Claim ledger
Use AI to Learn Without Getting Worse at It
Answer-giving AI left learners worse off once it was gone. Three setups built to keep the thinking with you, and a way to check what stuck.
The evidence, row by row
Each row is a claim the essay makes, then the evidence recorded for it and the source it was checked against. For verified rows that is usually the exact words, where they sit and the day they were read. The label says how far it was checked.
What the labels mean
- Verified
- Checked against the primary source itself by a checker who did not draft the piece, usually with the exact words, where they sit and the date read.
- Executed
- A run or observation done for the piece: code, a query, a count, or a check made live in an app. The claim is what it returned. Run records are not published.
- Checked
- Checked against its source by a separate checker in the older ledger format, which recorded no quote, locator or date.
- Reported
- Not verified word for word against a primary source: carried from a secondary source, from one that could not be opened in full, or supported only in part. The essay words it accordingly.
- Struck
- Drafted, checked and removed before publication. Kept here, crossed out, with the reason.
- Excluded
- Considered and deliberately left out of the piece. Kept here with the reason.
Sources (14)
- Bastani et al. 9 rows
- Kestin et al. 9 rows
- Shen & Tamkin 10 rows
- De Simone et al. 7 rows
- OpenAI 10 rows
- LearnLM Team 6 rows
- Claude Help Center 3 rows
- Google Gemini Help 2 rows
- First-party runs 6 rows
- Anthropic "learn" skill 1 row
- The seven studies' own publication records 1 row
- Microsoft Support 1 row
- Claude Code docs 1 row
- Lehmann 4 rows
-
In a Turkish high-school trial, a plain ChatGPT-style assistant raised practice scores by 48% and a hint-only tutor by 127%
"GPT Base and GPT Tutor would increase performance on the assisted practice sessions by 48% and 127%, respectively"
Bastani et al., PNAS 2025 · Main Results; Table 1 col (1) · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Main Results; Table 1 col (1), https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B1 (source read 2026-09-24). -
Once the AI was taken away, the plain-assistant group scored 17% lower on the exam than students who never had it
"GPT Base diminished the average control student's performance on the unassisted exam by 17%."
Bastani et al., PNAS 2025 · Main Results; Table 1 col (2) · pmc.ncbi.nlm.nih.gov
LinkReport an errorCite
Bastani et al., PNAS 2025, Main Results; Table 1 col (2), https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B2 (source read 2026-09-24). -
The analysis the authors registered in advance finds a smaller exam harm than the headline (−0.035, CI −0.057 to −0.012), still negative
"GPT Base vs. Control ... −0.035 [−0.057, −0.012] 0.003"
Bastani et al., PNAS 2025, SI · SI Appendix A.6, Table A.2 · pmc.ncbi.nlm.nih.gov
LinkReport an errorCite
Bastani et al., PNAS 2025, SI, SI Appendix A.6, Table A.2, https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B3 (source read 2026-09-24). -
The hint-only tutor removed the exam harm but produced no exam gain
"this negative effect is essentially eradicated in the GPT Tutor arm, though we still do not observe a positive effect."
Bastani et al., PNAS 2025 · Main Results; Introduction · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Main Results; Introduction, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B4 (source read 2026-09-24). -
The tutor's prompt carried the correct solution and told it not to give the whole solution away
"GPT-4 is given a prompt including the solution to each problem (to mitigate hallucinations) as well as instructions to avoid giving away the entire solution"
Bastani et al., PNAS 2025 · Footnote †; Experimental Design · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Footnote †; Experimental Design, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B5 (source read 2026-09-24). -
Without the answer key, the plain assistant got the practice problems right only 51% of the time
"GPT Base gives a correct answer only 51% of the time on average"
Bastani et al., PNAS 2025 · "GPT Errors vs. Student Performance"; Fig. 2; SI C.1 · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, "GPT Errors vs. Student Performance"; Fig. 2; SI C.1, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B6 (source read 2026-09-24). -
Students used the plain assistant as a crutch, asking for and copying solutions
"students often use GPT Base as a "crutch" by asking for and copying solutions"
Bastani et al., PNAS 2025 · Introduction; Potential Mechanism · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Introduction; Potential Mechanism, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B7, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B7 (source read 2026-09-24). -
The trial covered nearly 1,000 students
"We conducted four 90-min sessions for about fifty 9th, 10th, and 11th-grade classes, comprising nearly 1,000 students."
Bastani et al., PNAS 2025 · Experimental Design · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Experimental Design, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B8, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B8 (source read 2026-09-24). -
The trial measured short-term outcomes only
"we focus on short-term outcomes due to limitations imposed by our partner school"
Bastani et al., PNAS 2025 · Discussion, final paragraph · pnas.org
LinkReport an errorCite
Bastani et al., PNAS 2025, Discussion, final paragraph, https://www.pnas.org/doi/10.1073/pnas.2422633122. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row B9, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B9 (source read 2026-09-24). -
In a Harvard physics trial, students' median post-test was 4.5 with the AI tutor against 3.5 in an active-learning class
"higher median (M) post-score (M = 4.5, N = 142) compared to those in the in-class active learning group (M = 3.5, N = 174)"
Kestin et al., Scientific Reports 2025 · Results § Learning gains · nature.com
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Results § Learning gains, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K1 (source read 2026-09-24). -
Median learning gains with the AI tutor were more than double the class's
"in the AI-tutored group were over double those for students in the in-class active learning group"
Kestin et al., Scientific Reports 2025 · Results § Learning gains · pmc.ncbi.nlm.nih.gov
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Results § Learning gains, https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K2 (source read 2026-09-24). -
The regression effect size was 0.63
"While the linear regression suggests an effect size of 0.63, this is an underestimation due to ceiling effect"
Kestin et al., Scientific Reports 2025 · Results § Linear regression model; Table S1 · pmc.ncbi.nlm.nih.gov
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Results § Linear regression model; Table S1, https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K3 (source read 2026-09-24). -
The Harvard tutor's instructions released one step at a time and kept replies brief
"Only give away ONE STEP AT A TIME, DO NOT give away the full solution in a single message" / "Keep responses BRIEF (a few sentences or less) but helpful."
Kestin et al., Supplementary Material 1 · Supplement, system prompt · nature.com
LinkReport an errorCite
Kestin et al., Supplementary Material 1, Supplement, system prompt, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K4 (source read 2026-09-24). -
The Harvard team loaded each question with its step-by-step worked answer
"we enriched our prompts with comprehensive, step-by-step answers"
Kestin et al., Scientific Reports 2025 · Discussion § Designing successful student-AI interactions · nature.com
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Discussion § Designing successful student-AI interactions, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K5 (source read 2026-09-24). -
194 students took part
"Of the 233 enrolled students, 194 were eligible for inclusion in the study."
Kestin et al., Scientific Reports 2025 · Methods § Study population · nature.com
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Methods § Study population, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K6 (source read 2026-09-24). -
Each student learned one lesson with the AI tutor and another in the active-learning class (crossover)
"Given that our experiment is a crossover design in which each student experiences both conditions"
Kestin et al., Scientific Reports 2025 · Results (regression controls); Methods § Study design · pmc.ncbi.nlm.nih.gov
LinkReport an errorCite
Kestin et al., Scientific Reports 2025, Results (regression controls); Methods § Study design, https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K9, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K9 (source read 2026-09-24). -
In an Anthropic trial, 52 developers learned a new Python library, half with an AI assistant
"In our main study, 52 participants completed the task, 26 for each of the control and treatment groups."
Shen & Tamkin, arXiv 2601.20245 · §5.2.1 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §5.2.1, https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A1 (source read 2026-09-24). -
The AI group scored 4.15 points lower on a 27-point quiz, about two grade points
"There is a 4.15 point difference between the means ... For a 27-point quiz, this translates into a 17% score difference or 2 grade points."
Shen & Tamkin, arXiv 2601.20245 · §5.2.2 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §5.2.2, https://arxiv.org/html/2601.20245v2. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A2 (source read 2026-09-24). -
Nobody could use AI during the quiz
"All participants were not allowed to use AI in the comprehension check."
Shen & Tamkin, arXiv 2601.20245 · Figure 4 caption · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, Figure 4 caption, https://arxiv.org/html/2601.20245v2. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A3 (source read 2026-09-24). -
Using AI did not make them significantly faster
"While using AI to complete our coding task did not significantly improve task completion time, the level of skill formation gained ... is significantly reduced."
Shen & Tamkin, arXiv 2601.20245 · §1 / §5.2.2 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §1 / §5.2.2, https://arxiv.org/html/2601.20245v2. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A4 (source read 2026-09-24). -
High-scoring ways of using the AI averaged 65% to 86% on the quiz; low-scoring ways 24% to 39% (post-hoc, small groups)
"high-scoring interaction patterns (65%-86% quiz score) vs low-scoring interaction patterns (24%-39% quiz score)"
Shen & Tamkin, arXiv 2601.20245 · §6; Figure 11 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §6; Figure 11, https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A5 (source read 2026-09-24). -
Asking the AI conceptual questions was a high-scoring pattern, and the fastest of them
"On average, this mode was the fastest among high-scoring patterns and second fastest overall after the AI Delegation mode."
Shen & Tamkin, arXiv 2601.20245 · §6; Figure 11 (Conceptual Inquiry) · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §6; Figure 11 (Conceptual Inquiry), https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A6 (source read 2026-09-24). -
In a World Bank trial in Nigeria, students had twelve 90-minute sessions guided by teachers
"Those assigned to the intervention attended twelve 90-minute sessions in computer labs, engaging in curriculum-aligned activities guided by teachers."
De Simone et al., World Bank PRWP 11125 · §1, p. 2 · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, §1, p. 2, https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N1 (source read 2026-09-24). -
English scores rose 0.238 standard deviations, which the authors equate to 1.5 years of ordinary schooling
"Our ITT effect of 0.238 standard deviation in English is equivalent to increasing 1.5 years of 'business-as-usual' schooling in Nigeria"
De Simone et al., World Bank PRWP 11125 · §4.1, p. 18 · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, §4.1, p. 18, https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N2 (source read 2026-09-24). -
In the sample week's starting prompt, the chatbot answered the student's question, then set exercises on which it was told to hint after a wrong reply and give the answer only if the reply was still wrong
"provide encouraging words and provide a hint" … "if my reply is still incorrect"
De Simone et al., World Bank PRWP 11125 · Figure 11, PDF p. 51 (Week 7 starting prompt, read from rendered image) · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, Figure 11, PDF p. 51 (Week 7 starting prompt, read from rendered image), https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N3 (source read 2026-09-24). -
The control group got no programme at all, so extra time and attention are not separated from the AI
"the control group, which did not receive any intervention but continued their regular learning in the classroom"
De Simone et al., World Bank PRWP 11125 · §2.2, p. 9 · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, §2.2, p. 9, https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N4 (source read 2026-09-24). -
The authors credit the whole package (prompts plus teacher guidance), not the chatbot alone
"we interpret that the intervention as a whole -which includes the interaction with the LLM and teacher guidance with specific prompts- is driving the results."
De Simone et al., World Bank PRWP 11125 · §4.3, p. 21 · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, §4.3, p. 21, https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N5 (source read 2026-09-24). -
Nigeria's third-term exam (effect 0.206 SD) came after the intervention on content not limited to it; it ran the day after the last session, so it is not a test of whether the learning lasted
"the third term exam score, with an effect size of 0.206 standard deviation (SE = 0.067), although this exam was not limited to the intervention's specific content"
De Simone et al., World Bank PRWP 11125 · §4.1; timing from Appendix timeline (pilot sessions end 7/11/24, third-term exam 7/12/24) · documents1.worldbank.org
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, §4.1; timing from Appendix timeline (pilot sessions end 7/11/24, third-term exam 7/12/24), https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row N6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N6 (source read 2026-09-24). -
OpenAI's own trial ran over 300 college students through study sessions of a nominal 40 minutes
"we ran a randomized study with over 300 college students preparing for neuroscience and microeconomics exams" / "the nominal 40 minute sessions"
OpenAI, "New tools for understanding AI and learning outcomes" (2026-03-04) · "Origins and early research"; "Study design" · openai.com
LinkReport an errorCite
OpenAI, "New tools for understanding AI and learning outcomes" (2026-03-04), "Origins and early research"; "Study design", https://openai.com/index/understanding-ai-and-learning-outcomes/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row O1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O1 (source read 2026-09-24). -
Study mode scored roughly 15% higher in microeconomics
"roughly a 15% higher score relative"
OpenAI (2026-03-04) · "Findings", bullet 2 · openai.com
LinkReport an errorCite
OpenAI (2026-03-04), "Findings", bullet 2, https://openai.com/index/understanding-ai-and-learning-outcomes/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row O2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O2 (source read 2026-09-24). -
In neuroscience, study mode was not distinguishable from studying with ordinary online resources
"results were not distinguishable from students studying with traditional online resources"
OpenAI (2026-03-04) · "Findings", bullet 1 · openai.com
LinkReport an errorCite
OpenAI (2026-03-04), "Findings", bullet 1, https://openai.com/index/understanding-ai-and-learning-outcomes/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row O3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O3 (source read 2026-09-24). -
The comparison group used ordinary online resources, not ordinary ChatGPT
"a control group studied using traditional online resources such as Google Search and YouTube, with AI generated overview features disabled"
OpenAI (2026-03-04) · "Study design", para 1 · openai.com
LinkReport an errorCite
OpenAI (2026-03-04), "Study design", para 1, https://openai.com/index/understanding-ai-and-learning-outcomes/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row O4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O4 (source read 2026-09-24). -
OpenAI calls the results early; analysis is still underway
"While analysis is still underway, early results give us confidence"
OpenAI (2026-03-04) · "Origins and early research" · openai.com
LinkReport an errorCite
OpenAI (2026-03-04), "Origins and early research", https://openai.com/index/understanding-ai-and-learning-outcomes/. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row O5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O5 (source read 2026-09-24). -
Google's own trial of Gemini Guided Learning in Sierra Leone raised maths scores 0.258 standard deviations (CI 0.027 to 0.488)
"yielding a gain of +0.258 standard deviations across the Guided Learning classrooms (intent to treat; 95% confidence interval [0.027, 0.488]"
LearnLM Team, Google & Fab AI, tech report (May 2026) · p. 3; Table C.4 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team, Google & Fab AI, tech report (May 2026), p. 3; Table C.4, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G1 (source read 2026-09-24). -
It enrolled 1,763 students in 48 grade 7 and 8 classrooms
"The trial enrolled 𝑁= 1763 students aged 13 or older in 48 grades 7 and 8 classrooms."
LearnLM Team tech report · p. 2 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team tech report, p. 2, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G2 (source read 2026-09-24). -
Students shared devices in pairs, in lessons teachers ran
"Students accessed the Gemini app on tablets or desktop computers, sharing at a 2:1 student-to-device ratio."
LearnLM Team tech report · p. 2 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team tech report, p. 2, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G3 (source read 2026-09-24). -
The gain came from grade 8; the grade 7 estimate was slightly negative (−0.078; grade 8 interaction 0.429)
"Arm: Treatment –0.078* (0.037)"; "Arm: Treatment × Grade: Grade 8 0.429** (0.142)"
LearnLM Team tech report · Table C.14, p. 28 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team tech report, Table C.14, p. 28, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G4 (source read 2026-09-24). -
Gemini gave a direct solution in 2.1% of its messages (Gemini-classified)
"far more frequently than providing direct solutions (2.1%)"
LearnLM Team tech report · p. 4 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team tech report, p. 4, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G5 (source read 2026-09-24). -
The report has not been peer reviewed
"Please note works submitted as a preprint have not undergone a peer review process."
LearnLM Team tech report · p. 8 · storage.googleapis.com
LinkReport an errorCite
LearnLM Team tech report, p. 8, https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row G6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G6 (source read 2026-09-24). -
OpenAI's own help page says Study mode may still give a direct answer
"there may be times when it gives a direct answer"
OpenAI Help Center, "Using study mode in ChatGPT" · Article body · help.openai.com
LinkReport an errorCite
OpenAI Help Center, "Using study mode in ChatGPT", Article body, https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D1 (source read 2026-09-24). -
Study mode is not available inside ChatGPT Projects
"The Study option is not available in Temporary Chats, GPTs, or Projects."
OpenAI Help Center · Article body · help.openai.com
LinkReport an errorCite
OpenAI Help Center, Article body, https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D2 (source read 2026-09-24). -
A ChatGPT scheduled task made inside a project cannot read that project's files
"If you create a task in a project, it cannot access uploaded files or files stored in that project."
OpenAI Help Center, Tasks · Article body · help.openai.com
LinkReport an errorCite
OpenAI Help Center, Tasks, Article body, https://help.openai.com/en/articles/10291617. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D3 (source read 2026-09-24). -
Claude's free plan allows up to five Projects
"Free users can create a maximum of five projects."
Claude Help Center · Article body · support.claude.com
LinkReport an errorCite
Claude Help Center, Article body, https://support.claude.com/en/articles/9519177. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D4 (source read 2026-09-24). -
Gemini Guided Learning sits under Add files, then More tools
"In the text box, click Add Files." / "At the bottom, click More tools" … "Guided Learning"
Google Gemini Help · "Use Guided Learning", steps 2–3 · support.google.com
LinkReport an errorCite
Google Gemini Help, "Use Guided Learning", steps 2–3, https://support.google.com/gemini/answer/16448384. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D5 (source read 2026-09-27). -
Study appears as a composer mode in the ChatGPT web app (Plus)
Result: The author's screenshot (not published): composer shows "Study" mode chip
First-party: ChatGPT web app (Plus), the author's run
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: ChatGPT web app (Plus), the author's run). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X1 (source read 2026-09-24). -
In 8 runs of the final Level 2 prompt (7 across Claude, Gemini and GPT, 1 in the ChatGPT app), the tutor withheld the answer on a direct demand and, after the second wrong try, gave the answer (some models per step) and named the missed step; once (GPT, code) it described the fix in words while withholding the code
First-party: python3 run.py, run_gemini.py, run_codex.py + the author's ChatGPT run
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: python3 run.py, run_gemini.py, run_codex.py + the author's ChatGPT run). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X2 (source read 2026-09-24). -
In a simulated project (instructions plus file given to the model directly), the Level 3 setup stuck to the file, said when the file did not cover a question, and scored a no-feedback quiz correctly on all three model families; the Claude and Gemini apps themselves were not tested
First-party: python3 l3.py, l3x.py
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: python3 l3.py, l3x.py). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X3 (source read 2026-09-24). -
In the ChatGPT app, our first Level 3 instructions answered a "just tell me" question at once from general knowledge; the revised instructions held it back and answered from the file
First-party: ChatGPT web (Plus), Project "Cold Check Test", the author's run (v2) + Claude via Claude in Chrome (v3)
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: ChatGPT web (Plus), Project "Cold Check Test", the author's run (v2) + Claude via Claude in Chrome (v3)). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X4 (source read 2026-09-24). -
In the ChatGPT app, the cold-check prompt held all feedback to the end and scored a 3-of-6 quiz correctly against the file
First-party: ChatGPT web (Plus), Project "Cold Check Test", driven by Claude via Claude in Chrome
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: ChatGPT web (Plus), Project "Cold Check Test", driven by Claude via Claude in Chrome). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X5, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X5 (source read 2026-09-24). -
In Claude.ai, the old Styles menu is gone; Anthropic now offers an official "learn" skill (Customize → Skills → Discover → learn → Add) whose stated goal is to help the learner answer it themselves, and which does not trigger on coding tasks or factual lookups
"The goal is not to answer the learner's question but to help them be able to answer it themselves" / "Don't trigger for: Tasks: coding, writing, calculation, translation, factual lookup"
Anthropic "learn" skill (updated Sep 14), SKILL.md + Overview, read in the live Claude.ai app · Skill Overview + Contents › SKILL.md "Learning Mode" · claude.ai
LinkReport an errorCite
Anthropic "learn" skill (updated Sep 14), SKILL.md + Overview, read in the live Claude.ai app, Skill Overview + Contents › SKILL.md "Learning Mode", https://claude.ai/new#customize/skills/id/learn. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D6 (source read 2026-09-24). -
They were Python programmers recruited through a crowd-work platform; 29 of the 52 had seven or more years of coding experience
Paper Table 1 (experience bands; 7+ years = 29 of 52) / "the small group size of the 1-3 year participant group (n=4)"
Shen & Tamkin, arXiv 2601.20245 · Table 1; §5.2.2 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, Table 1; §5.2.2, https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A7, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A7 (source read 2026-09-24). -
Only two of the seven studies went through journal peer review (PNAS and Scientific Reports); the others are two preprints, a working paper, a vendor tech report and a vendor blog post
Result: PNAS 122(26); Scientific Reports 15, 17458; arXiv 2601.20245 and arXiv 2409.09047 (preprints); World Bank Policy Research Working Paper 11125; Google tech report "have not undergone a peer review process"; OpenAI "While analysis is still underway"
The seven studies' own publication records
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (The seven studies' own publication records). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row S2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-S2 (source read 2026-09-24). -
Microsoft Copilot has a "Study and learn" mode, chosen from the Quick response menu under the prompt
"Press (or tap) 'Quick response' under your prompt and select an appropriate mode"
Microsoft Support, "Conversation modes in Microsoft Copilot" · Article body · support.microsoft.com
LinkReport an errorCite
Microsoft Support, "Conversation modes in Microsoft Copilot", Article body, https://support.microsoft.com/en-us/microsoft-copilot/conversation-modes-in-microsoft-copilot. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D7, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D7 (source read 2026-09-24). -
Claude Code has a Learning output style, switched on with /output-style learning
"/output-style learning" (Learning style: Claude "stops and waits" at TODO(human) markers)
Claude Code docs, Output styles · Built-in styles section · code.claude.com
LinkReport an errorCite
Claude Code docs, Output styles, Built-in styles section, https://code.claude.com/docs/en/output-styles. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D8, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D8 (source read 2026-09-24). -
The Harvard tutor's instructions allowed it to give the answer if the student demanded it
"DO NOT not tell them the answer UNLESS they demand you to give them the answer"
Kestin et al., Supplementary Material 1 · Supplement, system prompt · nature.com
LinkReport an errorCite
Kestin et al., Supplementary Material 1, Supplement, system prompt, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K7, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K7 (source read 2026-09-24). -
The Harvard tutor's instructions asked students to try first
"encourage them to give it a try first"
Kestin et al., Supplementary Material 1 · Supplement, system prompt · nature.com
LinkReport an errorCite
Kestin et al., Supplementary Material 1, Supplement, system prompt, https://www.nature.com/articles/s41598-025-97652-6. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row K8, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K8 (source read 2026-09-24). -
The highest-scoring pattern (2 people) had the AI generate code and then asked it questions to understand it, averaging 86%; asking conceptual questions averaged 65%; the lowest patterns, which delegated or relied on it, averaged 24% to 39%
"high-scoring interaction patterns (65%-86% quiz score) vs low-scoring interaction patterns (24%-39% quiz score)"
Shen & Tamkin, arXiv 2601.20245 · §1.1; §6; Figure 11 (Generation-Then-Comprehension n=2 86%; Conceptual Inquiry n=7 65%) · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, §1.1; §6; Figure 11 (Generation-Then-Comprehension n=2 86%; Conceptual Inquiry n=7 65%), https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A8, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A8 (source read 2026-09-24). -
The low-scoring patterns were delegation, progressive reliance and iterative AI debugging (relying on the AI to debug or verify their code)
"Iterative AI Debugging (n=4): Participants in this group relied on AI to debug or verify their code."
Shen & Tamkin, arXiv 2601.20245v2 · §6, interaction patterns · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245v2, §6, interaction patterns, https://arxiv.org/html/2601.20245v2. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A10, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A10 (source read 2026-09-24). -
In pre-registered lab experiments on learning to code, giving students an AI chatbot had no overall effect on learning
"we find no effect of LLMs on overall learning outcomes"
Lehmann, Cornelius & Sting, arXiv 2409.09047v2 · Abstract; Tables 5, 7 · arxiv.org
LinkReport an errorCite
Lehmann, Cornelius & Sting, arXiv 2409.09047v2, Abstract; Tables 5, 7, https://arxiv.org/abs/2409.09047. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row L1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L1 (source read 2026-09-24). -
Students who used it to substitute for their own work (e.g. generating solutions to exercises) covered more topics but understood less; students who used it to complement their work (e.g. asking for explanations) understood more (exploratory, not randomised)
"Students who substitute some of their learning activities with LLMs (e.g., by generating solutions to exercises) increase the volume of topics" / "Students who complement their learning activities with LLMs (e.g., by asking for explanations) do not increase topic volume but do increase their understanding."
Lehmann et al., arXiv 2409.09047v2 · Abstract; §9 · arxiv.org
LinkReport an errorCite
Lehmann et al., arXiv 2409.09047v2, Abstract; §9, https://arxiv.org/abs/2409.09047. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row L2, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L2 (source read 2026-09-24). -
42% of messages asking for a solution were sent without a single attempt
"42% of messages asking for a solution were sent without a single attempt"
Lehmann et al., arXiv 2409.09047v2 · §9.1 (Study 3) · arxiv.org
LinkReport an errorCite
Lehmann et al., arXiv 2409.09047v2, §9.1 (Study 3), https://arxiv.org/abs/2409.09047. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row L3, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L3 (source read 2026-09-24). -
The chatbot raised how much students felt they had learned by more than their actual learning explains
"LLMs increase perceived learning by more than can be explained by actual differences in learning"
Lehmann et al., arXiv 2409.09047v2 · §9.3 · arxiv.org
LinkReport an errorCite
Lehmann et al., arXiv 2409.09047v2, §9.3, https://arxiv.org/abs/2409.09047. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row L4, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L4 (source read 2026-09-24). -
Claude Skills need code execution switched on (Settings > Capabilities > Code execution and file creation on Free, Pro and Max)
"This feature requires code execution to be enabled."
Claude Help Center, "Use Skills in Claude" · Article body · support.claude.com
LinkReport an errorCite
Claude Help Center, "Use Skills in Claude", Article body, https://support.claude.com/en/articles/12512180-use-skills-in-claude. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D9, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D9 (source read 2026-09-24). -
ChatGPT's Study mode is chosen by typing @study or from the + menu, and is on every plan
"Available across ChatGPT plans globally on web, iOS, and Android"
OpenAI Help Center, "Using study mode in ChatGPT" · Article body · help.openai.com
LinkReport an errorCite
OpenAI Help Center, "Using study mode in ChatGPT", Article body, https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D10, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D10 (source read 2026-09-24). -
The cold check with a topic line kept to the named topics and scored correctly in the ChatGPT app (3 of 6, the planted misses exactly), and kept to topic on 17 of 18 questions across Claude, Gemini and GPT
First-party: ChatGPT web (Plus) via Claude in Chrome + l3_cc2.py, l3x_cc2.py
A run or observation done for this piece. The run record is not published.
LinkReport an errorCite
A run or observation by The Durability Curve (First-party: ChatGPT web (Plus) via Claude in Chrome + l3_cc2.py, l3x_cc2.py). Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row X6, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X6 (source read 2026-09-24). -
Tutor prompts built to make students work before giving answers were published by Ethan and Lilach Mollick in 2023, and the Nigeria programme borrowed from them
"Some of the prompt structures were derived from Mollick and Mollick (2023a)"
De Simone et al., World Bank PRWP 11125, fn 3; Mollick & Mollick, "Assigning AI" (SSRN 4475995) · §2.1, pp. 7 to 8 (main text) · papers.ssrn.com
LinkReport an errorCite
De Simone et al., World Bank PRWP 11125, fn 3; Mollick & Mollick, "Assigning AI" (SSRN 4475995), §2.1, pp. 7 to 8 (main text), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4475995. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row P1, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-P1 (source read 2026-09-24). -
The no-AI group averaged about 65% on the quiz
Figure 6 plots about 65.4% for the control group (read from marker positions); the Anthropic blog gives 67%
Shen & Tamkin, arXiv 2601.20245, Figure 6; Anthropic blog · Figure 6 · arxiv.org
LinkReport an errorCite
Shen & Tamkin, arXiv 2601.20245, Figure 6; Anthropic blog, Figure 6, https://arxiv.org/abs/2601.20245. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row A9, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A9 (source read 2026-09-24). -
ChatGPT project: New project in the sidebar; instructions via the project's settings
"Select New project in the sidebar." / "Select the more options menu (•••), then select Project settings to add instructions for the project."
OpenAI Help Center, Projects in ChatGPT · "Create a project"; "Add project instructions" · help.openai.com
LinkReport an errorCite
OpenAI Help Center, Projects in ChatGPT, "Create a project"; "Add project instructions", https://help.openai.com/en/articles/10169521. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D11, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D11 (source read 2026-09-27). -
Claude project: Projects, then New Project, with instructions and knowledge files
"click “Projects,”" … "Click "+ New Project" in the upper right corner." / "Anything you upload to this space will be used across all of your chats within that project." / "Click on "Set project instructions.""
Claude Help Center, projects article · "How to create a project"; "Add content to project knowledge"; "Add project instructions" · support.claude.com
LinkReport an errorCite
Claude Help Center, projects article, "How to create a project"; "Add content to project knowledge"; "Add project instructions", https://support.claude.com/en/articles/9519177. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D12, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D12 (source read 2026-09-27). -
Gemini: a Gem (Gems, then New Gem) with files added under Knowledge
Help-page path: Gems → New Gem → Knowledge → Add files
Google Gemini Help · Article body · support.google.com
LinkReport an errorCite
Google Gemini Help, Article body, https://support.google.com/gemini/answer/15146780. Checked in: Claim Ledger, "Use AI to Learn Without Getting Worse at It", The Durability Curve, row D13, https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D13 (source read 2026-09-24).
No rows match. Clear the search or pick another label.
Cite and reuse
- A single claim
- Use the row's Cite button. It names the original source first, then this row, which carries its own link.
- This ledger
Floyd, Harry (2026). Claim Ledger: Use AI to Learn Without Getting Worse at It. The Durability Curve. https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/- The essay
Floyd, Harry (2026). Use AI to Learn Without Getting Worse at It. The Durability Curve. https://durabilitycurve.com/blog/learn-with-ai-without-getting-worse/
The ledger data is licensed CC BY 4.0: reuse it, quote it or train on it, with credit to The Durability Curve and a link. Read the licence. Machine copies: JSON, CSV, Markdown.