Three Free AI Chatbots Did Every Graded Task in One Intro Accounting Course. All Three Earned an A

A student pastes your homework into a free chatbot and submits the reply. Could it pass your course alone?

Orrin Swift of Lamar University in Texas tested the tools, not students (pp. 135, 137).

What he tested

Swift gave three free AI chatbots every graded task in an introductory financial accounting course (p. 137). He says ChatGPT 5, Gemini 2.5 and Claude Sonnet 4.5 were chosen by summer 2025 leaderboard ranking (p. 138).

These were 76 homework assignments, two exams and the Jackson Cycle project, a full accounting cycle from source documents to closing (p. 137).

Each model got only the students' instructions and materials, with no worked examples or corrective feedback (p. 138). A graduate research assistant marked them with the student rubrics, and the instructor verified them (p. 138).

What the models scored

All three averaged above 99% on homework, with minor errors mostly in formatting, subtotals or labels (p. 139). Exams were perfect or near-perfect, and the summary table shows 100 for all three (pp. 139, 141).

On the project, Gemini scored 98; Claude scored 85 after an early cash misstatement carried through several statements (p. 140). ChatGPT first left out the unadjusted trial balance and scored 85, or 93 with it included (p. 140). The paper does not say how that step was added, and its tables give different scores (pp. 138, 141).

Final grades ran from 96.6 to 99.4, all an A, against a historical student average of 78.4 (pp. 138, 141–142). Even the lowest project score beat the students' project average of 74 (p. 138).

Swift concludes that long chains of dependent steps remain these tools' relative weakness (pp. 141–142).

Limits and context

This measured the tools on one course's tasks, not anyone's learning. No students took part, and neither student average has a stated size.

Swift notes the course's tabular format may favour AI pattern recognition (p. 137). He calls his case study unable to support formal hypothesis testing (p. 142). Testing dates and attempts per model go unreported.

Evidence strength: limited and preliminary. One course and three free models, marked by rubric, with no statistical tests. It shows what the tools produced, not how students learn.