The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2480
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:40:13.794348Z
status: published
lens: g_space
sector: education
headline: Let’s Chat: Leveraging Chatbot Outreach for Improved Course Performance
dek: This study provides pre-registered, experimental evidence on the use of non-generative artificial intelligence (AI) chatbots to support students in large-enrollment undergraduate courses. We find the chatbot messaging increased students’ final grades and engagement with academic supports, such as tutoring. Treatment effects were generally consistent across student demographics, with the exception of treated women in a Microeconomics course, who earned final grades that were seven percentage points higher than wo…
gain_title: In randomized trials covering 2,483 students at Georgia State University, students assigned to course chatbot texts were four percentage points more likely to earn an A or B than the 61% of control students who did.
problem_title: (none)
trace_subject: (none)
gain_reading: In randomized trials covering 2,483 students at Georgia State University, students assigned to course chatbot texts were four percentage points more likely to earn an A or B than the 61% of control students who did.
gain_evidence: we randomized a total of 1,568 students enrolled in Government | we randomized 915 students enrolled in Microeconomics
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers ran randomized controlled trials of a text-message chatbot with non-generative AI in two large online courses at Georgia State University: Introduction to American Government and Principles of Microeconomics. Students assigned to the chatbot were four percentage points more likely to earn an A or B than the 61% of control students who did, about a 7% relative increase.

Effects varied by course and group. In Microeconomics, women assigned to the chatbot earned final grades seven points higher than control women, were 11 percentage points more likely to earn an A or B and were six percentage points less likely to drop the course. Men in that course showed no effect. In Government, treated students were three percentage points less likely to earn a D or F or withdraw, against 18% of controls.

The chatbot did not appear to change students' wider academic habits. Researchers found no strong evidence of effects on other courses taken that term or on enrolling in another course in the subject the next term. Treated students were about two percentage points more likely to attend supplemental instruction tutoring, which explained a small share of the grade gain.
limitation: Only about half of Government students completed the end-of-course survey, with response rates lower for Black students, course retakers and students with lower high school GPAs, and no survey was fielded in Microeconomics; the authors also caution against over-interpreting the mechanism results.
tag: Evidence-backed gain
key_points: The trials randomized 1,568 Government students across fall 2021, spring 2022 and fall 2022, and 915 Microeconomics students during 2022-23; all had opted in to university texting. | Treated students received 2-3 customized texts a week (about 40 a semester) covering due dates, missing-assignment nudges and invitations to ask questions. | Chatbot answers came from a pre-programmed knowledge base; low-confidence questions went to a teaching assistant, who also reviewed bot replies daily. | Treated students were four percentage points more likely to earn an A, against 36% of controls, though after multiple-comparison correction only the gain in earning a B or higher stayed significant. | In Microeconomics, treated students were three percentage points less likely to drop the course than the 8% of controls who did. | Treated women in Microeconomics completed practice assignments 8-9 percentage points more often than control women; assignment completion explained 48-73% of their grade gain. | Supplemental instruction attendance rose about two percentage points against baselines of 7% in Government and 2.4% in Microeconomics. | On an end-of-course survey in Government, 82% of respondents across the pooled results recommended continuing the chatbot.
rundown: Georgia State University, which enrolls more than 52,000 undergraduates and where 63% of students identify as Black, Hispanic or of two or more races, chose large, asynchronous online courses with high rates of D, F or withdrawal. Half of the consenting students in each course were randomly assigned to chatbot texts, and all students still received the usual instructor emails.

The texts were personalized by whether a student had completed the prior week's work. Students could message the bot at any hour and get answers from a knowledge base built with administrators and the instructional team. The Government version added a #quizme function that offered multiple-choice practice before exams; about 38% of surveyed treated students said they used it.

The headline grade effect was similar across the two courses and across most demographic subgroups, with formal equality tests finding five significant differences in 42 comparisons. The main exception was women in Microeconomics, who ended up outperforming both control and treated men; control women and men had averaged 68.9 and 70.5.

Beyond grades, the researchers found no consistent treatment effects on attitude measures in the Government survey. Only about half of Government students answered it, and response rates were lower for Black students, course retakers and students with lower high school GPAs.
sources:
- peer_reviewed | NBER Working Paper | https://doi.org/10.3386/w35397 | 2026-07-01
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

b770b0d2fb562ccc48935ec9506717dce3c8c45e41d0fc30c9371d2f51cb42af

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000