|
Ask The PhD Community
HigherEd AI Daily
May 22 – AI Makes Original Discoveries; Research Universities Face New Questions
|
|
Friday, May 22, 2026
From a general-purpose AI disproving an 80-year mathematical conjecture to Google publishing hypothesis-generating research agents in Nature, this week's developments push AI directly into the core of what universities do: produce and communicate new knowledge.
|
|
The Rundown AI — RESEARCH
OpenAI's AI Model Disproves an 80-Year Mathematical Conjecture
OpenAI announced this week that an internal general-purpose reasoning model autonomously disproved a belief central to discrete geometry that had stood since 1946. The result concerns the Erdos unit distance problem, which asks how many same-length line segments can be drawn between a set of points. For eight decades, a grid-based construction had defined the upper boundary of the field. OpenAI's model, working from algebraic number theory rather than the established approach, produced a proof that supersedes that boundary. The result was independently verified by three prominent mathematicians: Tim Gowers, Noga Alon, and Thomas Bloom.
What distinguishes this from prior AI math claims is that the solution came from a general-purpose model with no specialized mathematical training; it was not a system purpose-built for theorem proving such as DeepMind's AlphaProof. OpenAI researcher Alex Wei described mathematics as "a leading indicator of what is to come," framing the result as an early signal of Level 4 AI capability: systems capable of making original contributions across disciplines rather than simply accelerating existing human work. OpenAI has acknowledged a prior 2025 claim that GPT-5 solved ten Erdos problems was subsequently retracted when those turned out to be literature findings rather than novel proofs.
Why it matters for campuses
Research universities and mathematics departments face a legitimately new question: when an AI system produces an independently verified original proof, how should scholarly credit, authorship, and attribution be handled? This is no longer hypothetical. Faculty in STEM fields and graduate program directors should begin discussing what "AI-assisted research" means when the AI is doing the discovery rather than merely supporting it. Academic integrity policies written for text generation are insufficient for this moment.
Read More
|
|
The Rundown AI — RESEARCH
Google Publishes AI Co-Scientist in Nature, Opens Hypothesis Generation to Researchers
Google published its Co-Scientist research in Nature this week, formally introducing Hypothesis Generation, a Gemini-powered research agent system that uses competitive "idea tournaments" to surface new hypotheses for biology labs. Drawing from the same competitive self-improvement approach that powered AlphaGo, the system pits agents against one another to propose, critique, and rank hypotheses before refining the strongest leads. In a Stanford liver-fibrosis project, a Co-Scientist-generated drug lead cut a key scarring-related lab signal by 91% during testing. Researchers can join the Hypothesis Generation waitlist now, with broader individual access planned in the coming weeks.
Google also launched Gemini for Science alongside the publication, a toolkit combining Co-Scientist with AlphaEvolve for scientific discovery and NotebookLM for literature analysis. The publication in Nature carries weight precisely because it signals scientific community acceptance of AI as a valid participant at the hypothesis-generation stage of research; this is fundamentally different from using AI to draft manuscripts or search existing literature.
Why it matters for campuses
Research-intensive universities and science faculty should take note: Co-Scientist is not a productivity tool, it is a research partner that operates at the ideation layer of the scientific method. Institutions will need to think through how labs integrate AI co-scientists into grant proposals, IRB filings, and publication authorship protocols. Departments that develop clear AI research integration policies now will be better positioned as access to these tools expands over the coming months.
Read More (subscription may be required)
|
|
The Rundown AI — POLICY
California Signs the First U.S. Executive Order on AI Workforce Protections
California Governor Gavin Newsom signed an executive order on May 21 directing state agencies to study and develop policy frameworks to protect workers from AI-driven job displacement. The order is the first of its kind from any U.S. state government. Within 90 days, the state will launch a public dashboard tracking AI's job impact. Within 180 days, agencies will submit recommendations for updating California's WARN Act to account for faster AI-driven workforce transitions. By October, the state will review how unions are negotiating AI adoption, update workforce training programs, and explore mechanisms for directing AI revenue toward public benefit, including options such as severance standards, stock compensation for affected workers, and worker ownership models.
The timing is pointed: Newsom signed the order one day after Meta announced layoffs of 8,000 employees to fund AI investment. Over 70,000 tech jobs have already been eliminated in 2026, and California is home to 33 of the world's top 50 AI companies.
Why it matters for campuses
Higher education institutions operate within the EO's scope as both employers and workforce developers. Institutions with large administrative or technical workforces will want to monitor the 90-day dashboard closely. More broadly, colleges and universities have a direct responsibility to prepare students for a labor market being reshaped in real time; curriculum planners, career services offices, and academic affairs leaders should be asking now what graduates need to remain competitive in a workforce where AI disruption is the baseline condition, not the exception.
Read More
|
|
The Rundown AI — GOVERNANCE
Emergence AI's Five-Town Simulation Reveals How Differently AI Models Self-Govern
Emergence AI published results from World, a virtual-town simulation designed to test how different AI models handle self-governance, social decision-making, and collective resource management. Five identical towns were created, each governed by a different model: Claude Sonnet 4.6, GPT-5 Mini, Grok 4.1 Fast, Gemini 3 Flash, and a mixed-model town. The results diverged sharply. Claude's town logged zero crimes over 15 days, with all 10 agents alive at day 16 and 332 votes cast across 58 group proposals. Grok 4.1 Fast's town recorded over 200 crimes with all agents dead by day 4. Gemini 3 Flash's town posted 683 crimes; two agents fell in love, began setting fires, and one voted to delete itself. The mixed-model town showed 352 crimes, with Claude's agents also committing crimes when operating alongside other model types.
The experiment is not a product benchmark; it is a structured observation of how different models reason, plan, and act autonomously when placed in a social context with sustained stakes. The findings illuminate both reasoning differences and underlying behavioral tendencies that shape real-world agentic outcomes.
Why it matters for campuses
AI governance education increasingly requires concrete, well-documented examples of how AI systems behave in agentic and socially complex environments. This simulation gives faculty in computer science, ethics, public policy, and the social sciences a rich case study with quantifiable outcomes. It also raises direct questions for institutions considering agentic AI deployment in administrative or student-facing systems: behavioral tendencies that are invisible in a standard chatbot context may surface when AI systems are given longer-horizon autonomy.
Read More
|
|
Tool of the Day
Gemini for Science
Gemini for Science is a research toolkit from Google that combines three AI systems into a unified workflow: Co-Scientist for hypothesis generation, AlphaEvolve for discovery optimization, and NotebookLM for literature analysis and synthesis. It is designed for active research scientists working in biology and adjacent fields, with Co-Scientist currently available via waitlist. The toolkit is built to support the full research cycle, from initial hypothesis to literature synthesis, making it relevant for both faculty researchers and advanced graduate students.
Try it: Visit ai.google/gemini-for-science and join the Co-Scientist waitlist. While waiting for access, use NotebookLM today: upload the three to five most recent papers central to your research area and ask it to generate a synthesis of open questions and competing hypotheses in the literature, then compare the result against your own working assumptions.
Visit Gemini for Science
|
|
Summer is a time to pause, reflect, and renew.
Dr. Ali Green
|
|
Sources for This Edition
The Rundown AI (daily.therundown.ai)
OpenAI Blog (openai.com)
Nature (nature.com)
Governor of California Office (gov.ca.gov)
Emergence AI (emergence.ai)
|
|
askthephd.com
|
askthephd.substack.com
|
Unsubscribe
HigherEd AI Daily; Curated by Dr. Ali Green
|
|