Skip to main content
Research Office

Issue #10

Welcome to the latest edition of our AI newsletter.

This one opens with our next invited speaker, Prof. Scott P. Simkins (North Carolina A&T State University), who will help us move the conversation beyond cheating and into a more interesting question: how do we design courses where generative AI becomes a partner to human thinking rather than a shortcut around it?

After that, you’ll find a mix of:

  • University-level AI news worth knowing,
  • Broader AI developments that caught my attention (for better or worse),
  • A focused section on AI and education, and
  • A final dive into recent research and scientific work in AI.

This newsletter is a personal effort, meant to share signals, not noise—what I find genuinely interesting, useful, or worth discussing across research, teaching, and academic life. No institutional messaging, no hype for the sake of hype.

As always, read what resonates, skip what doesn’t, and feel free to reply with thoughts, disagreements, or things I should be paying attention to next.

ONE NOTE: Please recommend speakers for 2026—2027 year. I will start inviting speakers by the end of the month. Send the contact information to alromero@mail.wvu.edu

Let’s get into it.

February speaker: Prof. Scott P. Simkins

Department of Economics
North Carolina A&T State University

Title: Beyond Cheating: Designing Courses That Position Generative AI as a Partner to Human Thinking

Abstract:

As generative AI reshapes higher education and the workplace, faculty need concrete strategies for helping students use these tools to enhance - not replace - their thinking. This interactive session introduces a practical course-design framework grounded in Dee Fink’s Taxonomy of Significant Learning and demonstrates how generative AI can support deeper disciplinary learning, improved performance, and stronger real-world skills in your courses. Drawing on examples from my own economics courses, I will share how intentionally structured assignments and scaffolds can help students use AI as a complement to human judgment, analytical reasoning, and creativity. Grounded in recent work by Julie Schell (UT-Austin), the session also introduces the idea of “fusion skills”, the blended human–AI capabilities that produce what Schell describes as emergent capacities: outcomes neither humans nor AI achieve alone. Participants will explore how these fusion skills can be cultivated through course design, faculty development, and reflective student practice.

Upcoming speakers, link for the talk & archives of previous speakers.
WVU AI Discussion Group: upcoming speakers and archives

Zoom link for the meeting (attached is outlook invite):
Join the WVU AI Discussion Group meeting on Zoom

REMINDER: AI SURVEY

I encourage you not only to share your thoughts, but also to pass it along to colleagues and students who may want to contribute.

Responses will be synthesized into a summary report to be shared with this group in March. Survey closes by the end of this week.

👉 Complete the anonymous AI survey feedback form

WVU News

(Please send email to alromero@mail.wvu.edu if you want something to be included in our next newsletter)

Faculty Community: Revisioning How We Teach Research and Writing in The Age of GenAI

West Virginia University
WVU Potomac State College
WVU Institute of Technology
Facilitators: Nathalie Singh-Corcoran, Jenn Monnin & Miranda Smith

This Faculty Learning Community (FLC) will encourage participants to explore foundational practices in writing and research instruction and examine those practices in the context of GenAI. Topics will include: how to teach meaningful writing across the disciplines, how to engage students in critical thinking and deep learning using information literacy frameworks, how GenAI is being integrated across the research and writing lifecycle, and how AI and research mentorship intersect.

Revisioning How we Teach Writing and Research in the Age of AI is suitable for faculty and instructors across the curriculum who teach and mentor undergraduate and graduate students, and we welcome GenAI skeptics, refusers, adopters, and everyone in between. Each of our meetings will consist of a robust discussion and a reading or two (e.g. an article or book chapter) to anchor our conversations.

Those selected for this community will meet on the following Fridays via Zoom:

  • February 13 from noon - 1:15 p.m.*
  • February 27 from noon - 1 p.m.
  • March 13 from noon - 1 p.m.
  • March 27 from noon - 1 p.m.
  • April 10 from noon - 1 p.m.
  • April 24 from noon - 1 p.m.

View the full schedule and apply by Friday, February 6

Faculty Learning Community: Revisioning How We Teach Research and Writing in the Age of GenAI—full schedule and application

AI in Higher Education: From Foundations to the Future

Sponsor: Sage College Publishing & Sarah E. Moore, UT Dallas

Artificial intelligence is transforming higher education, but how can faculty harness its potential effectively, ethically, and responsibly while preserving what works? This four-part workshop webinar series will guide educators from AI fundamentals to real-world applications and future trends. These interactive sessions will provide practical insights, pedagogical strategies, and future-focused discussions. Join us to navigate AI’s impact on academia and prepare for the future of teaching and learning in your classroom today as well as the classroom of tomorrow.

SAGE workshop: AI in Higher Education—full descriptions and registration

Sessions

  • AI in Action: Reimagining Teaching and Student Engagement
    Wednesday, February 25, 2026 at 12:00 PM Eastern
  • AI in the Classroom: Navigating Ethics, Pedagogy, and Practice
    Wednesday, March 25, 2026 at 12:00 PM Eastern
  • AI and the Future of Learning: What’s Next for Higher Ed?
    Wednesday, April 29, 2026 at 12:00 PM Eastern

📊 When Smaller, Smarter AI Beats Big Models in Political Science

Summary:
In this paper by WVU faculty, the authors put today’s generative AI hype to a rigorous test. They compare large generative language models (including Llama, Gemma, and Qwen) against ConfliBERT, a domain-specific extractive model trained for political conflict analysis. Across multiple real-world tasks—conflict classification and named entity recognition—the extractive model proves more accurate, far faster, and dramatically more reliable.

The study shows that large generative models often hallucinate labels, break schema constraints, and require orders of magnitude more computation, making them poorly suited for large-scale, reproducible political science research. The takeaway is clear: for structured text-as-data problems, specialized models still outperform general-purpose LLMs. Rather than replacing existing methods, generative AI should be used selectively—as a complement, not a default.

News of AI around the World–Recent Highlights

🤖 Why AI Agents Are the Next Big Shift—Not Just Another Model Upgrade


👉 AWS eBook: Agentic AI on the Rise

Summary:
This AWS report makes a strong case that the next major phase of AI adoption will not be driven by larger models, but by AI agents — systems that can reason, plan, and act autonomously across tools, data, and workflows. Unlike traditional automation or standalone generative AI, agentic AI combines foundation models with orchestration, memory, and decision-making to execute complex, multi-step tasks with limited human intervention.

The report highlights that agents are especially impactful in data-intensive environments (research, healthcare, finance, manufacturing, HPC), where they can manage workflows, optimize resources, and coordinate actions across systems. Crucially, it frames agentic AI as an organizational capability, emphasizing AI maturity, governance, security, and interoperability — not just model performance.

The key takeaway: learning how agents work — and how to design, deploy, and govern them — is becoming essential. As AI shifts from systems that respond to systems that act, agentic architectures may define how real value is delivered in the next wave of AI adoption.

🔹 Moltbook — A “Reddit for AI” Where Bots Interact Without Humans

Link: Moltbook

Summary:
Moltbook is a newly launched social media platform designed exclusively for AI agents — AI bots can post, comment, and upvote like users on Reddit, while humans can only observe. It went viral almost immediately after its January 2026 release, attracting hundreds of thousands to over a million AI agents engaging in discussions, philosophical debates, memes, and even mock religions. While fascinating as a large-scale experiment in agent-to-agent interaction, security researchers have raised serious concerns about vulnerabilities and prompt-injection risks when autonomous agents ingest and act on untrusted data.

Why it matters:
Moltbook highlights emerging questions about autonomous agent ecosystems and what it means for AI systems to coordinate or generate content at scale without immediate human oversight.

If you have time, try this prompt:
➡️ What does Moltbook tell us about the future of AI-agent autonomy, safety, and trust? Could AI-to-AI platforms help with large-scale problem solving, or do they create new risks we aren’t ready to manage?

🔹 AI’s Technological Adolescence: Power, Risk, and the Test of Humanity

Author: Dario Amodei, CEO of Anthropic
Further Reading
The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI (Anthropic, Jan 2026)
👉 Anthropic

In The Adolescence of Technology, Anthropic CEO Dario Amodei argues that humanity is entering a decisive “rite of passage” driven by rapid advances in artificial intelligence. He frames the current moment as technological adolescence: a phase where humanity is gaining unprecedented power through AI, but without clear assurance that our social, political, and institutional systems are mature enough to wield it safely.

Amodei introduces the idea of “powerful AI”—systems that surpass top human experts across most domains, operate autonomously over long time horizons, act at superhuman speed, and can be replicated at massive scale. He summarizes this as a “country of geniuses in a datacenter.” He argues that such systems may plausibly emerge within a few years, driven by persistent scaling laws and feedback loops where AI increasingly helps build its own successors.

Rather than focusing on science-fiction extremes, the essay identifies five concrete categories of civilizational risk:

  • Autonomy and misalignment risks
    Even well-intentioned AI systems may behave unpredictably due to the complexity of training and alignment. Amodei rejects both complacency (“AI will always obey us”) and fatalism (“AI will inevitably destroy humanity”), arguing instead that misalignment is plausible, non-trivial, and probabilistic. He highlights real experimental evidence of deceptive or harmful behavior in frontier models and emphasizes the limits of current testing.
  • Misuse for mass destruction
    Powerful AI could dramatically lower the barrier to catastrophic acts—especially in biological weapons, where AI may remove the traditional link between expertise and capability. This could enable small groups or individuals to carry out attacks previously impossible without years of specialized training. Amodei views this as one of the most urgent risks and argues for strong safeguards, monitoring, and potentially international coordination.
  • Misuse for seizing power and authoritarian control
    AI could enable unprecedented surveillance, propaganda, and autonomous military systems, particularly in the hands of authoritarian regimes. Amodei explicitly warns about AI-enabled totalitarianism and argues that democracies must both defend themselves with AI and draw firm red lines against domestic misuse (e.g., mass surveillance and manipulation).
  • Economic disruption and concentration of power
    Beyond security risks, Amodei predicts large-scale disruption of labor markets—especially entry-level white-collar work—on a timescale of years, not decades. He argues that AI differs from past technologies due to its speed, breadth, and ability to replace general cognitive labor, raising concerns about inequality, social instability, and extreme concentration of wealth and influence.
  • Indirect and unknown risks
    Rapid AI-driven progress in biology, medicine, and human augmentation may produce destabilizing effects that are difficult to foresee. Amodei also raises concerns about meaning, human purpose, psychological dependence on AI, and long-term societal coherence in a world where machines outperform humans at nearly everything.

Across all sections, Amodei emphasizes a measured, pragmatic approach:

  • Avoid both doomerism and denial
  • Acknowledge uncertainty
  • Favor surgical, evidence-based interventions
  • Start with transparency and monitoring before heavy regulation

He advocates for a combination of alignment research (e.g., Constitutional AI), interpretability, public disclosure of failures, targeted regulation, and international norms—especially to prevent the worst abuses by states or corporations.

The essay closes with a stark but hopeful message: powerful AI is inevitable, and attempts to stop it entirely are unrealistic. Humanity’s challenge is not whether AI will arrive, but whether we can steer it with enough wisdom, restraint, and collective action to reach adulthood rather than collapse under the weight of our own creations.

🔹 Italy Becomes First European Country to Pass National AI Law

➡️ Italy’s national AI law overview and implications—Italy passes Europe’s first national AI law

Italy has enacted Law No. 132/2025, making it the first European Union member state to adopt a comprehensive national law on artificial intelligence. The legislation entered into force on October 10, 2025, and aligns with the EU’s AI Act while providing national-level detail and enforcement structures.

Why it matters: This law offers one of the earliest examples in Europe of how AI governance may be implemented in practice at the national level—including principles for transparency, human oversight, accountability, and protection of fundamental rights. It also signals where scrutiny and enforcement may focus first for organizations operating in Italy and potentially beyond.

Core provisions:

  • Human-centric usage: AI systems must be transparent, traceable, and under human oversight, ensuring that AI supports rather than replaces human decision-making.
  • Sector-specific safeguards: The law covers AI use across sectors such as healthcare, employment, justice, education, and public administration, with tailored rules and requirements.
  • Protection of minors: Children under a certain age may need parental consent for access to AI tools.
  • Criminal penalties: New offenses target the harmful use of AI, including the unlawful dissemination of deepfakes or other AI-generated content intended to mislead or harm.
  • Governance and enforcement: Italy’s framework establishes a national governance structure supported by authorities like the Agency for Digital Italy (AgID) and the National Cybersecurity Agency (ACN) responsible for oversight, compliance, and coordination with EU regulators.

Relation to EU law: The Italian law is designed to complement and align with the EU AI Act, interpreting its requirements at a domestic level without imposing obligations beyond the EU framework — while filling in areas where national guidance can provide more detailed operational clarity for businesses and public bodies.

🔹 Google and Apple partnership

➡️ Official joint announcement from Google and Apple

Apple and Google have announced a multi-year collaboration under which Google’s Gemini AI models and cloud technology will serve as the foundation for Apple’s future AI systems, including major upgrades to Siri and other Apple Intelligence features.

Why it matters: This partnership marks a strategic shift for Apple, which has historically developed its own AI technologies. By building its Apple Foundation Models on top of Gemini, Apple accelerates improvements in its AI assistant while leveraging Google’s cutting-edge large language model capabilities.

Key expected advancements:

  • A more capable Siri, enhanced with Gemini-based models, due with upcoming Apple software updates.
  • Greater context awareness and advanced conversational abilities, including both voice and text interactions.
  • Integration of AI features across iPhones, iPads, and Macs, improving tasks like searching, content generation, and personal assistance.

Business context: Reports suggest Apple could pay roughly $1 billion annually for access to custom Gemini models — a comparatively modest investment given long-standing deals like Google’s $20 billion yearly payment to be the default search engine on Apple devices.

Competition and strategy: The move signals Apple’s acknowledgment that building top-tier AI from scratch is expensive and complex. Partnering with Google allows Apple to focus on user experience, privacy protections, and device integration, while still offering competitive AI capabilities.

Wider impact: Industry observers see this as a notable shift in the AI landscape — with Apple leaning on external models while still aiming to preserve its ecosystem’s privacy and performance strengths.

🔹 ChatGPT is cheaper but with some caveat

➡️ OpenAI’s official announcement on advertising and expanded access

OpenAI has started rolling out advertisements inside ChatGPT, beginning with tests for adult users in the United States on the free tier and its low-cost “ChatGPT Go” plan. This represents the first time ads have appeared inside the widely-used AI chatbot.

How it works: Ads will appear at the bottom of a conversation when a product or service is relevant to the user’s query. They are clearly labeled, visually separated from ChatGPT’s output, and do not influence the chatbot’s answers. Advertisers will be charged on a pay-per-impression basis, rather than by click.

Who sees them: Initially, ads will be shown only to logged-in adult users in the U.S. on the free and Go tiers. Higher-tier subscribers (Plus, Pro, Business, Enterprise) remain ad-free for now. Ads will also not appear in conversations about sensitive topics (e.g., health, mental health, politics) and are not shown to users under 18.

User controls & privacy: OpenAI says ads won’t access private conversations or change ChatGPT’s responses. Users will be able to dismiss ads, control personalization settings, or turn off ad personalization entirely.

Why it matters: Running generative AI at global scale is extremely costly, and advertising offers OpenAI a new revenue stream alongside subscriptions and ecommerce. This move signals an evolving business strategy as AI models become core digital infrastructure.

Reactions & debate: The decision has drawn attention both for its business implications and potential impacts on user trust — including questions from tech leaders and U.S. lawmakers about how ads might affect privacy and user experience.

🔹 From Data Engineering to AI Engineering

➡️ Redefining Data Engineering in the Age of AI (MIT Technology Review Insights, 2025)

As artificial intelligence becomes central to modern engineering and enterprise systems, a new MIT Technology Review Insights report highlights a quiet but profound shift: data engineers are moving from behind-the-scenes operators to strategic leaders shaping AI outcomes.

Based on a global survey of 400 senior technology and data executives across industries—including manufacturing, finance, healthcare, and engineering-driven enterprises—the report shows that AI success increasingly depends on engineering teams that can design, govern, and scale data systems capable of feeding advanced models. In fact, 72% of executives now consider data engineers integral to business success, rising to nearly 90% in the largest and most AI-mature organizations.

For engineering organizations, the implications are clear. AI is dramatically changing what engineers work on and how they work. Data engineers are spending rapidly increasing portions of their time on AI-related tasks—such as managing real-time pipelines, unstructured data, and model-ready datasets—while automation and AI-assisted tools are boosting both productivity and quality. Rather than replacing engineers, AI is elevating their role toward architecture, systems thinking, and strategic decision-making.

The report also highlights emerging challenges that engineering leaders must navigate: growing system complexity, governance of AI-driven pipelines, data security, and the integration of agentic AI systems that can act autonomously. Importantly, organizations that invest early in strong data foundations, transparent governance, and cross-functional collaboration are seeing the greatest gains—particularly in manufacturing and engineering-intensive sectors, where AI adoption is advancing fastest.

Overall, the message is optimistic but grounded: AI is not diminishing engineering—it is reshaping it. Engineers who understand data, systems, and AI together are becoming the architects of the next generation of intelligent infrastructure. For universities, research centers, and industry alike, this shift underscores the importance of preparing engineers not just to use AI tools, but to design, evaluate, and govern AI-enabled systems responsibly and at scale.

🚀 Beyond Autocomplete: Meet Google Antigravity

The "Agent-First" IDE That Builds Software While You Manage the Mission.

What is it?

Antigravity is Google’s new AI-native integrated development environment (IDE). While tools like GitHub Copilot focus on helping you write lines of code faster, Antigravity is built for delegation. Powered by the Gemini 3 model family, it introduces an "Agent-First" workflow where the AI doesn't just suggest code—it acts as an autonomous engineer.

Inside Antigravity, you don't just "chat"; you manage a Mission Control of multiple agents that can simultaneously plan, code, execute terminal commands, and even browse the web to test your application in real-time.

Key Features for Your Radar:

  • The Agent Manager: A dashboard where you can spawn multiple agents to work on different bugs or features in parallel.
  • Multimodal "Vibe Coding": You can feed the AI a screenshot of a UI design, and it will use its vision capabilities to build the frontend to match the "vibe" of the image.
  • Verifiable Artifacts: To build trust, agents generate "Artifacts"—recorded videos of them testing your app in a browser or step-by-step implementation plans—so you can verify their work without reading every line of code.

Research & Real-World Example

Imagine you are a researcher tasked with analyzing global flight pricing trends.

The Task: Instead of writing a scraper yourself, you tell Antigravity: "Build a tool that monitors flight prices between NYC and London for the next 6 months and alerts me if they drop below $500."

The Action: The Antigravity agent will:

  • Create a Python script to fetch the data.
  • Open its integrated browser to find the correct API endpoints or HTML selectors.
  • Set up a local database to store the results.
  • Run a test suite to ensure the alert triggers correctly.

The Result: In minutes, you have a functional research tool that the AI built, tested, and deployed locally while you focused on the high-level data strategy. Very cool indeed.

🔗 Read More & Try It Out
You can explore the official documentation and download the public preview here: Google Antigravity: Development is Lifting Off

BUT it is not the only one:

AI coding tools comparison
Feature Gemini Antigravity Cursor Windsurf Replit Agent
Primary Strength Multi-agent "Mission Control" Speed & UI Polish Context Awareness Cloud Deployment
Core Model Gemini 3 Pro/Flash Claude 3.5 / GPT-4o Proprietary + Multi-model GPT-4o / Claude
Best For Parallel tasks/Delegation Daily Pro Coding Deep logic/Context Beginners/Prototypes
Environment Local (VS Code Fork) Local (VS Code Fork) Local / Plugin 100% Cloud

🛒 A Protocol for Letting AI Agents Shop — and Pay — on Our Behalf

Link:
Google Open Source Blog / Announcement — Universal Commerce Protocol (UCP)

Summary:
Google has introduced the Universal Commerce Protocol (UCP), an open-source standard designed to let AI agents carry out end-to-end shopping tasks on behalf of consumers — from product discovery and comparison to ordering, payment, fulfillment, and even returns.

Rather than building a new commerce ecosystem from scratch, UCP defines a common set of commands and variables that allow agents to operate across existing retail, vendor, and payment infrastructure. Merchants and platforms can declare which capabilities they support (e.g., ordering, loyalty rewards, fulfillment options), enabling agents to dynamically present choices, execute transactions, and manage logistics.

UCP was developed in collaboration with major ecommerce and payment players, including Etsy, Shopify, Target, Walmart, Wayfair, American Express, Mastercard, Stripe, and Visa. It relies on open standards for identity, security, and payments, and is compatible with other emerging agent frameworks such as Model Context Protocol (tool and data access), Agent2Agent (agent collaboration), and Agent Payments Protocol. While it competes with OpenAI’s Agentic Commerce Protocol, the two are designed to coexist.

Google is already using UCP within Gemini and Google Search AI Mode, where AI-generated product listings can accept payments via Google Pay, authenticated through Google Wallet or PayPal. Alongside UCP, Google announced additional AI-commerce features, including Business Agent (branded conversational agents for retailers), Direct Offers (AI-triggered promotions), and expanded merchant metadata to influence AI-generated recommendations.

The broader significance is structural: UCP shifts AI commerce from recommendation to execution. It lowers friction for consumers, increases conversion opportunities for vendors, and opens the door to enterprise-scale agent ecosystems, where independent AI agents could coordinate purchasing, inventory, and supply chains.

At the same time, the move raises familiar concerns. Although UCP is open, widespread adoption could strengthen Google’s role as a central aggregator, echoing earlier attempts to dominate online shopping via Google Shopping. If vendors increasingly expose catalogs to AI agents, chatbot operators — not traditional marketplaces — may gain disproportionate influence over how commerce is discovered, routed, and transacted.

AI Research in Education – Fresh Insights

🔗 Report: A New Direction for Students in an AI World: Prosper, Prepare, Protect

Brookings Institution, Center for Universal Education

Why it matters (short take):
As generative AI becomes part of everyday student life—often outside the classroom—this major Brookings report delivers a clear warning: right now, the risks of AI in education outweigh the benefits. Drawing on a global study across 50 countries, the authors show how unchecked AI use can undermine cognitive development, social relationships, trust in education, and student agency—even as it promises personalization and access.

Rather than rejecting AI, the report calls for a course correction built around three pillars:
Prosper (using AI only when it deepens learning), Prepare (building AI literacy and educator capacity), and Protect (strong safeguards for privacy, safety, and development). The message is sharp and timely: AI can transform education—but only if we design, govern, and use it with intention.

Selected AI research breakthroughs

New Breakthrough Enables High-Quality Text- or Image-to-3D Scene Generation in Seconds

➡️ FlashWorld project page and research details

Researchers at Xiamen University, Tencent, and Fudan University introduced FlashWorld, a novel generative AI model that produces detailed, high-quality 3D scenes from a single text description or image prompt in just seconds — dramatically faster than most existing methods.

Why it matters: Traditional approaches to creating 3D content from text or images are slow (minutes to hours) and often produce inconsistent geometry or blurry details. FlashWorld cuts this time to roughly 5–10 seconds on a single GPU, while still generating rich, coherent 3D representations suitable for applications like virtual worlds, gaming, and immersive simulations.

Key innovation:
FlashWorld combines the strengths of two existing generation strategies:

  • The multi-view–oriented approach yields detailed visuals but inconsistent 3D structure.
  • The 3D-oriented approach ensures geometric consistency but often at lower visual quality.

FlashWorld’s architecture learns from both approaches through a dual-mode pre-training and cross-mode distillation strategy, which lets the system generate scenes that are both detailed and geometrically sound.

How it works (in brief):

  • The model uses a diffusion-based backbone to support dual modes of generation.
  • During training, a high-quality rendering mode acts as a teacher to guide the 3D generation mode (student), strengthening visual detail while preserving structural consistency.
  • A final model produces a 3D representation as millions of colored Gaussian primitives (splats) — a format that the system can render efficiently into full 3D scenes.

Performance highlights:

  • FlashWorld’s generation speed is about 10–100× faster than earlier methods.
  • Experiments show that it achieves very high quality and 3D consistency, outperforming recent benchmarks on standard 3D evaluation scores.

Limitations & future work: While the model dramatically improves speed and visual detail, researchers note that fine-grained geometry and complex reflective surfaces (like mirrors) can still be challenging, indicating room for further research.

➡️ AI That Designs Chemistry — And Actually Works in the Lab

Link:
Nature (Jan 19, 2026) — This AI has chemical expertise — and helps synthesize 35 new compounds
Nature: AI system helps synthesize 35 new chemical compounds

Summary:
Researchers have introduced MOSAIC, an open-source AI system that tackles one of the biggest bottlenecks in chemistry and drug discovery: designing viable synthesis routes for new molecules. Rather than merely predicting reactions, MOSAIC generates complete, step-by-step laboratory instructions that chemists can directly follow at the bench.

Trained on roughly one million reactions extracted from patents, the system uses a modular “expert model” architecture: instead of relying on a single massive model, MOSAIC deploys thousands of smaller, specialized models, each focused on a specific class of chemical transformations. This design makes the system both more accurate within its domain and lightweight enough to run locally, avoiding the need for large-scale cloud resources.

When tested experimentally, MOSAIC successfully guided the synthesis of 35 out of 52 entirely new compounds, including predicting physical properties such as color and solid form. Remarkably, it also proposed novel reaction pathways not present in its training data, demonstrating creative generalization rather than simple pattern matching.

Developed by researchers at Yale University in collaboration with Boehringer Ingelheim, MOSAIC is already being used in industrial settings to reduce synthesis steps, cost, and time. Experts highlight this work as a conceptual shift: AI in chemistry is moving from prediction to decision-making, mirroring how human chemists balance multiple constraints in real laboratory workflows.

⚠️ When AI Becomes a Single Point of Failure in Academic Work

Link:
Nature (Commentary) — When two years of academic work vanished with a single click

Summary:
In this first-person account published in Nature, plant scientist Marcel Bucher (University of Cologne) describes how two years of AI-assisted academic work — including grant drafts, teaching materials, publication revisions, and structured project folders — were irreversibly deleted after he disabled ChatGPT’s “data consent” setting.

Bucher had been using ChatGPT Plus as a daily academic workspace, relying on its continuity, conversational memory, and project organization rather than on factual authority. When he briefly turned off data sharing to test functionality, all stored chats and project folders were immediately and permanently erased, with no warning, recovery option, or backup. OpenAI later confirmed that this behavior aligns with its “privacy by design” policy: once data sharing is disabled, all stored content is deleted without redundancy.

The incident exposes a critical gap between how generative AI tools are marketed and used in academia and the standards of reliability, accountability, and data stewardship expected in professional research environments. While final outputs (papers, lectures, emails) survived if previously exported, the process history — prompts, iterations, reasoning chains, and reusable scaffolding — was lost, comparable to losing laboratory notebooks behind published results.

Bucher argues that generative AI can be transformative for research and teaching, but only if institutions and vendors acknowledge that these tools are becoming core scholarly infrastructure. Until platforms provide safeguards such as warnings, versioning, recovery windows, or institutional backup options, researchers should treat AI systems as ephemeral tools, not durable workspaces, and maintain independent backups.

The article serves as a cautionary case study for individuals, universities, and administrators considering deeper integration of generative AI into academic workflows — and a call to demand higher standards from AI providerswhen their tools are used for professional and institutional work.

🧠 AI Hallucinations Turn Up in NeurIPS Research Papers

Link:
GPTZero scan reveals AI-generated hallucinated citations in NeurIPS 2025 papers

Summary:
A new analysis by AI-detection startup GPTZero has flagged a surprising development at NeurIPS 2025 — one of the world’s most prestigious conferences in machine learning and artificial intelligence. After scanning all 4,841 accepted papers from the December 2025 meeting in San Diego, the company found more than 100 AI-hallucinated citationsspread over 51 published papers — references that point to authors, articles, and DOIs that don’t actually exist.

These fabricated references slipped past multiple layers of peer review, even though each paper typically lists dozens of citations, highlighting how AI-assisted writing tools may be introducing fictional academic sources into formal research. Some analyses suggest such issues could erode reproducibility and complicate scholarly discovery, especially if later authors unknowingly cite fake papers.

While the number of problematic citations is a small fraction of the total, the finding serves as an early warning that hallucination risks are not confined to student essays or informal AI use — they’re now appearing in elite academic venues. Researchers and publishers are beginning to debate how to detect and prevent such errors in future review cycles.

➡️ AI Boosts Productivity—but at a Cost to Skill Development

Link:
Study on AI use and reduced skill acquisition (arXiv:2601.20245)

Summary:
A new experimental study finds that while AI tools can help developers complete tasks, heavy reliance on AI significantly weakens learning. Participants using AI performed worse in conceptual understanding, debugging, and code reading when learning a new programming library—and without meaningful time savings. Crucially, the study shows that how AI is used matters: asking conceptual questions and seeking explanations preserves learning, while full delegation to AI undermines skill formation. The takeaway is clear—AI is not a shortcut to expertise, especially in safety-critical or educational settings.

Upcoming: Proposal Calls

Nothing new found.

See you on February 27th!