Bar-Ilan University AI Agent Helps Students Catch Up, but Faculty Still Do the Teaching

Bar-Ilan University expanded an AI learning agent after a 71-course pilot reached about 2,400 students, despite limited evidence about its academic impact. The Bar-Ilan University AI agent searches approved course materials, creates summaries, and generates practice questions inside Moodle.

That controlled design separates the project from students pasting lecture notes into a public chatbot. Faculty members decide which recordings, transcripts, presentations, and articles enter the system. Students then question a course-specific knowledge base rather than the open web.

The distinction also creates the project’s central tension. Restricting the agent’s sources can reduce irrelevant answers, but it cannot guarantee accurate explanations or better learning. Bar-Ilan must now show that convenient retrieval supports understanding instead of replacing the difficult work that produces it.

The Bar-Ilan University AI Agent Moves Beyond a Small Pilot

Bar-Ilan has turned an experimental learning assistant into a broader test of course-grounded AI.

The university piloted the agent in February 2026, according to Microsoft’s reported rollout. The initial deployment covered 71 courses and approximately 2,400 students.

Bar-Ilan plans to add another 100 courses during the first semester of the 2026/27 academic year. That expansion is meaningful because it moves the system beyond a few enthusiastic instructors.

The university has more than 20,000 students in Ramat Gan, near Tel Aviv. The expanded deployment still represents only part of the institution, but it creates a larger operational test.

Students reach the agent through Moodle, the learning management system already used for course activities. The familiar location matters because students do not need to adopt a separate study platform.

The agent supports Hebrew, Arabic, and English. Its course knowledge can include recorded lectures, transcripts, slides, academic articles, assignments, and other faculty-selected resources.

A student can ask where a lecturer discussed a specific concept. The agent can retrieve the relevant material, condense it, and produce questions for self-testing.

Those functions address a common problem after an absence. Missing one lecture creates more than an empty page in a notebook. Later material often depends on definitions, examples, and arguments introduced during that session.

Yarden Shay, an undergraduate studying criminology and psychology, described that accumulating gap in Microsoft’s account. She had previously relied on classmates’ notes while also managing six to eight hours of new lectures each day.

Shay now uses the agent to review course material and test herself with multiple-choice and open-ended questions. She told Microsoft that her daily study time after lectures fell from five or six hours to about three.

That report offers a concrete use case, but it remains one student’s experience. It does not establish that every participant saved time or learned more effectively.

The project also has a longer history than the latest rollout suggests. Bar-Ilan announced an original 2024 pilot involving 15 courses and 15 lecturers across fields including law, psychology, physics, engineering, and history.

That earlier assistant grew from a tool developed for reserve soldiers who needed help recovering missed coursework. The design therefore began with continuity and access, not automated essay production.

Its expanded form reflects the same goal. The agent helps students navigate materials that their instructors have already approved, rather than producing an independent curriculum.

The system uses Azure AI Services and Microsoft Copilot Studio. KPMG and OpenApp worked on its development, while Bar-Ilan’s teaching staff supplied and reviewed the academic content.

This combination makes the deployment less autonomous than the word “agent” might imply. The software retrieves, summarizes, and generates study activities, but faculty members define its informational boundaries.

That narrow role is the reason the project deserves attention. Bar-Ilan is testing whether universities can make generative AI useful by limiting what it knows.

Grounded Answers Change the Campus AI Debate

The strongest feature is not unrestricted intelligence, but a deliberately restricted knowledge base.

General-purpose chatbots can answer questions about nearly any subject. That breadth is useful, yet it creates a poor fit for courses with specific terminology, assigned readings, or contested interpretations.

A broad model might explain a legal concept using another jurisdiction. It might emphasize a theory the lecturer never assigned or present an outdated definition.

Bar-Ilan’s agent instead uses retrieval grounding. This process supplies selected source material to a language model so its response follows that material more closely.

Microsoft’s grounding controls let builders connect agents to specified files and other knowledge sources. Builders can also decide whether an agent should use general model knowledge.

Microsoft says Bar-Ilan’s implementation is grounded exclusively in materials uploaded by teaching staff. The agent therefore operates more like a conversational interface to each course than an independent online tutor.

That approach improves relevance, but “grounded” does not mean infallible. A model can still summarize a passage badly, merge separate ideas, or express uncertainty with unwarranted confidence.

Source quality also becomes decisive. Sparse slides may work as prompts during a live lecture, yet provide too little context for an automated explanation.

Lecturer Eliyahu Biton encountered that limitation while preparing healthcare law and ethics courses. He expanded slide bullets, added detailed papers, supplied lecture transcripts, and repeatedly tested answers.

His experience reveals the hidden mechanism behind the polished student interface. The agent becomes useful only after instructors turn implicit classroom knowledge into explicit, machine-readable material.

This is a familiar challenge in personal knowledge management. Retrieval quality depends on how clearly information is captured, organized, and connected before a question arrives.

For students, the controlled corpus offers another advantage. A summary can remain aligned with the lecturer’s teaching sequence instead of pulling definitions from unrelated websites.

That consistency can be especially helpful when students study in a second language. It can also reduce the time spent switching among video recordings, slide decks, readings, and classmates’ notes.

The three-language interface increases accessibility, although the available reporting does not compare answer quality across Hebrew, Arabic, and English. Translation and multilingual retrieval remain important areas for testing.

Course grounding also changes how instructors can evaluate errors. When an answer looks wrong, they can inspect the approved source material and determine whether the problem began with the content or the model.

A public chatbot offers less control. Its answer may reflect training data, live web results, user uploads, system instructions, or several sources blended together.

Bar-Ilan’s approach gives faculty a smaller error surface. It does not remove the error surface.

This distinction should guide how students use the tool. The agent can help locate and rehearse course knowledge, but disputed answers still require a lecturer or authoritative reading.

The design also avoids one common misunderstanding about educational AI. More information is not always the objective.

Students already have recordings, slides, articles, and online databases. Their problem is often finding the right explanation at the moment of confusion.

A course-grounded agent changes the interface to that information. Whether it changes learning depends on how students respond after retrieval.

An AI Teaching Assistant Creates More Work Before It Saves Any

The AI teaching assistant shifts faculty effort from answering repeated questions to preparing and supervising reusable course knowledge.

The Bar-Ilan project is often described through student convenience, but Biton’s preparation shows the institutional cost. An effective agent requires more than uploading an old folder of slides.

Faculty must review recordings, clean transcripts, expand shorthand, resolve contradictions, and identify obsolete readings. They must also test questions that students are likely to ask.

That work resembles building a new course resource. The difference is that students can explore it through conversation instead of following a fixed document.

For Biton, the investment appears to have reduced routine and administrative questions. He also reported that students seemed less anxious before examinations.

Students can ask the agent to create new practice exams instead of relying only on archived questions. They receive immediate feedback explaining why an answer is correct or incorrect.

This use moves the system beyond passive summarization. Retrieval finds information, while generated quizzes ask students to recall and apply it.

Practice can support learning when students genuinely attempt an answer. It becomes less useful when they immediately request a solution and mistake recognition for mastery.

The AI teaching assistant therefore needs deliberate constraints. It should prompt effort, show evidence, and direct ambiguous questions back to human instructors.

The lecturer’s role changes under this model, but it does not disappear. Faculty become responsible for both subject teaching and the information environment surrounding the agent.

Biton described his evolving role as more focused on mentoring. He continues to hold 90-minute sessions four to six times annually for deeper discussion.

Those meetings preserve something a retrieval system cannot reproduce reliably. An instructor can examine a student’s reasoning, challenge assumptions, and adapt to confusion that was never anticipated.

That division of labor is more credible than a replacement narrative. The agent handles repeated access and practice tasks, while the lecturer handles judgment and intellectual development.

However, the arrangement may place unequal burdens on faculty. A lecturer with extensive digital materials can prepare an agent more easily than someone teaching through discussion, laboratories, or changing fieldwork.

Course design also affects suitability. A fact-heavy introductory module offers clearer grounding material than a seminar built around interpretation and disagreement.

Universities must therefore avoid treating agent readiness as a proxy for teaching quality. Some excellent instruction does not produce a large archive of reusable documents.

Maintenance is another cost. Laws change, research advances, and lecturers revise their arguments. An answer grounded in last year’s trusted file can still be wrong today.

Every course needs ownership rules for updates. Someone must decide when a recording expires, whether a transcript is accurate, and which version of a reading remains authoritative.

Faculty members also need a way to see where the agent struggles. Repeated unanswered questions can reveal missing content, confusing instructions, or weaknesses in the course structure.

That feedback loop offers a meaningful institutional benefit. The agent can expose gaps that remain invisible when individual students quietly search elsewhere.

Yet analytics must be handled carefully. A student’s questions can reveal academic difficulties, health concerns, political beliefs, or sensitive personal circumstances.

Institutions should specify who can inspect conversations, how long logs remain available, and whether interactions influence assessment. The available public reporting does not provide those details.

An AI teaching assistant becomes trustworthy through governance, not branding. The interface may look simple, but the operating model requires content ownership, review procedures, privacy controls, and human escalation.

Course-Grounded AI Pressures Open-Ended Chatbots

The main contest is between institution-controlled course agents and open-ended chatbots that students already use without faculty supervision.

Universities cannot realistically prevent every student from consulting a general chatbot. The practical question is whether they can offer a better academic alternative.

A course-specific agent competes on relevance and trust. It knows the assigned materials and can reflect the vocabulary used in a particular classroom.

A public chatbot competes on range and convenience. It can introduce outside examples, compare competing theories, and answer questions beyond the formal syllabus.

Neither route wins automatically. The open system risks unsupported claims, while the closed system risks reproducing incomplete or biased course materials.

Bar-Ilan’s strategy favors the closed route for routine learning support. Students can still consult lecturers and wider sources, but the agent begins from an approved academic boundary.

Other universities are testing broader arrangements. Arizona State University has integrated ChatGPT Edu into teaching, research, and operations through a campuswide AI program.

That model invites experimentation across many university functions. Bar-Ilan’s implementation is narrower because its most visible student use centers on specific course repositories inside Moodle.

Earlier campus chatbots followed another route. Georgia State University’s Pounce began by answering questions about enrollment, financial aid, and registration.

Georgia State later tested course messages and reported improved outcomes in a political science class. Its chatbot study found that first-generation participants receiving messages earned final grades about 11 points higher than peers.

Pounce mainly demonstrated the value of timely nudges and administrative support. Bar-Ilan’s AI learning agent tackles the content of lectures and readings more directly.

These examples show how the category has changed. Campus chatbots once answered predefined questions or sent reminders. Newer systems generate explanations, summaries, and assessments from unstructured academic material.

That transition increases both usefulness and risk. A registration deadline has one correct date, while a legal or historical question can support several interpretations.

Course grounding helps by narrowing the evidence. It cannot decide which interpretations deserve emphasis unless the supplied material and instructions make that choice clear.

The pressure on open-ended chatbots is therefore not about raw model intelligence. It concerns whether an institution can create a safer default with better context.

Students also need clarity about the boundary. If the agent answers only from approved sources, it should disclose when those sources do not contain an answer.

A confident response without adequate evidence would defeat the purpose of grounding. A useful system must sometimes say that the course materials are insufficient.

Citation behavior matters for the same reason. Students should be able to move from a generated summary to the lecture segment, slide, or reading behind it.

Without that traceability, the agent becomes another authority students are expected to trust. With it, the agent can serve as a map back to primary course materials.

This design supports a healthier study pattern. A student can ask for orientation, inspect the evidence, and then work through the original source.

The alternative is dependence on compressed answers. Fast summaries can conceal disagreement, remove context, and make difficult concepts appear settled.

Institutions should measure which pattern actually develops. Usage volume alone cannot distinguish productive inquiry from answer outsourcing.

Bar-Ilan’s expansion gives it an opportunity to evaluate those behaviors across different disciplines. The results will matter more than the number of conversations.

The Early Results Are Promising but Far From Conclusive

The public evidence shows adoption and encouraging anecdotes, not proof that the agent improves learning at scale.

Microsoft’s account offers several positive signals. Shay reported shorter study sessions and greater independence, while Biton observed fewer routine questions and calmer students.

Biton also said the average final score among approximately 50 students was three percentage points higher than in previous years. He acknowledged that the cohort itself may have been stronger.

That caveat is essential. A year-to-year comparison cannot isolate the agent’s effect from teaching changes, exam difficulty, student composition, or ordinary variation.

The deployment also lacks a publicly reported control group. Without comparable students who did not use the agent, the score difference remains suggestive rather than causal.

Self-reported time savings require similar caution. Reducing study time can indicate efficiency, but it can also indicate less engagement with full lectures and readings.

Researchers would need to compare retention, transfer, examination performance, and student behavior over time. Immediate satisfaction alone cannot answer whether deeper learning occurred.

The agent’s quiz feature raises another measurement question. Generating unlimited practice can improve preparation, but only if questions accurately reflect course objectives.

Poor questions may reward surface recall. Repeated exposure to similar generated items may also create confidence without flexible understanding.

Faculty review becomes especially important in fields involving safety, law, health, or contested evidence. A subtle mistake can mislead students even when most of an answer is correct.

Privacy is another unresolved area in the public account. Microsoft says the agent operates inside Bar-Ilan’s secure Azure environment, but that statement does not explain every data practice.

Readers still need details about retention, access, model improvement, auditing, and deletion. Institutions should also explain whether instructors can view individual questions.

UNESCO’s education guidance emphasizes data protection, human oversight, ethical validation, and pedagogical design. Those principles apply even when software runs within an institutional cloud environment.

Academic freedom introduces a separate concern. A centralized platform can standardize support, but each lecturer needs control over how their subject is represented.

Students also need a clear way to challenge an incorrect answer. Reporting a problem should be simple, and corrections should reach everyone affected.

Accessibility deserves direct evaluation rather than assumption. Multilingual access is valuable, but performance can vary across languages, document formats, and speech transcription quality.

Recorded lectures may contain specialized terms, accents, formulas, or names that automatic transcripts misread. A grounded agent can faithfully retrieve an inaccurate transcript.

Coverage across disciplines may reveal further limits. Text-centered courses fit the system more naturally than studio work, laboratory technique, musical performance, or clinical judgment.

Equity also has two sides. An always-available assistant can help students with work, caregiving duties, military service, illness, or language barriers.

However, students with stronger prompting skills may extract more value. Others may not know when to distrust a fluent answer or request supporting evidence.

Universities should teach those skills explicitly. Giving students an AI agent without AI literacy would transfer too much responsibility to the user.

There is also a risk of institutional overreach. Once a platform exists, administrators may see it as a way to increase class sizes or reduce human support.

The current Bar-Ilan account does not claim such a plan. Still, the project’s value depends on preserving the lecturer-agent relationship described by Biton and Shay.

The safest interpretation is therefore limited. Bar-Ilan has shown that a course-grounded agent can operate across dozens of courses and support plausible study tasks.

It has not yet shown that the system consistently improves grades, retention, critical thinking, or long-term knowledge. Those are the standards that should guide the next phase.

What the Next 100 Courses Need to Prove

The expansion should be judged by learning evidence, answer quality, and sustained faculty participation, in that order.

The first signal is comparative academic performance. Bar-Ilan should examine whether agent users improve on common assessments after accounting for prior achievement and course differences.

The strongest evaluation would compare similar groups while measuring more than final grades. Delayed assessments could show whether students retain knowledge after the examination period.

Results should also be separated by activity. Searching for a lecture segment, reading a summary, and completing generated quizzes are different behaviors with different educational effects.

If stronger outcomes follow active practice but not passive summaries, Bar-Ilan could redesign the agent around questioning and retrieval. That finding would strengthen the project’s learning case.

If usage grows without measurable improvement, the agent may still offer convenience. It should not be presented as evidence of better education.

The second signal is answer reliability. The university needs a repeatable evaluation set covering factual retrieval, synthesis, ambiguous questions, missing information, and multilingual responses.

Faculty members should score answers against course materials before deployment and throughout the semester. Error categories should distinguish wrong retrieval from flawed generation and defective source content.

The system should also record appropriate refusals. An agent that declines unsupported questions can be safer than one optimized to answer everything.

Expansion across 100 additional courses will provide a much harder test. A method that works for a carefully prepared law course may struggle with equations, diagrams, laboratory procedures, or interpretive debates.

Published error rates would help students and instructors calibrate trust. Even internal reporting can improve deployment decisions if results guide course approval and revision.

The third signal is faculty participation after the novelty fades. Preparing an AI learning agent requires sustained editorial labor, not a one-time technical connection.

Bar-Ilan should track how long instructors spend preparing materials, checking answers, responding to reports, and updating sources. It should also ask whether those costs decline in later terms.

High participation would suggest the workflow creates enough value to justify that effort. Faculty withdrawal would indicate that the system shifts too much maintenance onto individual lecturers.

Support teams can reduce that burden through transcript correction, content templates, evaluation tools, and clear governance. They should not remove instructors from decisions about academic meaning.

Students deserve the same transparency. Each course should explain what the agent knows, what it records, and which tasks remain unsuitable.

A study companion should direct students toward evidence and reflection. It should not become an invisible shortcut that replaces lectures, readings, or conversations.

Students already building their own study systems can also use a structured student knowledge hub to connect notes and sources. The same principle applies: organization should support thinking, not substitute for it.

The Bar-Ilan University AI agent is a credible attempt to make campus AI more accountable. Its approved sources, Moodle integration, and faculty oversight create a stronger foundation than an unrestricted chatbot.

The next stage must test outcomes rather than celebrate access. Do students remember more, question sources more carefully, and recover from missed classes without becoming dependent on summaries?

Those answers will determine whether course-grounded agents become durable academic infrastructure. Universities considering similar systems should watch the evidence from those next 100 courses before treating deployment itself as success.