Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership
Adults / lifelong learning AI and learning Attention and memory
Links to claims
Supports (1)
| Claim | Basis |
|---|---|
| C2 | HE experiment: ChatGPT during coding → higher task scores but lower immediate and 48 h recall without the tool; kill-condition test survived 2026-09-29 (arXiv preprint). |
Review state: first-pass draft, pending human review.
Source note (Swedish, verbatim from the Kunskapsbank)
Syntes: Bergh et al. (2026) — ChatGPT programming performance vs retention
Källa: Bergh, C., Tag, B., Vassar, A., & Renzella, J. (2026). Your programming students' cognition with ChatGPT: Higher performance, lower retention, and reduced ownership. arXiv. https://doi.org/10.48550/arXiv.2609.21194
Metadata
- ID:
AI-017 - Tier: RCT (preprint)
- N / design: between-subjects; 59 first-year CS undergraduates, 55 analysed; ChatGPT-4.5 vs Google search with GenAI blocked; cued recall immediate + 48 h
- Population: vuxen / HE (introductory C programming)
- Subject: answer-capable GenAI during practice coding vs non-GenAI web search; learning measured without the tool
Finding
- ChatGPT raised coding performance (89% vs 69%; Hedges' g = −1.24 favoring AI on performance) but lowered immediate recall (41% vs 53%; g = 0.71) and 48-hour recall (39% vs 52%; g = 0.77).
- No group × time interaction on forgetting; the retention gap was set by end of session.
- Self-attributed ownership of submitted code lower with ChatGPT (45% vs 81%; g = 1.05); mental-effort trajectory flatter (tool × task interaction, Holm p = .047).
Mechanism
Cognitive offloading during answer-capable GenAI use: higher assisted performance, lower durable learning without the tool, reduced ownership.
Limits
Single university cohort; arXiv preprint (not yet peer-reviewed journal version at bank fetch); how students used ChatGPT (copy-paste solutions vs debug) not experimentally controlled; physiological channels mostly inconclusive after quality gates.
This claim dies if …
A replicated HE/K–12 study measuring without the tool finds typical answer-capable GenAI use during practice does not lower retention vs a non-GenAI control.
Swedish implication
Measure coding knowledge without the assistant after GenAI-assisted practice; performance with the tool is not learning.
Links
- DOI: https://doi.org/10.48550/arXiv.2609.21194
- Related: bastani-etal-2024-genai-harm-learning, barcaui-2025-chatgpt-cognitive-crutch, kreijkes-etal-2026-llm-note-taking, fan-etal-2025-metacognitive-laziness
Metadata as recorded
- ID
- AI-017
- Tier
- RCT
- Closed-book follow-up
- ja (cued recall immediately and 48 h later, no ChatGPT)
- Titel
- Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership
- Författare
- Bergh, C.; Tag, B.; Vassar, A.; Renzella, J.
- År
- 2026 (arXiv preprint 2609.21194)
- Källa/DOI
- https://doi.org/10.48550/arXiv.2609.21194
- URL
- https://arxiv.org/abs/2609.21194
- URL (PDF)
- https://arxiv.org/pdf/2609.21194.pdf
- Typ
- between-subjects controlled experiment (N=59; n=55 analysed; ChatGPT-4.5 vs conventional web search without GenAI)
- Åldersgrupp
- vuxen / HE (first-year CS undergraduates)
- Teman
- ai-och-larande; uppmarksamhet-och-minne
- Sparade fil
- Bergh-etal-2026-ChatGPT-programming-retention.pdf
- Hämtad
- 2026-09-29
- Paywall
- OA arXiv preprint (not yet journal-published at fetch)
Original files held in the project (not published): Bergh-etal-2026-ChatGPT-programming-retention.pdf