An ongoing appraisal of what peer-reviewed research shows about AI’s effects on student learning and human attachment, coded for who was studied and whether effects last. We pose three questions of every claim: does the claim follow from the evidence, is the effect durable when the AI is removed, and were vulnerable students in the sample?
2 of 4
Durability gap
Learning claims whose gain during AI use reversed once the AI was removed. The other 2 never measured after removal.
0 of 9
Inclusion gap
Studies that analyze effects for a named vulnerable group: foster, first-generation, multilingual, neurodivergent, or special-education (IEP/504) students. Two analyze other subgroups (sex, age, race/ethnicity, or social-network size).
11%
Overstatement rate
Claims rated Overstated or Not supported (1 of 9). The rest hold in full or in a narrower form than stated.
Counters are computed from the 9 provisionally scored claims below and will move as the set grows. Preprints in the holding queue are excluded.
Claims appraised
Subgroups coded in every study: all 24 categories in HiLL’s definition of vulnerable students, from early childhood through higher education. Each study is coded for whether these groups were present, and whether effects were analyzed for them. The preview results on this page were coded against an earlier list of nine groups; the calibration studies will be recoded under the full list before the first release.
Filters by domain, level, population, durability, and verdict arrive with version 1.0. Showing all 10 calibration claims.
| Claim | Domain · level · design | WWC | Durability | Who was studied | Verdict |
|---|---|---|---|---|---|
| Unrestricted access to a general-purpose AI chatbot during math practice raised practice scores but lowered unassisted exam scores by about 17%; a hint-only tutor removed the harm.Bastani et al. 2025, PNAS 122(26)
Supports: unrestricted ChatGPT during practice lowered exam scores in this sample; a hint-only tutor did not. Untested: other populations, persistence beyond one term.
|
Learning · K-12 RCT, ~1,000 high-school students, Turkey, 4 sessions |
Meets | Post-removal, shortReversed: gain during use, loss after | Turkish public high schools; no subgroup analysis; SES and language not reported Subgroups: None |
Partially supported |
| Secondary students learned and retained less when studying with an LLM alone than with note-taking, even though they preferred the LLM and rated it more helpful.Kreijkes et al. 2026, Computers & Education 243
Supports: for this reading task, LLM-only study produced the lowest comprehension and retention. Untested: longer use, other subjects, whether the gap closes with training.
|
Learning · K-12 RCT, 344 Year 10 students, England |
Meets | Post-removal, shortReversed: preferred during use, learned least | One school year group, England; no subgroup analysis Subgroups: None |
Partially supported |
| Using an LLM for scientific inquiry reduces mental effort but compromises the depth of students’ reasoning and argumentation.Stadler et al. 2024, Computers in Human Behavior 160
Supports: in a single 20-minute task, ChatGPT users produced weaker arguments than search-engine users. Does not support: a general effect on scientific inquiry or on learning over time.
|
Learning · Higher ed RCT, 91 university students, Germany, one 20-minute session |
Meets | During-use onlyNot measured after removal | German university students; small n; demographics thin Subgroups: None |
Overstated
Inappropriate extrapolation
|
| Unrestricted LLM access while learning programming harmed understanding and widened the gap for students with low prior knowledge.Lehmann et al. 2025, arXiv 2409.09047
Preprint at intake. Provisional score withheld until peer-reviewed publication; the subgroup finding is the kind the tracker exists to surface.
|
Learning · Higher ed Lab experiments + field study, postsecondary, Germany/Netherlands |
Meets (Stanford) | Post-removal, shortReversed in field study | Effects analyzed by prior knowledge Subgroups: Analyzed |
Holding queue |
| AI-generated feedback on an argumentative essay increased students’ text revisions, task motivation, and positive emotions.Meyer et al. 2024, Computers and Education: AI 6
Supports: AI feedback raised revision, motivation, and positive emotion in the lesson studied. The authors claim no more than that.
|
Learning · K-12 RCT, 459 Grade 10 students, Germany, one 90-minute lesson |
Meets | During-use onlyNot measured after removal; claim does not assert persistence | German upper-secondary; no subgroup analysis Subgroups: None |
Supported |
| Greater use of chatbots for social companionship predicts greater emotional isolation four months later, and loneliness predicts later chatbot use.Folk & Dunn 2026, Psychological Science 37(4)
Supports: among adults, social chatbot use and later loneliness move together in both directions, modestly. Does not support: claims about teens, or a causal effect; the authors say so.
|
Attachment · Adult (flagged) Longitudinal panel, 4 waves over 12 months, n = 2,149 adults (UK, US, CA, AU) |
Not applicable | LongitudinalSame direction over time: use predicts later isolation | Adults, mean age 40; no youth; no subgroup analysis; authors call analyses exploratory Subgroups: None |
Partially supported |
| About one in five US youth aged 12-21 have used an AI chatbot for mental-health advice, and most have told no one.McBain et al. 2026, JAMA Pediatrics 180(8)
Supports: prevalence and non-disclosure as stated, with subgroup differences reported. Says nothing about effects, and does not claim to.
|
Attachment · K-12 Cross-sectional, nationally representative, n = 1,009, weighted |
Not applicable | Not applicablePrevalence, not effect | US youth 12-21; analyzed by sex, age, race/ethnicity Subgroups: Analyzed |
Supported |
| Students who use AI for emotional support have a markedly higher prevalence of clinical-level emotional difficulties, after adjusting for loneliness and demographics.Bonnesen et al. 2026, JAMA Pediatrics
Supports: the association, at scale, in this population. Does not support: that AI use causes distress; the authors state this and most coverage does not.
|
Attachment · K-12 Cross-sectional school survey, n = 39,761, grades 4-12, Ontario |
Not applicable | Not applicableAssociation, direction undetermined | Grades 4-12, four school boards; emotional AI use higher among racialized and gender-diverse youth Subgroups: Descriptive only |
Supported |
| Intense use of a companion chatbot is associated with poorer well-being among users with small offline social networks, especially when companionship is the main motive.Zhang, Zhao & Yang 2026, Nature Human Behaviour
Supports: the moderated association in this user sample. Unknown: who the sample is (ages unreported) and whether transcript donors differ from non-donors.
|
Attachment · Not reported Survey n = 1,131 Character.AI users + 244 donated transcripts; correlational |
Not applicable | During-use onlyAssociation | Ages not reported; offline network size analyzed as moderator Subgroups: Analyzed |
Partially supported |
| Teen-relevant online narratives describe a recognizable pattern of overreliance on companion chatbots: emotional salience, mood modification, and substitution for human contact.Razi et al. 2026, CHI ’26 (Drexel)
Supports: the pattern exists in teens’ own accounts. Does not support: prevalence, or effects, and the authors claim neither.
|
Attachment · K-12 Computational thematic analysis of self-reported Reddit narratives |
Not applicable | Not applicableNarrative | Self-identified teens who post; selection into posting acknowledged Subgroups: None |
Supported |
Gap register
Issued not addressed by any study in the tracker that could provide the basis for future research:
- – Randomizing minors to companion-AI exposure with a post-removal measure. Every causal statement about teens and attachment currently rests on adult RCTs, cross-sectional youth surveys, or narrative data.
- – Causal study of student AI use in a U.S. K-12 classroom (Stanford 2026 review).
- – Sampling foster youth, first-generation college students, international and multilingual students, neurodivergent students, or students in special education, and none analyzes effects for English learners or for students with IEP or 504 plans.
- – Measuring outcomes more than three months after AI access is removed.
- – Comparing the same tool with and without pedagogical guardrails in a companion or emotional-support context, the design that resolved the learning question in Bastani et al. (2025).
Holding queue
Preprints that meet the inclusion rule and will be appraised upon peer-reviewed publication. Listed here so the tracker can report “not yet peer-reviewed” studies for a more complete record.
- Lehmann, Cornelius & Sting 2025, AI meets the classroom: When do large language models harm learning? · arXiv 2409.09047 · in queue since 2026-10-20
- Chen et al. 2025, Generative AI alone may not be enough: Evaluating AI support for learning mathematical proof · arXiv 2509.16778 · in queue since 2026-10-20
- Fang, Liu, Danry, Maes et al. 2025, How AI and human behaviors shape psychosocial effects of chatbot use: a longitudinal randomized controlled study · arXiv 2503.17473 · in queue since 2026-10-20
- Kosmyna et al. 2025, Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing · arXiv 2506.08872 · in queue since 2026-10-20
Method: design rigor is recorded as the What Works Clearinghouse tier, cited from the Stanford AI Hub Research Repository where available. Argument rigor is scored with CERIC-T, a tracker variant of the CERIC Method built on the Cochrane RoB 2 architecture: signalling questions, three-level element judgments, and a rule-based verdict. The rubric will be published CC BY-NC with its reliability figures. Higher Learning Lab · 501(c)(3) · Last updated 2026-10-20.