AI in Education and Knowledge Work

AI in education and knowledge work is the use of language models and related systems to tutor students, assess and produce written work, and assist professionals in text-based tasks such as drafting, summarising and analysis. The central empirical question is whether such tools improve learning and durable skill, or only performance while the tool is available.

Education and knowledge work share a structure: output quality can improve immediately with an assistant, while the user's own skill develops, or fails to develop, only over time. A tool that raises task scores today may leave a learner worse off tomorrow. For that reason the evidence is organised here by what was measured:

Deployment numbers (how many students use which tool) are vendor-reported and fast-changing; they are not tabulated here.

Intelligent tutoring systems have a long history; Bloom (1984) framed the "2 sigma problem", reporting that students tutored one-to-one performed about two standard deviations better than students in conventional classes, and set the goal of finding group methods as effective [VERIFICATION NEEDED: the two-sigma figure is recalled from the paper's well-known framing; the full text was not read]. Generative language models changed the situation from 2022: general-purpose chat assistants became available to students without being designed for instruction. Early reviews catalogued opportunities and risks (Kasneci et al., 2023), and field experiments began to appear: a tutor-versus-classroom trial (Kestin et al., 2025), a school mathematics experiment (Bastani et al., 2025), and several workplace experiments (Noy & Zhang, 2023; Dell'Acqua et al., 2023).

Why the tool design matters. A general chat model optimised to be helpful will often produce the answer, which can substitute for the student's own effort. A tutor configuration constrains the model with instructions: ask guiding questions, give hints, withhold final answers, use teacher-supplied solutions. Evidence from a school field experiment suggests that this difference is large (Bastani et al., 2025).

Performance versus learning. Let $P_{\text{aid}}$ be performance with the assistant and $P_{\text{no-aid}}$ performance after it is removed. A tool can raise $P_{\text{aid}}$ while lowering $P_{\text{no-aid}}$ relative to a control that never had it. Studies therefore need a delayed or assisted-free assessment, and the effect size should be reported for both.

Detection and assessment. Detectors that classify text as machine-generated are statistical classifiers with non-zero false-positive rates. A published evaluation reported that GPT detectors are biased against non-native English writers (Liang et al., 2023), which makes detector output unsuitable as sole evidence in misconduct cases.

• Tutoring and feedback systems: constrained assistants delivering hints, worked explanations and formative feedback.
• General assistants used by students: unconstrained chat tools used for homework, writing and revision.
• Teacher-facing tools: lesson planning, rubric-based feedback, translation.
• Assessment integrity tools: AI-text detectors, process-based assessment (drafts, oral exams).
• Workplace knowledge assistants: drafting, summarising, analysis for professionals; see APP-BIZ-001 and APP-SW-001 for sector-specific cases.

#6.1 Validated results

• Tutor versus active-learning class (university). In a randomised controlled trial, college students learned significantly more, in less time, from a custom AI tutor designed around pedagogical best practices than from an in-class active-learning lesson, and reported being more engaged and motivated (Kestin et al., 2025). Setting: one course and one institution, a purpose-built tutor [VERIFICATION NEEDED: sample size and effect sizes were not read; only the abstract-level conclusion is used].
• School mathematics field experiment. Nearly a thousand high-school students were given a GPT-4-based assistant in practice sessions. With a standard chat interface, grades on practice problems improved by 48%, and with a version prompted to safeguard learning, by 127%. When access was later removed, students who had used the standard interface scored 17% worse than students who never had access, while the safeguarded tutor largely mitigated that harm (Bastani et al., 2025). Without guardrails the assistant improved performance and reduced learning.
• Professional writing. In a preregistered experiment with 453 college-educated professionals, ChatGPT reduced average task time by 40% and raised output quality by 18%, with a decrease in inequality between workers (Noy & Zhang, 2023). These are short incentivised tasks, not full jobs; two months after the experiment, exposed workers were 1.6 times as likely to report using the tool in their real job.
• Consulting tasks. In an experiment with 758 consultants, GPT-4 users did more tasks, faster, on tasks inside the model's capability, but were 19% less likely to be correct on a task outside it (Dell'Acqua et al., 2023; working paper).

#6.2 Research demonstrations

• Essay writing and brain activity. A preprint with 54 participants in the main sessions (18 in the final session) reported weaker EEG connectivity and lower self-reported ownership of essays in an LLM group than in search-engine and unaided groups (Kosmyna et al., 2025). It is small, not peer-reviewed at the time of drafting, and EEG connectivity is an indirect measure; it motivates further research and does not establish lasting cognitive harm.

#6.3 Deployed products and their problems

• AI-text detectors. In use at institutions, with documented bias against non-native writers (Liang et al., 2023).
• General assistants. Used widely by students; the school experiment above shows the learning risk when no guardrails are set (Bastani et al., 2025). Quantitative adoption statistics were not verified for this draft.

Advantages. Immediate, patient feedback; possible gains in learning time for well-designed tutors (Kestin et al., 2025); productivity and quality gains in short professional writing tasks, larger for lower performers (Noy & Zhang, 2023).

Limitations.

• Performance-learning gap. Assistance without guardrails can reduce later unaided performance (Bastani et al., 2025).
• Capability frontier. Gains on tasks inside the model's range can turn into errors outside it, and users may not recognise the boundary (Dell'Acqua et al., 2023).
• Hallucination. Language models produce fluent but unsupported statements (see CAP-LIMITS-001); in education this includes fabricated citations and incorrect worked solutions.
• Equity and integrity. Biased detectors can penalise non-native writers (Liang et al., 2023).
• Generalisability. The strongest positive result concerns a purpose-built tutor in one setting; the negative result concerns an unguarded general assistant in another. Neither establishes the effect of "AI" in general.

Unresolved questions. Long-run effects on skill and employment (especially for novices); optimal tutor design and teacher roles; how assessment should adapt; whether workplace productivity gains persist and who captures them (see ECON-LABOR-001).

As of 2026-10-15: the evidence supports a conditional conclusion. Purpose-built, pedagogically constrained tutors can raise measured learning in randomised settings (Preliminary to Well Supported, depending on replication, which this draft did not verify); unguarded general assistants can raise assisted performance while reducing later unaided performance (Preliminary; one large field experiment). Claims of broad cognitive harm or benefit from everyday use rest on small or non-peer-reviewed studies (Contested). Workplace experiments show average gains on short writing tasks with heterogeneous effects (Well Supported for the tasks studied).

Near-term (0-3 years): more randomised trials with delayed post-tests; shifts toward process-based and in-person assessment; institutional policies on permitted use. Medium-term (3-10 years): integration of tutors with curricula and teacher dashboards; evidence on long-run outcomes. Long-term effects on skill formation are uncertain and not rated here.

• Foundational Concepts:
  • TECH-LLM-001 Large Language Models: The model class behind assistants and tutors.
  • CAP-GENAI-001 Generative AI: General generative capabilities.
• Related Methods & Technologies:
  • CAP-LIMITS-001 Hallucination, Reliability and Calibration: Error behaviour relevant to instruction.
  • APP-BIZ-001 AI in Business and Finance: Sector field experiments.
  • APP-SW-001 AI in Software Engineering: Novice and expert effects in programming.
• Advanced Implications & Future Trajectories:
  • ECON-LABOR-001 AI, Productivity and Labor Markets: Labour-market implications.