Ask The PhD Community

HigherEd AI Daily

June 4 – When AI Tutors Outperform Faculty

Thursday, June 4, 2026

Today's edition covers a landmark Stanford study showing AI systems outperforming law professors in blind evaluations, a new open-weight language model designed to run on standard campus laptops, and a growing conversation about whether AI token costs are outpacing their return on institutional investment.

The Rundown AI — Research

Stanford Study: AI Systems Outperform Law Professors in Blind Tutoring Evaluations

A new study out of Stanford Law School has produced one of the most direct head-to-head comparisons of AI and human academic expertise to date. Sixteen law professors from fourteen institutions evaluated 2,918 anonymized matchups between AI-generated responses and answers written by their faculty peers to real contract-law office-hours questions. The professors preferred the AI responses 75 percent of the time, with only a single top-ranked professor holding level against the models.

Google's Gemini 2.5 Pro led the human evaluations at 75.92 percent, followed closely by Google's NotebookLM at 74.75 percent. When the research team extended the evaluation using an AI-based judge and tested nine additional systems, Claude Opus 4.7 ranked first overall; all models tested outperformed the human professors. Critically, faculty flagged AI responses as likely to harm student learning only 3.5 percent of the time, compared to an average of 12 percent for their colleagues' answers.

The researchers deliberately chose contract law because it demands judgment and tolerance for ambiguity rather than retrieval of a single correct answer. That the AI systems performed well precisely in this kind of nuanced environment distinguishes this study from earlier benchmarks that relied on multiple-choice or closed-domain tests.

Why it matters for campuses

This study has direct implications for how law schools and professional programs think about AI-assisted instruction. On-demand tutoring has always been unevenly distributed across institutions; students at well-resourced schools get more office hours access and more timely feedback. If AI systems can reliably provide high-quality academic support at scale, campuses that deploy them thoughtfully stand to close that gap. Academic leaders should approach these findings not as a challenge to faculty roles but as evidence that AI tutoring deserves serious, structured piloting alongside human instruction.

Read More

TLDR AI — Tools

Google's Gemma 4 12B Brings a Capable Open-Weight Model to Standard Laptops

Google has released Gemma 4 12B, a new open-weight multimodal language model designed to run on consumer hardware with as little as 16GB of RAM. Google reports that the model performs at nearly the level of its larger Gemma 4 26B variant, with efficiency gains attributed to a new approach to multimodality. It is the first Gemma model at this size class built for native audio, and it can be downloaded through Kaggle and Hugging Face.

The significance is practical: most faculty and student laptops purchased in the last two to three years with 16GB of RAM could run this model locally, without requiring a cloud API, subscription account, or institutional data-sharing agreement. For researchers handling sensitive data or institutions in the middle of AI governance reviews, a locally hosted, open-weight model changes the calculus around what AI tools are available right now.

Why it matters for campuses

Instructional technology teams and faculty researchers who have been waiting on institutional AI agreements have a new option worth evaluating. Running a capable model locally eliminates data-privacy concerns tied to cloud services and puts AI experimentation within reach of departments that cannot yet access centrally managed tools. Instructional designers and faculty who want to prototype AI-assisted assignments or research workflows can do so without waiting for enterprise contracts to clear.

Read More

TLDR Founders — Governance

The AI Token Reckoning: Only 18 Percent of Spending Yields Productive Output

AI token costs are rising sharply, and organizations across sectors are beginning to reassess their commitments. According to EntelligenceAI, only 18 percent of current AI spending results in measurably productive output. Companies including Uber have begun flagging the difficulty of justifying AI expenditures without demonstrable productivity gains; some organizations have scaled back or canceled AI initiatives outright in response to the mismatch between cost and return.

The analysis draws a distinction between AI investment that supports structured, well-defined workflows and the broader deployment of models as a general-purpose productivity layer. The former continues to show returns; the latter is where costs accumulate without clear accountability.

Why it matters for campuses

Higher education institutions investing in AI infrastructure, enterprise licenses, or AI-enhanced platforms should take note. Budget cycles that were built around 2024 projections may not account for the current trajectory of token pricing. Academic technology leaders and provosts reviewing AI spending will benefit from distinguishing between use cases that have demonstrated clear value, such as research support, accessibility tools, and writing assistance, and broader deployments where the productivity case remains unproven. A period of more focused, measurable AI investment is likely healthier for the sector than continued sprawl.

Read More

Tool of the Day

NotebookLM

NotebookLM is Google's AI-powered research and study assistant that lets users upload documents, PDFs, lecture notes, and readings, then ask questions, generate summaries, and create study guides directly grounded in those sources. Today's Stanford study ranked it second overall among AI tutoring systems, outperforming law faculty in blind evaluations 74.75 percent of the time; it is free to use and requires only a Google account, making it accessible to students and faculty at any institution.

Try it: Upload your course syllabus and three or four assigned readings to a new NotebookLM notebook, then ask it to generate a study guide that connects the key concepts across all sources. Compare what it produces to how you might have framed the synthesis yourself.

Visit NotebookLM

Have a great learning day!

Dr. Ali Green

Sources for This Edition

The Rundown AI (daily.therundown.ai)
TLDR AI (tldrnewsletter.com)
TLDR Founders (tldrnewsletter.com)
Stanford Report (news.stanford.edu)
Ars Technica (arstechnica.com)
Big Technology (bigtechnology.com)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green