web.

The Batch (deeplearning.ai) — GOVERNANCE

Europe Pauses Some AI Regulations

The European Union has moved to weaken and delay certain provisions of its landmark AI Act following sustained pressure from businesses and policymakers who argued the law placed European organizations at a competitive disadvantage. The European Parliament and member states reached agreement on amendments that defer enforcement of restrictions targeting applications deemed to carry significant risks, buying regulators and regulated entities more time to work out implementation details.

The amendments do retain and strengthen protections in specific high-harm areas: generation of sexually explicit images of children and non-consensual intimate imagery are explicitly prohibited under the revised framework. This targeted approach signals a shift from broad precautionary regulation toward more selective intervention, a pattern worth watching as similar debates play out in the United States and beyond.

The recalibration reflects a broader tension regulators worldwide are navigating: how to set meaningful guardrails without stalling the institutions, companies, and researchers that need to use AI tools now.

Why it matters for campuses

Institutions that have been waiting on the AI Act as a governance benchmark before finalizing their own campus AI policies should take note. The EU’s willingness to revise its framework mid-implementation reinforces that no single regulatory document will serve as a permanent blueprint; campus governance structures must be built for ongoing revision. Academic leaders developing AI use policies, IRB frameworks for AI-assisted research, and student-facing guidelines should design those frameworks with amendment cycles built in, rather than treating them as one-time documents.

Read More

The Batch (deeplearning.ai) — RESEARCH

Agents Surf the AI-Written Web

AI-driven traffic on the internet nearly tripled in 2025, according to a new report, and the surge is accelerating. In one striking finding, 80 percent of agentic web traffic in 2025 was directed to product search pages, pointing to the early dominance of commercial applications in shaping where autonomous AI systems spend their browsing cycles. Researchers note that 2025’s 80x rise in agentic traffic over the prior year is likely just the beginning of a longer-term structural shift in how information flows across the web.

This development has direct implications for content discovery and research. As AI agents increasingly serve as intermediaries between users and information, the mechanisms by which content is surfaced, ranked, and consumed are changing in ways that are not yet fully understood. The rise of AI-generated content feeding AI-navigated websites creates feedback loops that academic communities have only begun to examine critically.

Why it matters for campuses

Libraries, research services, and information literacy programs are facing an environment where the pathways through which students and faculty find sources are being fundamentally restructured. Academic librarians should begin incorporating agentic search patterns into their instruction; faculty designing research assignments should think carefully about whether traditional source-evaluation frameworks remain sufficient when a growing share of information discovery is mediated by AI. Research computing teams should also expect queries about responsible agentic research tools to increase substantially over the next academic year.

Read More

TLDR AI — TOOLS / ACADEMIC INTEGRITY

Various LLM Smells: How to Recognize AI-Generated Content

A widely circulated analysis catalogues the specific rhetorical and structural signatures that reliably appear in AI-generated text, the patterns that writers and readers are beginning to call “LLM smells.” These include predictable transitions (“In conclusion…”), overuse of bullet lists as a substitute for argumentation, a flattening of voice that avoids genuine specificity, and repetitive structural designs across genres. The piece is notable for its granularity; it does not merely assert that AI writing is detectable but demonstrates the patterns with precision.

The analysis arrives at a moment when detection technology remains unreliable and contested. Turnitin and other academic integrity tools have faced criticism for false positives and inconsistent results across languages and student populations. This human-readable taxonomy of LLM patterns offers faculty and writing instructors an alternative framework, one based on rhetorical analysis rather than probabilistic scoring.

Why it matters for campuses

Academic integrity offices and writing program administrators will find this taxonomy directly useful for professional development sessions with faculty who are struggling to assess AI-assisted student work. Rather than relying solely on software detection, instructors can develop their own close-reading capacity for LLM patterns and build rubrics that reward originality, specificity, and genuine argument, qualities that AI models still handle poorly. The analysis is also a valuable resource for students learning to use AI tools responsibly; understanding what makes AI writing recognizable can help them develop a clearer sense of what authentic scholarly voice looks like.

Read More

TLDR AI — RESEARCH / TOOLS

How Far Behind Are Open Models?

A detailed 17-minute analysis on LessWrong examines the current capability gap between open-source AI models and the leading closed commercial systems. The findings show that open models are currently running approximately four to six months behind their closed counterparts on public benchmarks; the gap was at its narrowest following the release of DeepSeek R1 earlier this year, but has since grown as frontier closed models have continued to advance at a faster pace.

The analysis also raises important questions about benchmark reliability, noting that public leaderboards can be gamed and that real-world capability differences may not map cleanly onto test scores. The author argues that the open/closed gap is real and consequential, even if its precise magnitude is difficult to measure with confidence.

Why it matters for campuses

Institutional decisions about which AI systems to deploy carry real stakes: open models can be run locally, offering data privacy advantages and lower long-term costs, but the capability gap documented here means students and faculty using open-source tools may be working with meaningfully less capable systems than their peers at institutions that have licensed commercial platforms. Research computing officers, CIOs, and faculty senate technology committees considering AI procurement decisions should treat this analysis as useful context for evaluating what the capability cost of open-source deployments currently looks like, and how quickly that calculus may shift as the landscape continues to evolve.

Read More

Tool of the Day

Claude Opus 4.8

Anthropic released Claude Opus 4.8 this week, bringing measurable benchmark improvements alongside two features with direct relevance for educators: adjustable effort controls, which allow users to dial between quick responses and thorough extended reasoning; and dynamic workflows that enable more reliable multi-step task completion. The model also includes a faster, lower-cost mode for routine queries. Educators and researchers who regularly push AI tools through demanding tasks, such as literature synthesis, syllabus design, or grant narrative drafting, will find the effort controls particularly useful for matching model depth to task complexity.

Try it: Upload a set of five recent articles from your research area and ask Claude Opus 4.8 with extended effort to identify emerging themes, note methodological tensions across the studies, and suggest two unanswered questions that could anchor a future literature review. Then run the same prompt on the fast mode and compare the depth of the two responses.

Visit Claude Opus 4.8

Dr. Ali Green

Sources for This Edition

The Batch by Andrew Ng (deeplearning.ai)
TLDR AI Newsletter (tldrnewsletter.com)
TLDR Dev Newsletter (tldrnewsletter.com)
Various LLM Smells (shvbsle.in)
LessWrong: How Far Behind Are Open Models? (lesswrong.com)
Anthropic: Claude Opus 4.8 (anthropic.com)

askthephd.com
|
askthephd.substack.com
|
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green