Teacher, student and AI: who does what in the coaching loop

The best evidence says test-score gains come from frequent, structured instruction and practice, and that the human role that most reliably helps is a proactive adult who starts the contact, follows up and is trusted. In the strongest tutoring RCT with adolescents, the gain came through instruction, not through a “mentoring effect” (Guryan et al. 2021), while mentoring and coaching RCTs move persistence and decisions more than grades (Bettinger and Baker 2011; Carrell and Sacerdote 2013; Oreopoulos and Petronijevic 2019). Hybrid tools that hand a human tutor or teacher AI suggestions or live analytics improve learning modestly, with the largest gains for weaker tutors and weaker students (Wang et al. 2024; Holstein et al. 2018). Generic one-way nudges do not move outcomes at scale (Bird et al. 2019), and goals work when they are about tasks the student controls, not about scores (Clark et al. 2017). So the spec’s split between Claude and teachers is broadly right, but three parts need to change: commitments should be weekly task goals rather than score targets, the 10-minute one-on-ones should not be treated as a learning engine, and “teacher time saved” should be measured net of the time the loop adds.

Scope note. Almost every study below is from the US, Canada or England. Only Guryan et al. (grades 9-10), Faria et al. (grades 9-10) and Carrell and Sacerdote (grade 12) study adolescents near the JNV age. None studies Indian students, residential schools or JEE/NEET preparation. Every transfer to Avanti below is an inference, not a finding.

Evidence table

Study	Design and population	n	Effect (unit)	What it means for the coach
Guryan et al. 2021 (Saga tutoring)	Two RCTs, grade 9-10 students in Chicago public schools; 2:1 tutoring one class period every day	2,633 and 2,710 students	+0.16 SD and +0.37 SD on math tests; no detectable effect on feeling supported by adults, grit, locus of control	Daily, structured instruction moves scores. The effect did not run through the relationship.
Nickow, Oreopoulos and Quan 2020	Meta-analysis of PreK-12 tutoring RCTs	96 studies	Pooled 0.37 SD; during-school about twice after-school; “little evidence” for once-weekly sessions	A 10-minute 1:1 every 3-4 weeks is not tutoring. Dosage must come from frequent practice sessions.
Kraft, Schueler and Falken 2024	Meta-analysis of tutoring RCTs, US, by programme scale	265 RCTs in the paper text (the Oct 2024 cover abstract says 282)	0.42 SD full sample; 0.21 SD for programmes with 400-999 students; 0.16 SD for 1,000+	Expect much smaller effects at scale than in pilots. Supports the spec’s 0.15-0.20 SD as an upper realistic target.
Wang et al. 2024 (Tutor CoPilot)	RCT of an LLM tool that suggests moves to human tutors in live chat; grades 3-6, US South, mostly Hispanic and low-income	783 tutors (tutors randomised), 1,013 students	+4 pp topic mastery (62% to 66%); +9 pp for lower-rated tutors; no significant end-of-year test effect	AI that advises the human, who keeps the choice, works best for less experienced adults. Proximal gains need not show on the big test.
Thomas et al. 2024 (CMU hybrid tutoring)	Quasi-experiment, three low-income US middle schools; remote human tutors plus math software	125, 385 and 75 students	More time on software and more skills mastered; round-robin tutoring may have slowed students at or above grade level; 1:4 beat 1:8 (+0.36 workspaces per hour)	Direct human attention by need, not by rotation. Not randomised.
Holstein, McLaren and Aleven 2018 (Lumilo)	Classroom RCT, 3 arms, US middle school math with an intelligent tutor; teachers wear glasses showing live analytics	286 students, 18 classes, 8 teachers	Analytics vs business-as-usual r = 0.21 on learning gain; monitoring alone r = 0.13; analytics narrowed the gap between low and high prior-knowledge students	Teacher dashboards that say who is stuck help. Students knowing the teacher can see their work has its own effect.
Holstein and Aleven 2021	Design account of the Lumilo pilots and trial	14 teachers, 304 students in pilots	Teachers spent about 24 seconds per visit; often prompted reflection or encouraged rather than taught content	Short human moves in response to an AI flag are a realistic unit of teacher work. Descriptive only.
Roy et al. 2024 (EEF ChatGPT lesson prep)	School-randomised RCT, Year 7-8 science teachers in England	259 teachers, 68 schools	Prep time 56.2 vs 81.5 min/week (25.3 min saved, 31%); no quality difference on blind review; use declined over the trial	AI can save about 25 minutes a week on preparation without lowering quality. No student outcomes measured.
Bettinger and Baker 2011 (InsideTrack)	RCT, US college students, mostly non-traditional; professional coaches by phone, email and text	13,555 students	+5.2 pp enrolled at 6 months; +3.4 pp at 24 months	Proactive coaching on goals, time management and study skills raises persistence.
Oreopoulos and Petronijevic 2016/2018	RCT, Canadian university; online exercise vs one-way texts vs personal coaching	4,000+ students; coaching arm only 24	Coaching +0.35 SD GPA; texts and online exercise no effect	Contact that is proactive, two-way and remembers past conversations is the difference. Coaching estimate is from a very small arm.
Oreopoulos and Petronijevic 2019	Five years of RCTs, three Canadian campuses; online, text and face-to-face coaching	~25,000 students	No effect on grades or persistence (rules out above 0.07 SD); study time +2 hours/week; better well-being	Coaching can change habits and well-being without changing grades. Students study 5-8 hours/week less than they plan.
Carrell and Sacerdote 2013	RCT, New Hampshire grade 12 students at the margin of applying to college; weekly in-person mentors	Statewide, multiple schools	Women +15 pp college-going (ITT), ~30 pp TOT; information-only and cash-only arms no effect	Human mentoring works where it substitutes for missing adult time; information alone does not.
Clark et al. 2017	Two RCTs, US college course; self-set performance goals vs self-set task goals	3,971 students	Task goals: practice exams completed +~0.1 SD, course performance +~0.1 SD; performance goals: small, not significant	Commit to tasks the student controls (sets done), not to scores.
Bird et al. 2019 (nudging at scale)	Two RCTs, national and state FAFSA campaigns in the US	800,000+ students	No effect on aid receipt or enrolment; no arm (framing, timing, 1:1 advising) worked	Generic, one-way messages from a distant sender fail at scale.
Page and Gehlbach 2017	RCT, AI text chatbot for students admitted to Georgia State; unanswered questions routed to human counsellors	7,489 students	+3.3 pp on-time enrolment	AI handles routine questions; humans answer what it cannot, and their answers grow the knowledge base.
Faria et al. 2017 (EWIMS)	Cluster RCT, early warning system, US high schools, grades 9-10	73 schools, 37,671 students	Chronic absence 10% vs 14%; course failure 21% vs 26%; no effect on credits or on school “data culture”; implementation low	Flags help even when used imperfectly, but do not change how staff use data by themselves.
Fang et al. 2025	Four-week RCT of chatbot modes, adult US online panel	981 adults	No effect of assigned condition; heavier voluntary daily use associated with more loneliness and emotional dependence (correlational)	Do not make Claude the student’s confidant. Adults, not adolescents.

What the evidence says

Learning gains come from dosage and instruction, not from the relationship alone

Guryan et al. (2021) ran two RCTs of daily 2:1 tutoring with Chicago 9th and 10th graders. Math scores rose 0.16 SD in the first and 0.37 SD in the second. The authors tested whether this was a “mentoring effect” and found no detectable change in students feeling close to adults, grit, conscientiousness or locus of control. They conclude the gain came from the instruction. Nickow et al. (2020) pooled 96 RCTs to 0.37 SD, found during-school tutoring roughly twice as effective as after-school, and found “little evidence” that once-weekly sessions produce large effects. They also note too few studies from grade 6 upward to give reliable estimates there.

Kraft et al. (2024) add a scale correction. Across 265 RCTs the pooled effect is 0.42 SD, but it falls to 0.21 SD for programmes serving 400-999 students and 0.16 SD for 1,000 or more. Their high-school math estimate is large (0.55 SD) but rests on few studies.

AI that advises the human works, modestly

Tutor CoPilot gave human tutors LLM-generated suggestions during live chat sessions. Tutors chose whether to use them. Students of treated tutors were 4 pp more likely to pass the session’s exit ticket (62% to 66%), and 9 pp more for lower-rated tutors (Wang et al. 2024). The authors did not detect an end-of-year test effect and say the two-month window and low exposure may explain it. Their limitations section says the tool may help experienced educators less.

Lumilo showed teachers live analytics from a math tutor through smart glasses. Learning gains rose against business-as-usual (r = 0.21) and against a monitoring-only version (r = 0.11). The analytics shifted teacher time toward low prior-knowledge students and flattened the gap between weaker and stronger students (Holstein et al. 2018). The monitoring-only arm also beat business-as-usual (r = 0.13), and the authors suggest students’ awareness of being watched reduced gaming of the software. In the design pilots teachers spent about 24 seconds per visit and often prompted reflection rather than teaching content (Holstein and Aleven 2021).

Thomas et al. (2024) combined math software with remote human tutors in three low-income middle schools. Hybrid tutoring increased time on the software and skills mastered. A round-robin rule that gave every student equal tutor time may have slowed students at or above grade level, and moving from 1:8 to 1:4 raised lessons completed per hour. The design is quasi-experimental.

AI saves teacher preparation time, by a known amount

In the EEF trial, Year 7-8 science teachers given ChatGPT and a short guide spent 56.2 minutes a week on lesson preparation against 81.5 for controls, a 25.3-minute saving, with no quality difference on blind expert review (Roy et al. 2024). Teachers used it for one or two activities, mostly writing questions and quizzes, and used it less as the trial went on. The trial did not measure student outcomes.

Coaching moves persistence and decisions more reliably than grades

Proactive professional coaching raised US college enrolment by 5.2 pp at six months and 3.4 pp at 24 months (Bettinger and Baker 2011). In-person weekly mentoring raised college-going among New Hampshire girls by 15 pp, while information-only and cash-only arms did nothing; the authors read this as the mentor substituting for missing parent or teacher time (Carrell and Sacerdote 2013).

The grade evidence is mixed. A 24-student coaching arm at Toronto raised GPA 0.35 SD while texts and an online exercise did nothing (Oreopoulos and Petronijevic 2016/2018). The authors attribute the difference to coaches who started conversations, remembered earlier ones and built trust. But the same team’s five-year programme with about 25,000 students found no grade or persistence effect from online, text or face-to-face coaching, while study time rose about two hours a week and well-being improved (Oreopoulos and Petronijevic 2019). Students studied 5-8 hours a week less than they planned, and coaching did not close that gap.

Goals: task goals beat score goals

Clark et al. (2017) randomised goal-setting in a college course. Self-set goals for a task (number of practice exams) raised practice completed and course performance by about 0.1 SD. Self-set performance goals (a target score) had small, non-significant effects. Only about a quarter of students hit a performance goal, against about half for task goals. The correlation between ambitious performance goals and grades (0.203) was much stronger than the causal effect, which the authors say makes correlational evidence on goals misleading. Task goals helped men and did little for women in their sample.

Nudges: generic and one-way fail; personal and backed by humans can work

Two FAFSA campaigns reaching over 800,000 students had no effect on aid or enrolment, and no variation in framing, timing or one-on-one advising changed that (Bird et al. 2019). The authors suggest the messages were generic, one-way and from a sender students had little connection to. By contrast, a Georgia State chatbot that answered students’ own questions, and routed what it could not answer to human counsellors, raised on-time enrolment by 3.3 pp (Page and Gehlbach 2017).

Early-warning flags help a little, even when used badly

In a 73-school RCT, an early warning system cut chronic absence from 14% to 10% and course failure from 26% to 21% in year one, even though implementation was low in nearly all schools. It did not change credits earned, and did not change how often staff reviewed data or what interventions they offered (Faria et al. 2017). The report says it is not clear which staff actions caused the gains.

What should stay human

Three findings bear on the split. First, the adult’s academic value in Guryan et al. (2021) came from instruction, which Claude can partly supply through practice. Second, the adult’s value in Carrell and Sacerdote (2013) and Oreopoulos and Petronijevic (2016/2018) came from proactive contact and trust, which Bird et al. (2019) suggest a distant, generic sender cannot reproduce. Third, in a four-week RCT with adults, heavier voluntary chatbot use was associated with more loneliness and emotional dependence, though usage was not randomised (Fang et al. 2025). Together these support keeping commitment conversations, pastoral care and the post-bad-mock conversation with teachers, and keeping Claude to practice, diagnosis and preparation.

Where the evidence is weak or contested

• No Indian or JEE evidence in this strand. No RCT tests an LLM coach plus teacher debrief with Indian adolescents. Residential JNV students differ from US commuter college students on almost every dimension that drives coaching effects, such as parental contact and time use.
• Small and fragile coaching results. The 0.35 SD coaching effect (Oreopoulos and Petronijevic 2016/2018) comes from 24 students. The same authors’ larger programme found no grade effect (2019).
• Proximal vs distal outcomes. Tutor CoPilot moved exit tickets, not the end-of-year test (Wang et al. 2024). Lumilo measured a short pre/post test. An AI coach may move practice accuracy without moving AIAT or JEE scores.
• Hybrid tutoring evidence is mostly quasi-experimental (Thomas et al. 2024).
• Monitoring vs privacy. Lumilo’s monitoring arm suggests being watched reduces gaming (Holstein et al. 2018), while the spec worries students will feel watched. No study here measures the cost side for adolescents.
• Teacher time. The 25-minute saving is for lesson preparation in England (Roy et al. 2024). It says nothing about the time a weekly debrief and 5-6 one-on-ones will cost.
• Gender. Task goals helped men, not women (Clark et al. 2017); mentoring helped women, not men (Carrell and Sacerdote 2013). Neither result should be assumed to carry over.
• Claims in the input documents not verified here. Pritam’s review cites a “35% GPA improvement” from mentoring, a systematic review of 82 studies finding hybrid workflows beat autonomous tutors, advisors “saving over 100 hours” and “5 to 10%” persistence gains. This strand did not locate or open the primary sources for these, so they should not be repeated as findings.

Design rules for the Avanti coach

1. Make the weekly commitment a task goal the student sets, with the teacher. “Finish 3 practice sets on rotational motion” rather than “score 120 in the next AIAT”. Keep score targets as the long-range anchor in the teacher’s goal conversation, not the thing checked each week (Clark et al. 2017).
2. Treat device sessions as the dosage engine and spread them across at least 3 days a week. One 2-3 hour block a week is the pattern the tutoring evidence finds weakest (Nickow et al. 2020; Guryan et al. 2021).
3. Do not count the 10-minute one-on-ones as academic intervention. Use them for commitment, follow-up and pastoral checks. Their academic value is in setting the next week’s tasks, not teaching (Nickow et al. 2020; Oreopoulos and Petronijevic 2019).
4. Route teacher attention by need, not by rotation, for academic help. The flag list should rank students who are stuck. Keep the rotation floor, but make it a short pastoral check, not a full slot (Thomas et al. 2024; Holstein et al. 2018).
5. Write the 1:1 brief as options, not a script. Give the teacher a one-line diagnosis, two or three probing questions to choose from, and the student’s own words from the session log. The teacher decides what to use (Wang et al. 2024). Expect the largest benefit for newer teachers.
6. Every flag carries a reason and a suggested action, and the system records whether the teacher acted. Report flag-to-action rate in the weekly review. A flag nobody acts on is the EWIMS failure mode (Faria et al. 2017).
7. Tell students plainly that their teacher reads the session log, and what it contains. Visible teacher review is part of what makes the loop work (Holstein et al. 2018), and it answers the spec’s transparency worry honestly rather than hiding it.
8. Between-session contact is specific, two-way and references the teacher’s plan. Generic reminders and motivational lines will not move outcomes (Bird et al. 2019; Oreopoulos and Petronijevic 2016/2018). Messages should name the student’s own task and ask a question the student can answer.
9. When Claude cannot answer, route to the teacher and feed the answer back. Keep a queue of unresolved student questions in the teacher view, and add the teacher’s answers to the coach’s content (Page and Gehlbach 2017).
10. Do not let Claude become the confidant. Short, task-focused sessions; any personal or distress content routes to a named human; do not optimise for longer chat time (Fang et al. 2025; Carrell and Sacerdote 2013).
11. Measure teacher time net. Count the 1-2 hours a week the loop adds against time saved on preparation and diagnosis. The best RCT benchmark for saving is about 25 minutes a week (Roy et al. 2024).
12. Judge the pilot on distal outcomes, read proximal ones as early signals only. Plan completion and practice accuracy can rise without AIAT or JEE moving (Wang et al. 2024). At scale, 0.16-0.21 SD is what strong human tutoring achieves (Kraft et al. 2024); an AI coach at 2-3 hours a week should be expected to do no better.

Where the current spec conflicts with the evidence

• “Teachers own setting goals and score targets.” The goal evidence favours student-set task goals; performance targets had small, non-significant effects (Clark et al. 2017). Keep teacher ownership of the conversation, change what gets committed to. The handoff concept’s monthly level targets (L2 to L3) are also performance goals; they need weekly task goals underneath them.
• Core problem 3, “Better teacher one-on-ones”, as a lever on learning. At about 10 minutes every 3-4 weeks, the 1:1 is far below the dosage at which tutoring shows effects (Nickow et al. 2020). It is a commitment and pastoral tool. The spec should say so, so nobody expects score gains from it.
• Device time of 2-3 hours a week, shape unspecified. If delivered as one weekly block, it sits in the weakest dosage pattern (Nickow et al. 2020). The handoff concept’s daily slot is closer to what the evidence supports.
• “Teacher time saved” as a success metric. The loop adds 1-2 hours of teacher connector time plus 5-6 one-on-ones a week. The only RCT benchmark found saves about 25 minutes a week on preparation (Roy et al. 2024). Net time may rise. Measure it rather than assume it.
• Rotation floor for everyone, plus everyone after each mock. Equal-attention rules may slow students who do not need help (Thomas et al. 2024, quasi-experimental). Keep the floor short and pastoral; give academic time by need.
• Open question “what happens between sessions: printed plans, teacher reminders, or nothing?” Reminders alone are the generic-nudge pattern that fails (Bird et al. 2019; Oreopoulos and Petronijevic 2019). What has worked is a human who starts the contact and follows up on the specific task (Oreopoulos and Petronijevic 2016/2018).
• Risk 3, students “may feel watched.” The spec treats being watched only as a cost. Lumilo suggests it also reduces gaming (Holstein et al. 2018). The design should be transparent about it, not minimise it.
• 8c, “0.15-0.20 SD vs matched control is a realistic pilot target.” That is the range for large human tutoring programmes (Kraft et al. 2024). For a 2-3 hour a week coach it is ambitious, and a matched (not randomised) comparison of one centre cannot separate the coach from centre effects.

References

1. Guryan, J., Ludwig, J., Bhatt, M. P., Cook, P. J., Davis, J. M. V., Dodge, K., Farkas, G., Fryer, R. G., Mayer, S., Pollack, H., and Steinberg, L. (2021). Not Too Late: Improving Academic Outcomes Among Adolescents. NBER Working Paper 28531. https://www.nber.org/papers/w28531. Opened 2026-10-26 (copied from gs://avantifellows-research-sources/talent-identification/standards__guryan_tutoring.pdf). Local file: guryan2023_saga_tutoring.pdf. A revised version was later published in a journal; this review read the 2021 working paper only.
2. Nickow, A., Oreopoulos, P., and Quan, V. (2020). The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence. NBER Working Paper 27476. https://www.nber.org/papers/w27476. Opened 2026-10-26 (copied from gs://avantifellows-research-sources/talent-identification/csf_claim_audit__nickow2020.pdf). Local file: nickow2020_tutoring_meta.pdf.
3. Kraft, M. A., Schueler, B. E., and Falken, G. (2024). What Impacts Should We Expect from Tutoring at Scale? Exploring Meta-Analytic Generalizability. EdWorkingPaper 24-1031, Annenberg Institute at Brown University. https://doi.org/10.26300/zygj-m525 (PDF: https://edworkingpapers.com/sites/default/files/ai24-1031.pdf). Opened 2026-10-26. Local file: kraft2024_tutoring_at_scale.pdf.
4. Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., and Demszky, D. (2024, v3 2026). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. arXiv:2410.03017. https://arxiv.org/abs/2410.03017. Opened 2026-10-26. Local file: wang2024_tutor_copilot.pdf.
5. Thomas, D. R., Lin, J., Gatz, E., Gurung, A., Gupta, S., Norberg, K., Fancsali, S. E., Aleven, V., Branstetter, L., Brunskill, E., and Koedinger, K. R. (2024). Improving Student Learning with Hybrid Human-AI Tutoring: A Three-Study Quasi-Experimental Investigation. LAK ’24. https://doi.org/10.1145/3636555.3636896 (arXiv:2312.11274). Opened 2026-10-26. Local file: thomas2024_hybrid_tutoring.pdf.
6. Holstein, K., McLaren, B. M., and Aleven, V. (2018). Student Learning Benefits of a Mixed-Reality Teacher Awareness Tool in AI-Enhanced Classrooms. AIED 2018, LNCS 10947, 154-168. https://doi.org/10.1007/978-3-319-93843-1_12 (author copy: https://www.cs.cmu.edu/~bmclaren/pubs/HolsteinMcLarenAleven-StudentLearningBenefitsMixedRealityTeacherAwarnessTool-AIED2018.pdf). Opened 2026-10-26. Local file: holstein2018_lumilo.pdf.
7. Holstein, K., and Aleven, V. (2021). Designing for human-AI complementarity in K-12 education. arXiv:2104.01266. https://arxiv.org/abs/2104.01266. Opened 2026-10-26. Local file: holstein2021_complementarity.pdf.
8. Roy, P., Poet, H., Staunton, R., Aston, K., and Thomas, D. (2024). ChatGPT in Lesson Preparation: A Teacher Choices Trial. Evaluation Report. Education Endowment Foundation / NFER. https://files.eric.ed.gov/fulltext/ED680921.pdf. Opened 2026-10-26. Local file: roy2024_eef_chatgpt_lesson_prep.pdf.
9. Bettinger, E., and Baker, R. (2011). The Effects of Student Coaching in College: An Evaluation of a Randomized Experiment in Student Mentoring. NBER Working Paper 16881. https://www.nber.org/papers/w16881. Opened 2026-10-26. Local file: bettinger2014_insidetrack_coaching.pdf.
10. Oreopoulos, P., and Petronijevic, U. (2016). Student Coaching: How Far Can Technology Go? NBER Working Paper 22630. https://www.nber.org/papers/w22630. Opened 2026-10-26. Local file: oreopoulos2018_coaching_technology.pdf.
11. Oreopoulos, P., and Petronijevic, U. (2019). The Remarkable Unresponsiveness of College Students to Nudging and What We Can Learn from It. NBER Working Paper 26059. https://www.nber.org/papers/w26059. Opened 2026-10-26. Local file: oreopoulos2019_unresponsive_nudging.pdf.
12. Carrell, S. E., and Sacerdote, B. (2013). Why Do College Going Interventions Work? NBER Working Paper 19031. https://www.nber.org/papers/w19031. Opened 2026-10-26. Local file: carrell2017_college_going_mentoring.pdf.
13. Clark, D., Gill, D., Prowse, V., and Rush, M. (2017). Using Goals to Motivate College Students: Theory and Evidence from Field Experiments. NBER Working Paper 23638. https://www.nber.org/papers/w23638. Opened 2026-10-26. Local file: clark2020_goals_college.pdf.
14. Bird, K. A., Castleman, B. L., Denning, J. T., Goodman, J., Lamberton, C., and Rosinger, K. O. (2019). Nudging at Scale: Experimental Evidence from FAFSA Completion Campaigns. NBER Working Paper 26158. https://www.nber.org/papers/w26158. Opened 2026-10-26. Local file: bird2021_nudging_at_scale.pdf.
15. Page, L. C., and Gehlbach, H. (2017). How an Artificially Intelligent Virtual Assistant Helps Students Navigate the Road to College. AERA Open, 3(4), 1-12. https://doi.org/10.1177/2332858417749220 (copy: https://eduq.info/xmlui/bitstream/11515/35719/2/page-gehlbach-how-artificial-intelligent-virtual-assistant-helps-students-navigate-road-to-college-2017.pdf). Opened 2026-10-26. Local file: page2017_chatbot_summer_melt.pdf.
16. Faria, A.-M., Sorensen, N., Heppen, J., Bowdon, J., Taylor, S., Eisner, R., and Foster, S. (2017). Getting students on track for graduation: Impacts of the Early Warning Intervention and Monitoring System after one year (REL 2017-272). IES, REL Midwest. https://ies.ed.gov/sites/default/files/migrated/rel/regions/midwest/pdf/REL_2017272.pdf. Opened 2026-10-26. Local file: faria2017_ewims.pdf.
17. Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., and Agarwal, S. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv:2503.17473 (v2). https://arxiv.org/abs/2503.17473. Opened 2026-10-26. Local file: fang2025_chatbot_psychosocial.pdf.

All local files are in scratchpad/sources/hybrid_human_ai/ with a .txt extraction beside each PDF. Bettinger and Baker’s published version (EEPA 2014) and Oreopoulos and Petronijevic’s published coaching paper (Journal of Human Resources) were not opened; the working papers were.

Leads, not opened

• Gehlbach, H., et al. (2016). Creating birds of similar feathers: leveraging similarity to improve teacher-student relationships and academic achievement. Journal of Educational Psychology. doi:10.1037/edu0000042. Tried Harvard DASH links listed by OpenAlex; downloads returned empty files. Relevant as RCT evidence that relationship quality can be raised and affects grades.
• Bergman, P. (2021). Parent-Child Information Frictions and Human Capital Investment: Evidence from a Field Experiment. Journal of Political Economy. doi:10.1086/711410. Tried the CESifo WP 5391 link; it returned an HTML page, not the PDF.
• Bergman, P., and Chan, E. W. (2021). Leveraging Parents through Low-Cost Technology. Journal of Human Resources. doi:10.3368/jhr.56.1.1118-9837r1. No open-access copy found.
• Demszky, D., et al. (2023/2024). Automated feedback to instructors on uptake of student ideas (RCT). Not searched for a copy; relevant to AI-generated feedback for teachers.
• The 82-study systematic review of hybrid human-LLM workflows and the “35% GPA improvement” mentoring study cited in Pritam’s Oct 2025 review. Primary sources not identified.
• Avery et al. (in progress at the time of Bird et al. 2019), a texting campaign comparing national and school-counsellor senders. Not located.