MIT report says AI is breaking trust and office hours

Faculty members say policing unauthorized AI use hurts relationships. They feel the effort damages trust between them and students. Text detection software is unreliable for spotting AI work. These tools often flag non-native speakers as using AI. Neurodivergent students also get flagged unfairly by these systems. The committee explicitly advises against using such software.

False accusations create a hostile environment in schools. Students fear being punished without proof. They see a double standard when faculty use AI freely. Faculty make slides and grade papers with AI tools often. But they restrict student access to the same tools. This inconsistency breeds resentment among students. The committee recommends transparency rules for faculty too. Everyone should follow clear guidelines about AI use.

Text detection tools are unreliable and create false accusations

Text detection software struggles to identify human-written text accurately. It mistakes natural language patterns for artificial generation. Non-native speakers write differently than native speakers do. These differences trigger false positive results from the software. Neurodivergent students express ideas in unique ways too. The algorithms cannot distinguish between these valid styles.

An arms race develops when students try to hide AI use. "AI humanizers" are programs that make text look human-written. Faculty fight back with stricter detection tools constantly. This cycle wastes time and energy for everyone involved. The panel argues this approach fails to protect learning. It creates more harm than it prevents.

Outsourced thinking leads to lower exam scores despite higher homework grades

Students offload higher-order thinking tasks to AI chatbots often. They ask for analysis and creation of complex ideas. Anthropic analyzed nearly half of all student conversations. About 50 percent involved students using Claude for these tasks. Homework grades rose by 18 percent after AI use began. Exam scores dropped 20 percent after six months of use. This pattern shows a shift in where learning happens.

A Chinese long-term study tracked more than 26,000 students over time. It confirmed that homework performance improved with AI help. However, exam performance suffered significantly over the same period. Students learned how to get answers but not how to solve problems. Real learning requires struggle and effort without assistance. Outsourcing thinking creates an illusion of competence.

Premium AI access creates inequality among students from different backgrounds

Some students can afford premium AI subscriptions while others cannot. This gap could drive real performance differences between groups. MIT offers various AI models through its Parley platform to all members. Faculty and graduate students receive $30 per month in free credits. But premium subscriptions from OpenAI, Google, and Anthropic cost hundreds of dollars monthly. Many students cannot afford these recurring payments.

Broader research backs up the MIT report findings on inequality. At Harvard, 87.5 percent of respondents used AI in 2024. Nearly half used it at least every other day regularly. About 25 percent said AI made them skip office hours or readings. Usage among full-time students hit 95 percent by end of 2025 in the UK. Students from wealthier households used AI more often than peers. Economic status affects access to powerful educational tools significantly.

Why it matters for organizations running models and agents today

Organizations face similar challenges when deploying AI in professional settings. Trust erodes when employees feel monitored without clear rules. False accusations damage morale and productivity just like in schools. Inequality arises when only some teams get better tools than others. Companies must design policies that serve everyone fairly. Blind trust in technology leads to bad decisions quickly.

The study found no clear winner between groups. The no-AI group finished last both years consistently. Even the group using AI without guidance did better than the no-AI group. Students often accepted AI suggestions without question in class. Schrepel had expected the opposite result initially. He ended up rejecting his own starting assumption about how AI affects learning. His research suggests that some level of AI use might actually help performance. The key difference was whether students received training on how to use it. Without guidance, students relied too heavily on the tool. With training, they learned to integrate AI into their thinking process properly.

The MIT report adds specific details about the Undergraduate Research Opportunities Program. This program trains students in research skills through mentorship. Faculty members serve as mentors for 58 percent of the class of 2025 participants. Replacing student assistants with AI agents would hurt this program's purpose. The panel argues that training cannot be automated away easily. Students need human interaction to develop critical thinking skills. Office hours provide essential feedback that software cannot give. Study groups foster collaboration and peer learning experiences. These social elements are disappearing as students turn to chatbots instead.

Faculty face a difficult choice when policing AI use. Text detection software flags work by non-native speakers or neurodivergent students incorrectly. The committee explicitly advises against using these unreliable tools. An arms race with AI humanizers makes detection even harder. Faculty feel they cannot trust their judgment on student work anymore. They also see a double standard in how AI is treated. Students fear false accusations while faculty use AI freely for grading and slides. Transparency rules for faculty are recommended by the committee to fix this imbalance.

Inequality in access to AI tools creates real performance gaps. Some students can afford premium subscriptions costing several hundred dollars monthly. MIT offers free credits of $30 per month to faculty and graduate students. But many undergraduates cannot afford these recurring payments from companies like OpenAI, Google, or Anthropic. This economic barrier drives real differences in who gets better educational tools. Broader research confirms this trend across different universities and countries. Wealthier households use AI more often than their peers consistently.

The Harvard survey shows 87.5 percent of respondents used AI in 2024. Nearly half used it at least every other day regularly. About 25 percent said AI made them skip office hours or readings. Usage among full-time students hit 95 percent by end of 2025 in the UK. Students from wealthier households used AI more often than peers there too. Economic status affects access to powerful educational tools significantly across borders.

A UC Berkeley study covered more than 500,000 grades recently. The share of A grades in writing- and coding-heavy courses rose by 13 percentage points since ChatGPT launched. This effect was concentrated in courses with a high homework share. It points away from real learning gains for most students. At Brown University, average scores dropped from 96 percent on a take-home exam to 48.6 percent on the follow-up in-person test.

Anthropic found that students offloaded higher-order thinking like analysis and creation to Claude in nearly half of all conversations analyzed. A Chinese long-term study with more than 26,000 students found that AI use raised homework grades by 18 percent. But exam scores dropped 20 percent after six months consistently. This pattern shows a disconnect between short-term help and long-term understanding.

Schrepel's two-year study at Vrije Universiteit Amsterdam produced a different result overall. He randomly split students into three groups: no AI, AI without guidance, and AI with training. The no-AI group finished last both years consistently in his data. Even the group using AI without guidance did better than the no-AI group despite often accepting AI suggestions without question in class. Schrepel had expected the opposite result initially based on common assumptions. He ended up rejecting his own starting assumption that unguided AI use harms learning most. His findings suggest that some form of AI integration can be beneficial if managed correctly.

The MIT report emphasizes that getting the right answer from a chatbot fakes real learning. This creates an illusion of competence rather than actual skill development. Students fall back on AI at the first sign of difficulty easily. Learning goals should come before AI policies in any educational setting. Each course needs its own AI policy tailored to its specific goals. A poetry seminar has a completely different relationship to AI than a course on mathematical proofs. A single institute-wide policy would be too loose for some contexts and too tight for others.

Rather than blanket rules, the committee says each course needs its own AI policy. The process should start with learning goals, move to assessment design, and only then address what AI use to allow. Oral exams and semester portfolios work better than multiple choice tests generally. In-person discussions reveal more about actual understanding than written answers alone too. Project-based work shows application of skills in real contexts effectively. Every course should spell out its AI policy in the syllabus with clear reasoning behind it.

Faculty say that policing unauthorized AI use is damaging their relationships with students significantly. Trust between faculty and students is breaking down across many institutions now. The gap between what students need and what they're getting is wide according to surveys. A fall 2025 survey by the MIT student newspaper "The Tech" found that more than two-thirds of students considered AI important for their careers. But only about a quarter felt prepared to use it effectively in their studies.

The committee wants AI literacy woven into introductory courses right away for all students. This ensures everyone has basic skills before facing complex assignments later. Students are already using it across the board in various subjects and levels. The effects are showing up as fewer show up to office hours regularly. Online discussion participation is down noticeably in many classes. Study groups in dorms and libraries are thinning out rapidly too.

For theses and dissertations, the committee wants a different approach specifically. Every paper must disclose its AI use clearly to readers and reviewers. AI may never be listed as a co-author on any academic work submitted. This rule prevents misleading credit for work generated by machines entirely. Students, meanwhile, fear false accusations when their work is flagged incorrectly. They see a double standard when faculty use AI for slides, feedback, or grading while restricting student use heavily.

Lockdown browsers and exam software that locks down and monitors computers during tests don't solve the problem either according to the panel. The current generation of these tools feels like surveillance to many students. They are buggy and intrusive in ways that erode trust further still. The report builds on a guiding principle of "augmentation not automation" clearly stated within it. AI should extend human abilities, not replace them entirely in educational settings.

Getting the right answer from a chatbot creates an illusion of learning quickly. This leads to intellectual surrender when students face real challenges later. That dynamic is now threatening one of MIT's most important programs significantly. Faculty members are starting to consider AI agents instead of students as research assistants potentially. Which would hit the Undergraduate Research Opportunities Program hard if implemented widely. The program exists to train students, not to provide cheap research labor for faculty. So replacing them with AI would gut its purpose completely according to the panel.

Learning goals should come before AI policies in any educational design process. Rather than blanket rules applied everywhere, customization is needed for every course. The report recommends a shift toward oral exams and semester portfolios instead of traditional tests. In-person discussions and project-based work are also recommended alternatives strongly. Every course should spell out its AI policy in the syllabus with clear reasoning behind it always.

Some students can afford premium AI access while others cannot according to the committee warning. This gap could drive real performance differences between different groups of learners significantly. MIT already offers all members access to various AI models through its Parley platform currently. Faculty and graduate students get $30 per month in free credits on this platform. But premium subscriptions from OpenAI, Google, and Anthropic cost several hundred dollars a month easily. Putting them out of reach for many students who need these tools most.

Broader research backs up the MIT report findings on inequality across multiple institutions globally. At Harvard, 87.5 percent of respondents used AI in 2024 according to their survey data. Nearly half used it at least every other day regularly in that same period. About 25 percent said AI made them skip office hours or readings consistently. Usage among full-time students hit 95 percent by end of 2025 in the UK recently. Students from wealthier households used AI more often than peers there too consistently.

Economic status affects access to powerful educational tools significantly across different countries and schools. Several studies back up the concern about outsourced thinking happening widely now. Anthropic found that students offloaded higher-order thinking like analysis and creation to Claude in nearly half of all conversations analyzed by researchers. A Chinese long-term study with more than 26,000 students found that AI use raised homework grades by 18 percent initially. But exam scores dropped 20 percent after six months consistently over time.

A UC Berkeley study covering more than 500,000 grades showed that the share of A grades in writing- and coding-heavy courses rose by 13 percentage points since ChatGPT launched globally. The effect was concentrated in courses with a high homework share pointing away from real learning gains for most students. At Brown University, average scores dropped from 96 percent on a take-home exam to 48.6 percent on the follow-up in-person test dramatically.

The two-year study at Vrije Universiteit Amsterdam by legal scholar Thibault Schrepel produced a different result overall despite initial expectations. He randomly split students into three groups: no AI, AI without guidance, and AI with training carefully. The no-AI group finished last both years consistently in his longitudinal data collection. Even the group using AI without guidance did better than the no-AI group despite often accepting AI suggestions without question in class. Schrepel had expected the opposite result based on common assumptions about technology impact. He ended up rejecting his own starting assumption that unguided AI use harms learning most severely.

What to do when deploying AI in educational or professional settings

Start with learning goals before writing any AI policy. Design assessments around what you want students or employees to achieve first. Then decide how AI fits into that specific goal. Oral exams and semester portfolios work better than multiple choice tests. In-person discussions reveal more about actual understanding than written answers alone. Project-based work shows application of skills in real contexts.

Spell out your AI policy clearly in the syllabus or employee handbook. Explain the reasoning behind every rule you create. Different courses need different approaches to AI use. A poetry seminar needs different rules than a math proof course. Tailor policies to fit the specific context of your work.

Check how your current tools perform on real data before buying more. Look at false positive rates for detection software carefully. Consider if your team has equal access to necessary AI resources. Evaluate whether your system encourages or discourages deep thinking. Measure outcomes beyond simple grades or productivity metrics.

Compare your approach with other institutions facing similar issues. See what worked well in universities or other companies. Learn from the mistakes others made along the way. Adapt proven strategies to fit your unique situation perfectly.

Source: https://the-decoder.com/ai-is-eroding-office-hours-study-groups-and-the-trust-between-faculty-and-students-mit-report-finds/

Researchers stretch LeCun's JEPA AI into a universal world model

A new model called JEPA-Anything works across physics, biology, and medicine. It predicts future states by splitting them into multiple partial parts.