In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48% but left students 17% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127% and left exam scores indistinguishable from controls.

SocioLens project,

Claim LD04-SL-C11 (v1.0)

Provenance and status

Claim type: Estimate
Publication status: Eligible for publication, with a qualifier
Verification state: Checked against the primary source. Who checked this claim is not recorded for it; claim checks so far were AI-led, and no person has checked claims against their sources.
Does not license: Long-term learning effects; adult or workplace learning
Limitations: The exam measures short-term learning in the same session (the authors say so); one school; five class sessions did not receive their assigned arm (intent-to-treat used); a 2025 correction fixes an affiliation only
Source (as cited): Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025
Source type: Peer-reviewed journal article (recorded as “Peer-reviewed journal article”)
Institutional position: Independent academic
Maturity of the evidence: Emerging evidence
Recorded date: 14 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025 (Peer-reviewed journal article)
Read the original

Version

This page shows version v1.0. Its record never changes, even when newer evidence arrives; the site may add a note to the page or withhold a figure not cleared for public use.

This pinned version is also the current version.

Versions: v1.0 · current

Raw record: every recorded field, with its coded values

All 25 canonical Claim fields, verbatim, with the coded values derived from them. The plain-language labels above are read from these.

Claim type (coded): ESTIMATE (as printed: “ESTIMATE”)
Publication status (coded): ELIGIBLE (QUALIFIED) (as printed: “ELIGIBLE WITH QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY
Source type (coded): PEER_REVIEWED_JOURNAL_ARTICLE
Maturity (coded): EMERGING

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD04-SL-C11	—
2	PUBLIC-FACING CLAIM	In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48% but left students 17% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127% and left exam scores indistinguishable from controls.	—
3	CLAIM TYPE	ESTIMATE	base_types: ESTIMATE
4	RESEARCH STRAND	B — Skill after the tool is removed	—
5	MATURITY	EMERGING	states: EMERGING
6	SOURCE	Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025	source_ids: SRC-SL-BASTANI-ET-AL-PNAS-2025
7	SOURCE TYPE	Peer-reviewed journal article	classes: PEER_REVIEWED_JOURNAL_ARTICLE
8	INSTITUTIONAL POSITION	Independent academic	—
9	SCOPE	One Turkish high school; about fifty 9th–11th-grade classes; Fall 2023; four 90-minute sessions	—
10	DESIGN	Preregistered field experiment; each session had assisted practice then a closed-book exam on similar problems	—
11	SAMPLE / DATASET	Nearly 1,000 students	—
12	MODEL / TOOL / VERSION	GPT-4 with a standard interface (GPT Base) and a tutor version with safeguards (GPT Tutor); interface details not reported here	—
13	COMPARATOR	Textbook-only practice	—
14	MAIN RESULT	Practice +48% (GPT Base) and +127% (GPT Tutor); exam: GPT Base 17% of the control mean lower, statistically significant; GPT Tutor indistinguishable from control; both tutors narrowed the spread of practice scores, but that narrowing did not persist when AI was removed; students did not perceive the harm	—
15	INFERENTIAL TARGET	Short-term learning after AI-assisted practice	—
16	DOES NOT LICENSE	Long-term learning effects; adult or workplace learning	—
17	VERIFICATION STATE	CONTENT VERIFIED	state: CONTENT_VERIFIED_PRIMARY · verified_by: UNKNOWN
18	LIMITATIONS	The exam measures short-term learning in the same session (the authors say so); one school; five class sessions did not receive their assigned arm (intent-to-treat used); a 2025 correction fixes an affiliation only	—
19	CONTRADICTORY / SUPPORTING SOURCES	Direct evidence that assisted compression need not survive tool removal; consistent with losses in a programming-library learning experiment	—
20	GEOGRAPHY / LANGUAGE	Türkiye / Not checked	—
21	DISTRIBUTIONAL LEDGER	Limited or no heterogeneity by ability, resources or effort	—
22	R01 / FIELD-MAP LINKS	LD-04, D2	—
23	WHAT WOULD CHANGE THIS CLAIM?	Delayed tests showing the exam loss fading or reversing	—
24	PUBLICATION STATUS	ELIGIBLE WITH QUALIFIER	state: ELIGIBLE · qualifiers: QUALIFIED
25	RECORDED DATE	2026-10-14	—

Field 22 references resolved against Field Map v0.2: LD-04 (LIVING_DEBATE); D2 (DOMAIN)

The sources as recorded, with their record IDs and coded source types:

Pinned Sources (field 6, version-specific)

“Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025” → Bastani et al. (2025), “Generative AI without guardrails…” (SRC-SL-BASTANI-ET-AL-PNAS-2025--v1.0) — Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025; class PEER_REVIEWED_JOURNAL_ARTICLE (explicit public provenance pin)

Cited by Living Debate versions

Cite this page

SocioLens project. (2026, October 14). In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48% but left students 17% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127% and left exam scores indistinguishable from controls [Claim LD04-SL-C11, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld04-sl-c11/v1.0/

@misc{sociolens_ld04_sl_c11_v1_0,
  author = {{SocioLens project}},
  title = {In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48\% but left students 17\% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127\% and left exam scores indistinguishable from controls.},
  year = {2026},
  date = {2026-10-14},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld04-sl-c11/v1.0/}},
  note = {Claim LD04-SL-C11, version v1.0}
}

The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.

Research-of-record provenance

Materialized by sociolens-materializer 0.6.0 at 2026-09-28T20:00:00.000Z; normalization contract SocioLens — Claim Value Vocabulary v1.0.2 — Pre-Spike Materialization Contract (v1.0.2). Content seal SHA-256 a96a8d78e92f9343299c76ddbb32c2a322afb9dd6ad4cb30a4624f1e7a5c36aa.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD04-SL-C11 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-14; no Drive identity exists. Authority: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 14 October 2026 to build the LD-04 debate; a new Claim, first version, written from the project's LD-04 source notes, checked against full texts or official abstracts. Block “ENTRY 11 — LD04-SL-C11” SHA-256 246ea388212348653f958b3fac51c263676abec16a8950c05ce365054cb25a91; fixture fixtures/authored/claims/LD04-SL-C11--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD04-SL-C11 v1.0); the first version of a new Claim of the LD-04 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD04-SL-C11 v1.0" (Project Lead (Claude, AI), 2026-10-14); basis: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 14 October 2026 to build the LD-04 debate; a new Claim, first version, written from the project's LD-04 source notes, checked against full texts or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-BASTANI-ET-AL-PNAS-2025--v1.0 — basis: field 6 of ENTRY 11 — LD04-SL-C11 cites "Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-BASTANI-ET-AL-PNAS-2025--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025".