# bok-ai-lab-20250404-news-and-links

# **Harvard Professor Uses AI to Replicate Himself for Tutor Experiment**

Harvard physics professor Greg Kestin has developed a new kind of teaching assistant—one that never sleeps. In a pioneering experiment, Kestin used generative AI to create a customized tutor modeled on his own course materials, voice, and pedagogical style. The result: students who learned more, felt more engaged, and had 24/7 access to Ivy League-level support.

### **Why This Matters for the Future of Teaching**

Higher education has long relied on the model of the live lecture, with the professor at the center. But as students increasingly turn to digital tools outside the classroom, Kestin asked: What if AI could bring the professor directly to them—anytime, anywhere?

* **AI as a Teaching Extension, Not a Replacement:** “It’s not about replacing myself or my lessons,” Kestin said. “I think of these AI tutors as an extension of me and my colleagues.”
* **On-Demand Learning:** Students were able to ask questions in the moment they got stuck, instead of waiting to raise a hand or attend office hours. As one student put it: “It was really helpful in the sense that it never turned off. A tutor that never turned off.”
* **Personalized and Playful:** The AI allowed for more candid interaction. “I definitely felt like I could be a little more snarky than usual,” another student shared. “You were snarky to the AI?” the interviewer laughed. “Yeah.”

The AI wasn’t just a chatbot—it was specifically designed to understand the curriculum, anticipate sticking points, and guide students through complex concepts in physics, like mechanical energy or delta (Δ) as a symbol for change. “It’s even better than just being able to Google something,” one student explained. “It’s tailored specifically to this class.”

### **The Classroom Experiment: AI vs. Traditional Teaching**

Kestin designed a controlled experiment comparing two sections of students learning identical material: one in a traditional lecture format, and the other guided by the AI tutor.

#### **Key Results**

* **Higher Test Scores:** Students who used the AI tutor outperformed their peers in assessments.
* **Increased Engagement:** They reported greater flow during learning—what one student described as “clicking while you’re literally clicking your mouse pad.”
* **Positive Emotional Response:** The AI provided celebratory feedback, like digital thumbs-ups and happy faces, creating a more affirming learning experience.

### **Balancing Innovation and Caution**

Despite the positive results, Kestin remains thoughtful about potential drawbacks.

* **Misuse Risks:** “If it is misused, then students might not learn the fundamentals. If they’re circumventing their thinking, they might think they know it, but they don’t.There are a lot of things that we currently do that will move to AI tutors,” Kestin acknowledged. “But there are a lot of things that we do that are inherently human.”

As AI tools evolve, experiments like Kestin’s suggest that human-centered design in education can enhance—not erode—the student experience.

“I hope that while technology will always evolve,” Kestin concluded, “humanity is what brings learning to life.”

# **Harvard Panel Launches GAI Dialogues Series with Debate on Authorship and Ethics**

Harvard’s Faculty of Arts and Sciences launched its spring GAI Dialogues series last week with a faculty panel titled **“Original Thought in the AI Era: A Faculty Dialogue on Authorship and Ethics.”** The panel addressed questions surrounding the ethical and practical implications of generative AI (GAI) in research, teaching, and university administration.

### **Opening Poll Reveals Mixed Opinions**

The event began with a live audience poll posed by Dean of Arts and Humanities Sean Kelly, asking whether it is appropriate to use generative AI for tasks such as grading student papers, writing letters of recommendation, or screening job applicants. Responses were divided, with no clear consensus among attendees.

### **Diverging Views on AI’s Impact on Science**

The first speaker, **Matthew Kopec**, program director and lecturer for Embedded EthiCS, opened the discussion with a critical view: “Science is less fun because of all these tools.”

**Gary King**, Albert J. Weatherhead III University Professor and director of the Institute for Quantitative Social Science, and **Michael Brenner**, Michael F. Cronin Professor of Applied Mathematics and Applied Physics at SEAS, offered counterpoints. Brenner argued that researchers should use all available tools to further human progress: “We’re in the business of making discoveries to improve the world for humans. We should use every tool that we have at our disposal to do that.” Both he and King suggested that generative AI may free up time for scientists to focus on more complex problems.

King emphasized that integrating new technologies into education is consistent with Harvard’s long-standing approach: “You should be the kind of person that uses whatever the best tools are to progress the fastest and go the farthest.”

### **Ethical Considerations and Environmental Concerns**

While endorsing GAI use in the sciences, Brenner acknowledged the importance of oversight regarding who controls such tools. King commented more broadly: “Yes, this technology can be used for harm. Any technology can be used for that. The causal factor isn’t the technology, it’s the humans that decided to use it.”

Audience members raised additional concerns, including AI’s environmental impact. All panelists agreed that the technology requires substantial energy. King responded that AI could either contribute to environmental degradation or accelerate the development of cleaner energy solutions. “Before you single-handedly eliminate these incredibly visible tools,” he said, “let’s just figure out the cost and benefits.”

### **Context and Upcoming Events**

The panel marked the beginning of a series of events organized as part of a broader FAS initiative on generative AI, a priority for Edgerley Family Dean of the Faculty of Arts and Sciences **Hopi Hoekstra**. The initiative is being led by her senior adviser on artificial intelligence, **Chris Stubbs**, the Samuel C. Moncher Professor of Physics and of Astronomy.

Future events in the GAI Dialogues series include:

- **“Teaching With Integrity in the Age of AI”** (April 1), a faculty workshop on academic integrity and classroom use of AI
- **“Critical Reading in the Age of AI”** (April 3\)
- **“Critical Writing in the Age of AI”** (April 24\)
- **“Preparing Students for the Future: AI Literacy in the Liberal Arts”** (May 5\)

These sessions aim to support faculty in adapting to the evolving academic landscape shaped by generative AI technologies.

# **New Protocol Proposes Standard for AI Tool Integration**

Anthropic has released the **Model Context Protocol (MCP)**, a new specification designed to simplify and standardize how large language models (LLMs) interact with external services and APIs. Introduced in late 2024, MCP aims to make AI integrations more reliable, scalable, and secure.

### **The Integration Problem**

While LLMs excel at individual tasks, coordinating actions across platforms like Slack, Gmail, and Notion remains difficult. Current tool integrations face challenges such as:

- Overloaded API endpoints with inconsistent formats.
- Limited model context windows.
- Frequent API updates that break existing workflows.
- Tight coupling between tools and specific models.

These issues result in brittle systems that are hard to maintain or scale.

### **What MCP Offers**

MCP introduces a standardized protocol to manage AI–tool interactions more efficiently. Key features include:

#### **1\. Tool Discovery**

MCP servers expose available tools in a consistent JSON-RPC format. This allows models to dynamically find and use capabilities like “create calendar event” or “send Slack message,” without memorizing custom API schemas.

#### **2\. Context Efficiency**

By standardizing parameter formats and error handling, MCP reduces the size of system prompts and helps models focus on task logic rather than syntax.

#### **3\. Secure Abstraction Layer**

MCP separates models from direct API access, using OAuth 2.0 to define permission scopes (e.g., read-only). This limits the risk of unauthorized or unintended actions.

#### **4\. Modular Architecture**

MCP divides the system into:

- **Clients** (interfaces like Claude Desktop) that connect models to users and tools.
- **Servers** that adapt service APIs to the MCP format.
- **Service providers** (e.g., GitHub, internal databases), which don’t need to modify their existing APIs.

This separation enables broad interoperability without requiring buy-in from API owners.

### **Beyond Tools: Resources and Prompts**

MCP also defines:

- **Resources** for persistent memory (e.g., files, preferences) accessible across sessions.
- **Prompts** that provide reusable guidance or behavioral instructions to AI models.

These features are designed to support more advanced, autonomous agent workflows, though implementation is currently limited.

### **Adoption Challenges**

MCP’s promise is tempered by practical barriers:

- Few official servers exist; most are community-built.
- Client support remains limited and fragmented.
- The protocol is complex, with minimal conceptual documentation.
- Most models lack native MCP support, relying instead on client-mediated integration.

### **Potential Impact**

If widely adopted, MCP could:

- Enable secure enterprise AI applications.
- Support plug-and-play AI workflows.
- Reduce vendor lock-in across LLM providers.
- Spark an ecosystem of specialized tool servers and prompt libraries.

### **Getting Started**

Developers can explore MCP through:

- [Anthropic’s official documentation](https://docs.anthropic.com/claude/docs/mcp-introduction)
- The [MCP GitHub repository](https://github.com/anthropics/mcp)
- Community tools like **Smithery** or the **Vercel AI SDK 4.2**

As the protocol matures, MCP may provide a foundation for scalable, model-agnostic AI agents capable of real-world task coordination.

# **OpenAI and Anthropic Launch Competing Higher Ed AI Initiatives**

This week, AI labs OpenAI and Anthropic announced nearly simultaneous education-focused rollouts aimed at U.S. and international universities, reflecting growing competition over who will become the default AI platform for college students.

### **Anthropic’s First Major Push into Academia**

On Wednesday, Anthropic announced **Claude for Education**, its first large-scale academic offering. The company is partnering with **Northeastern University**, **London School of Economics (LSE)**, **Champlain College**, and infrastructure providers **Internet2** and **Instructure** (maker of Canvas). The initiative aims to expand “equitable access” to AI tools in higher education.

Central to the new offering is **“Learning mode,”** a feature designed to shift Claude’s behavior from answer-giving to question-asking. Using a Socratic approach, Claude encourages students to reflect with prompts such as “How would you approach this?” or “What evidence supports your conclusion?” The company positions this as a way to foster **critical thinking** rather than enabling shortcut solutions.

### **OpenAI Expands Its Lead in Campus Deployments**

One day later, OpenAI announced an expansion of **ChatGPT Edu**, first launched in May 2024\. The company revealed that **ChatGPT Plus**, normally $20/month, will be available **for free to all U.S. and Canadian college students through May**.

This move builds on recent OpenAI partnerships, including:

- A February rollout across the **California State University system**
- The **NextGenAI Consortium**, launched in March with **$50 million in funding** to support AI research across **15 colleges**

ChatGPT Plus access includes premium features like large file uploads, Deep Research, and enhanced voice capabilities. “Supporting their AI literacy means more than demonstrating how these tools work,” said **Leah Belsky**, VP of Education at OpenAI.

### **Implications for the Future of Academic AI Use**

While OpenAI has been active in higher education for nearly a year, Anthropic’s entry marks its first direct effort to support university users. The near-identical timing of both announcements signals the growing strategic importance of universities as **early adopters of AI tools**.

# **Google Adds Web Discovery Feature to NotebookLM**

Google has introduced a new capability to its AI-powered note-taking tool, **NotebookLM**, allowing the platform to **automatically find and summarize web sources** based on a topic description. The feature, called **Discover**, began rolling out on April 2 and is expected to reach all users within a week.

### **Discover: AI-Assisted Source Retrieval**

Previously, NotebookLM users needed to upload source material manually—such as PDFs, documents, or video links. With the new Discover function, users can instead enter a brief description of the topic they’re interested in. NotebookLM then searches “hundreds of potential web sources in seconds,” returning a list of up to **ten relevant articles or webpages**, each accompanied by a summary.

### **New Tools for Contextual Research**

Once selected, sources are saved within the user’s NotebookLM account for citation, note-taking, and question-answering. Google describes the feature as the first in a set of **Gemini-powered updates** aimed at making NotebookLM more research-friendly and autonomous.

A related feature, **“I’m Feeling Curious,”** allows the tool to generate random topics and supporting sources—designed both to showcase Discover’s capabilities and encourage exploration, similar to Wikipedia’s “random article” function.

### **Positioning NotebookLM as a Research Assistant**

With these changes, NotebookLM is evolving from a static summarizer into a **self-directed research assistant**, capable of finding and organizing content without user uploads. Google has not yet specified when additional Gemini-based features will be released.

# **OpenAI Announces Plans to Release Open-Weight Language Model**

OpenAI CEO **Sam Altman** has announced that the company is preparing to release a **“powerful new open-weight language model with reasoning”** in the coming months. The model will be the company’s first open-weight release since **GPT-2**, marking a notable shift in strategy for a company that has previously favored closed-source deployments.

### **Open Model Announcement and Developer Engagement**

Altman shared the news via a [public post](https://openai.com/open-model-feedback), stating that the team is working to make the model “very, very good” and is actively **seeking developer input** to maximize its usefulness. OpenAI plans to hold **developer events** in San Francisco, followed by sessions in Europe and the Asia-Pacific region, to gather feedback and preview early prototypes.

### **Shift Toward Open Weights**

The announcement follows earlier remarks made by Altman in a Reddit AMA, where he acknowledged that OpenAI had been **“on the wrong side of history”** regarding open-weight models. The new model will be released under OpenAI’s **preparedness framework**, with additional safeguards due to the increased risk of post-release modification.

While no specific technical details or release date have been disclosed, the company has launched a dedicated feedback portal to engage developers and enterprise users who may prefer to **run models locally** rather than via API.

### **Context: Rising Demand for Open Models**

The move comes amid growing interest in **open-weight alternatives** across the AI ecosystem, including the recent rise of models like DeepSeek, Mistral, and Meta’s LLaMA series. Open-weight models allow users to download, fine-tune, and deploy models independently, providing more flexibility for enterprise, research, and government use cases.

Altman emphasized that this decision reflects shifting priorities within the company: “We’ve been thinking about this for a long time but other priorities took precedence. Now it feels important to do.”

# **DeepMind Releases 145-Page Report on AGI Safety**

Google DeepMind has published a comprehensive **145-page paper** outlining its approach to safety in developing artificial general intelligence (AGI)—a hypothetical AI capable of performing any intellectual task a human can. The report, co-authored by DeepMind co-founder **Shane Legg**, projects that AGI could emerge by **2030** and warns of possible “**severe harm**,” including so-called **existential risks**.

### **Key Claims and Frameworks**

The paper introduces the concept of an **“Exceptional AGI”**—a system matching the **99th percentile of adult human performance** on a broad range of non-physical and metacognitive tasks. While it does not offer a definitive safety roadmap, it recommends efforts to:

- Block bad actors from accessing AGI
- Improve interpretability of model behavior
- Strengthen environmental controls around AI deployment

The authors caution that these are **early-stage strategies** with **open research challenges**, but argue that **proactive planning** is essential given the “transformative nature” of AGI.

### **Comparison with Other AI Labs**

DeepMind positions its strategy in contrast to other major AI labs:

- **Anthropic**, it claims, underemphasizes robust training, monitoring, and security.
- **OpenAI**, the report says, leans too heavily on **automated alignment research** and is overly optimistic about achieving **superintelligence**.

While the paper is skeptical about the near-term emergence of superintelligent AI, it considers **recursive AI improvement**—where AI systems conduct AI research—as plausible and potentially dangerous.

### **Expert Reactions Highlight Skepticism**

Several AI researchers remain unconvinced by the paper’s assumptions:

- **Heidy Khlaaf** (AI Now Institute) argued that AGI remains too **ill-defined** for scientific evaluation.
- **Matthew Guzdial** (University of Alberta) dismissed recursive self-improvement as speculative, saying there is **no evidence** it works.
- **Sandra Wachter** (Oxford University) pointed instead to a **more immediate concern**: AI models increasingly **training on their own generated content**, which may contain **hallucinations** or **inaccuracies**.

### **Ongoing Debate Over AGI and Safety**

Though the report represents one of the most detailed public safety frameworks from a major AI lab, it appears unlikely to resolve broader debates around **AGI feasibility**, **timelines**, or the **most pressing risks** associated with advanced AI systems.

# **Microsoft Reportedly Scales Back Global Data Center Expansion**

Microsoft is **pulling back on planned data center developments** in several global locations, including the **U.K., Australia, and U.S. states like North Dakota, Wisconsin, and Illinois**, according to a report by *Bloomberg*. The decision signals a possible reassessment of the pace and scope of its AI and cloud infrastructure expansion.

### **Strategic Shifts in AI Infrastructure**

A Microsoft spokesperson told *Bloomberg* that the company’s data center plans are made **“years in advance”** and that recent changes reflect the **“flexibility”** of its strategy. Though Microsoft had projected **over $80 billion in capital expenditures** for 2025—primarily for AI-focused data centers—the company now appears to be prioritizing **upgrades to existing facilities** over new construction.

Factors influencing the pullback may include:

- **Infrastructure constraints**, such as **limited power availability** or **shortages in building materials**
- **Potential recalibration** of expected AI demand growth

Microsoft has not explicitly clarified whether the adjustments stem from changing market forecasts or logistical hurdles.

### **Implications for Higher Education and AI Tool Development**

This move could have **ripple effects** across sectors that rely on high-performance cloud computing, including **higher education**. Universities integrating **AI teaching assistants, research agents, and course-specific models** depend on scalable backend infrastructure provided by tech giants like Microsoft and OpenAI (in which Microsoft is a major investor).

Potential impacts may include:

- **Limited compute availability** for universities piloting large-scale AI deployments, such as campus-wide chatbots or research simulations.
- **Increased pressure on design patterns** to favor **efficiency and on-device processing**, reducing reliance on cloud-based inference for day-to-day academic applications.
- **Delays in AI-powered learning platforms**, particularly those that require large-scale, always-on models for writing support, code generation, or tutoring.

As cloud infrastructure investment becomes more selective, institutions may need to **prioritize AI solutions that use smaller, fine-tuned models**, take advantage of **open-weight alternatives**, or support **hybrid deployment** between on-premise and cloud systems.

###

# **AI Integration Speeds Up Prototyping in Engineering Education at University of Genoa**

A new academic paper from the University of Genoa documents how generative AI tools like ChatGPT and GitHub Copilot accelerated project-based learning in an advanced engineering course, reducing time-to-prototype while improving documentation, design, and student outcomes.

Published in the *AIxEDU 2024 Workshop Proceedings*, the study details how AI tools were embedded into the curriculum of a **Software Architecture for Embedded Systems** course in the **Master in Mechatronics Engineering** program. The approach enabled students to build and document complex systems—including SLAM-capable robots—within a single month.

### **AI as a Co-Designer in Engineering Workflows**

The course unfolded in two phases:

1. **Foundational learning** in embedded systems architecture and programming.
2. **Integration of AI tools** to support design iteration, debugging, and documentation.

Key tools included:

- **ChatGPT and Claude** for prompt-based design exploration
- **GitHub Copilot** for code generation and debugging
- **LM Studio** for secure, local AI deployment
- **PlantUML** for AI-assisted generation of architecture diagrams

Students used these tools to generate UML diagrams, test embedded systems, and troubleshoot real-time telemetry pipelines. Notably, the projects emphasized **AI as a support system** rather than a replacement for human reasoning.

### **Prompt-Based Learning and Reflective Practice**

A distinctive feature of the course was the emphasis on **prompt-based learning**, where students iteratively refined their prompts to improve AI output. As part of the final deliverables, students submitted:

- A **functional prototype**
- A **detailed project report**
- A **log of AI prompts and interactions**

These logs were evaluated for clarity, refinement, and evidence of metacognitive learning—aligning with theories of **distributed cognition** and **constructivist pedagogy**.

### **SLAM Project Example: Hardware-Software Co-Design with AI**

In one standout project, a student group built a robot with **Simultaneous Localization and Mapping (SLAM)** functionality using Arduino and Raspberry Pi systems. AI tools were used to:

- Generate telemetry and object-detection code
- Suggest architectural alternatives
- Produce accurate documentation diagrams (Use Case, State Chart, Class)
- Troubleshoot sensor fusion challenges

Students retained control over design decisions and used AI primarily to speed up repetitive or technically routine elements.

### **Outcomes and Implications for Higher Education**

All five student teams completed high-functioning prototypes in four weeks—nearly **double the productivity** compared to previous years. Student feedback was overwhelmingly positive, and instructors reported improved documentation quality and problem-solving depth.

The course demonstrates a **scalable model for AI integration in STEM education**, with broader implications:

- **Faster time-to-project** in engineering and design disciplines
- A potential shift toward **AI-augmented pedagogy** in applied fields like robotics, physics, and environmental engineering
- Increased focus on **prompt literacy** as a 21st-century academic skill
- Future emphasis on **secure, local AI tools** for privacy-sensitive settings

### **Future Directions**

The authors plan to expand the course in 2025 to include more teams, stronger prompting curricula, and additional lab-based resources such as 3D printing. A focus on **open-source and locally hosted AI systems** (e.g., LM Studio) will help address privacy concerns, especially for industry-aligned projects.

As AI-enhanced design becomes more common in academic settings, the course offers a blueprint for how **AI can shorten learning cycles while preserving student agency**—transforming not just engineering education, but AI literacy across disciplines.

# **Federated Learning Shows Promise for Privacy-Preserving AI in Education**

A new study from the **University of Bergen’s SLATE research group** explores how **federated learning**—a decentralized machine learning approach—can be used to protect student data privacy without significantly sacrificing model performance. The paper, presented at the *Learning Analytics and Knowledge Conference (LAK 2025\)*, provides one of the most detailed experimental comparisons between federated and traditional machine learning in educational settings.

### **Balancing Predictive Power and Privacy**

Federated learning (FL) enables AI models to be trained across multiple devices or institutions without aggregating raw data. The study tested FL on two open-source educational datasets—one on student performance, the other on dropout prediction—and compared results to non-federated learning, including under adversarial conditions such as **label-flipping attacks** (a form of data poisoning).

Key findings:

- FL models delivered **comparable accuracy** to non-federated models, particularly on larger datasets.
- FL demonstrated **greater resilience to adversarial attacks**, maintaining higher recall and AUC scores under label poisoning.
- Performance in FL settings reached a **saturation point** after a few training rounds, suggesting opportunities to optimize computational costs.

### **Why This Matters for Higher Ed AI Design**

As institutions experiment with predictive analytics, tutoring systems, and adaptive learning platforms, the ethical implications of student data use are under increasing scrutiny. Traditional approaches such as anonymization or differential privacy provide some safeguards, but **federated learning offers a new privacy layer**: students' data never leaves their local context.

For developers and academic institutions:

- FL allows for **collaborative model development** across campuses or departments without sharing raw data.
- It supports **distributed deployment patterns**, reducing reliance on centralized, cloud-based systems.
- The architecture is well-suited for use cases like **dropout prediction, academic performance alerts**, or **learning behavior analysis** in regulated environments.

### **Threat Modeling in Educational AI**

The study also highlights **label-flipping attacks** as a relevant concern in education. These attacks—where malicious actors mislabel data to degrade model performance—pose risks in high-stakes environments like financial aid or admissions predictions. The authors show that while FL is not immune, it **dampens the impact** of such attacks compared to standard machine learning.

### **Study Limitations and Future Directions**

The study is exploratory, using only two datasets and focusing on classification tasks. It does not address regression models, privacy metrics, or more advanced attack types like **membership inference**. Future research is needed to:

- Evaluate FL across a broader set of educational use cases.
- Incorporate **privacy guarantee assessments** alongside performance metrics.
- Study **stronger adversarial strategies** and corresponding defense mechanisms.

### **Implications for AI Governance in Education**

As educational institutions adopt AI tools for personalization, performance prediction, and resource allocation, federated learning may offer a viable foundation for **ethical, privacy-conscious design patterns**. Its ability to balance accuracy and protection is especially relevant in jurisdictions with strict data protection laws or equity-driven institutional mandates.