Study AI grows up: checked citations, grounded notes and one trial every student should read

What changed this week

Digital Science announced on 29 October 2026 that it has acquired Jenni AI. Jenni AI is an AI research and writing platform used by more than 8 million students and researchers in 189 countries. The deal will connect Jenni AI to Dimensions. Dimensions is Digital Science's database of 171 million research publications. The aim is that sources Jenni AI finds and cites are grounded in the published research record. [1][2][3][1][2][3]

The deal lands in a busy month for student AI. Google published three new study features for Gemini Notebook on 16 October 2026. A University of Maryland trial of an AI study assistant reached a wider audience the same month. A Jed Foundation survey of more than 5,500 school students landed on 27 October 2026. Each item points to the same question. Students already use AI heavily. The tools and the rules around them are now catching up. [1][6][11][13][1][6][11][13]

Why the Jenni AI deal matters for students

Jenni AI is built for academic work from the first idea to a finished draft. Users can upload papers or import reference libraries. They can explore sources, draft and refine writing and check citations against the original material. That workflow is the reason students adopted it. It feels as easy as a general chatbot. It is shaped around sources rather than open chat. [1][2][1][2]

Digital Science says Jenni AI will continue as a standalone product under its own name. The team, pricing and subscription plans stay the same. Jenni AI joins Overleaf, Papers AI and ReadCube in the company's portfolio. Overleaf supports collaborative writing in LaTeX. ReadCube provides reference management. Papers AI brings data, analysis and writing together. For a student the practical point is continuity. The tool they already use is not shutting down. Its sources are about to get a stronger backbone. [2][5][2][5]

What checked citations actually mean

Digital Science chief executive Stephen Leicht framed the deal around verifiable citations. He argued that a general-purpose tool cannot give a student a citation checked against the actual paper. Inside Digital Science every source Jenni AI suggests can be verified against Dimensions and its 171 million publications. That is the promise. It matters because citation errors are a real risk in AI-assisted work. A fluent paragraph with a fake reference is worse than no paragraph at all. [1][2][1][2]

Disclosure is a second problem the deal touches. A September 2026 Oxford University Press survey of more than 2,600 researchers found that 74 percent were unsure which uses of AI they should disclose. Almost half said they had not fully disclosed their AI use. Grounded tools do not solve disclosure on their own. They make one part easier. A student can point to a real paper behind a claim rather than a chatbot sentence with no source at all. [4][5][4][5]

None of this means the integration is finished. Digital Science says the Dimensions connection will arrive over the coming months. It starts with the publication database and full-text search. Until it ships, Jenni AI users should treat the deal as a direction rather than a finished feature. Check citations by hand in the meantime. That habit stays valuable after the integration lands. [1][2][1][2]

Gemini Notebook's three new study features

Google published three new study features for Gemini Notebook on 16 October 2026. The first is a 60-second Short Video Overview for a tricky concept. Google presents it as a quick explainer a student can watch on the go or share with a classmate. The second is an Interactive Learning Overview under Reports. It walks a student through notes, lectures and readings step by step. It weaves together summaries and studio outputs like slides, quizzes and flashcards. The third is a set of expanded quiz formats. [6][7][8][6][7][8]

The quiz formats now include short-answer, multi-select and fill-in-the-blank questions. After a quiz or a flashcard review a student can ask the notebook for feedback. The feedback points to weak spots and suggests where to focus next. The Interactive Learning Overview is grounded strictly in the student's own coursework. That grounding is the key difference from a general chatbot. Answers trace back to material the student supplied rather than to training data alone. [6][7][6][7]

The Review toggle: check before it renders

Google announced a second change on 19 October 2026. Gemini Notebook gained a Review toggle for Slide Decks and Video Overviews. With the toggle on, a student can open outline notes for slides or producer notes for video before anything renders. The student can add feedback, ask for a summary at the start or swap the order of a scene. Only then does the student select Proceed to produce the final version. Google said the co-creation capabilities are rolling out globally to all users over the following week. [9][10][9][10]

One discrepancy is worth noting. A check on 23 October 2026 found the help pages still listing the review step for AI Ultra subscribers only. The announcement described a global rollout to all users. Students should not assume the toggle is live in their account until they see it. The feature matters because it puts a human check between a plan and a finished output. That check is exactly where errors get caught. [9][10][9][10]

Grounded tools versus general chatbots

The distinction running through both stories is grounding. A general chatbot answers from patterns in its training data. It can sound confident about a paper that does not exist. A grounded tool answers from sources the student supplied or from a verified database. The answer still needs checking. The check is possible because the source is visible. [6][8][6][8]

Gemini Notebook was renamed from NotebookLM on 16 July 2026. The rename came with a wider set of study tools. A free tier still exists alongside paid tiers. Free use moved to compute-based limits after Google replaced daily quotas in September 2026. That detail matters for students on a budget. Small questions no longer burn a full day of allowance. [10][17][10][17]

The Maryland trial: access alone lowered grades

The caution comes from a large randomised trial at the University of Maryland. Researchers ran the trial across 30 instructors and 2,379 undergraduates in the autumn 2025 semester. Some sections received a course-integrated virtual study assistant built on GPT-4o. The assistant used retrieval from instructor-selected course materials. Comparable sections of the same courses did not receive it. The paper appeared as an Annenberg Institute working paper in October 2026. [11][12][11][12]

Students offered the tutor finished with lower final grades. In the exact-match sample the drop was 0.37 standard deviations. That is roughly four points on a 100-point scale. Participation in instructor-designed online activities fell further. The drop there was 0.90 standard deviations. Page views, active days and discussion responses all fell. Homework scores showed no measurable change. Students still submitted homework. They stopped doing much of the surrounding practice that builds a mark. [12][18][12][18]

Use was low. Only about 15 percent of students offered the assistant ever opened it. Most requests asked for answers or explanations rather than practice or feedback. Effects were larger for first-generation students. The university's own report notes that giving students access may have signalled broad acceptance of AI tools. Students may then have turned to other AI agents outside the course system. [11][12][18][11][12][18]

What the trial does not prove

This section is analysis, not a new finding. The Maryland result does not show that every AI tutor harms learning. It tested one assistant in one configuration. Most instructors left it in answer mode. The authors limit their conclusions to that assistant as configured. A tutor built around forced retrieval and feedback could behave differently. That design has not been tested at this scale in a trial like this one. [11][12][18][11][12][18]

The paper is a working paper. It has not completed peer review. Numbers can move as review proceeds. The result also measures access rather than use. Most students in tutor sections never opened the tool. The grade effect still appeared. That pattern suggests the signal of access mattered as much as the tool itself. Universities should read it as a warning about unguided rollout rather than a verdict on AI in study. [11][12][11][12]

The rules around students are tightening

Schools are also drawing firmer lines. New York City announced a one-year moratorium on student-facing generative AI on 2 September 2026. The moratorium covers children from pre-kindergarten through eighth grade in the 2026-27 school year. It affects nearly 600,000 students. Companion chatbots are banned across all grades. High school students receive AI literacy modules and a small number of supervised pilots instead of open access. [15][16][15][16]

The Jed Foundation survey shows why schools feel pressure. More than 5,500 middle and high school students across 21 United States schools answered questions between October 2025 and April 2026. Fifty-eight percent said they use chatbots for help with schoolwork. Fifty-seven percent turn to chatbots when sad, stressed, lonely or overwhelmed. Students who used AI for homework reported higher levels of loneliness and a history of suicidal thoughts. The charity is clear about limits. The findings show links rather than proof of cause. The survey covers 21 schools in the charity's programmes rather than a national sample. [13][14][13][14]

A study loop that keeps the practice in your hands

This section is practical guidance drawn from the evidence above. It is not a study finding. Start with your own notes. Upload lectures, readings and problem sets into a grounded notebook. Ask the notebook to quiz you before it explains anything. Answer from memory first. Then check the answer against the source the notebook cites. [6][11][6][11]

Use short video overviews only for concepts that resist reading. A 60-second video can unlock a diagram or a formula. It cannot replace working through problems. Use the feedback after a quiz to choose the next topic. Do not paste the quiz question into a general chatbot for the answer. That move recreates the Maryland pattern. The tool answers. The practice disappears. [6][11][12][6][11][12]

Keep the human parts of study intact. Attend office hours. Discuss problems with classmates. Write drafts from sources you have actually opened. Disclosure rules vary by course and institution. Read them before submitting AI-assisted work. When a rule is unclear, ask the instructor. A short question now beats a disclosure problem later. [4][5][4][5]

What to watch next

Three developments are worth watching. The Dimensions integration into Jenni AI arrives over the coming months. Students should watch whether it changes the free tier or the citation display. New York City's Technology in Schools Coalition will publish recommendations for future school years in April 2027. The Maryland team has shared guidance with instructors in courses that already offer the assistant. Watch whether guided use produces different results from the first trial. [1][11][15][19][1][11][15][19]

The wider question is whether grounded tools can show learning gains in a trial like Maryland's. No comparable trial evidence for grounded study tools has appeared yet. Claims about grades should be treated as marketing until a randomised study backs them. Students can run their own small test. Use a grounded tool for one topic and plain notes for another. Compare quiz scores after a week. The comparison is rough. It still beats trusting a sales page. [11][12][11][12]

What we still do not know

Several questions stay open. The Dimensions integration has no firm launch date. Pricing for Jenni AI after the deal is unchanged today. Future pricing is unannounced. The Review toggle rollout may still be incomplete in some accounts. Google has not confirmed completion. Each gap changes what a student actually meets on screen. [1][2][9][1][2][9]

The biggest unknown is learning itself. Grounded tools reduce one failure mode. They show sources. They do not force practice. The Maryland trial suggests practice is where learning happens. A tool that makes answers easier without making practice stick could still hurt grades at scale. That hypothesis needs its own trial. Until then the safest rule is boring. Practise first. Check second. Trust the source rather than the sentence. [11][12][11][12]