evidoria

← Back to browse

Good practice Imported

Coursera's AI-Graded Peer Reviews for Professional Certificate Courses

United States of America · Mountain View · See the United States of America profile · See the Mountain View profile

Evidence: Observational / pre–post Top 81% 33/100 · Ask Evidence Copilot about this practice

Coursera's beta AI grader used instructor rubrics to auto-score ~300,000 text-based peer-review submissions in Professional Certificate courses, cutting grading time from ~15 hours to about a minute while independent reporting found early accuracy and consistency problems.

300000 submissions
Peer-review submissions graded by the AI beta
~15 hours to ~1 minute
Average grading time, human vs AI
45x
Increase in written feedback per response
326 characters
Average AI feedback length
90 %
Learners rating AI feedback positively
16.7 %
Increase in same-day course completions after a peer review
88% to 72%
First-attempt pass rate before and after AI grading
~3 % lower
AI grades vs prior peer grades
7 %
Learners opting back to human grading
Coursera's AI-Graded Peer Reviews for Professional Certificate Courses

Details

Maturity
Pilot
Promoter
Coursera, Inc.
Period
2024–2025
Keywords
online education, MOOCs, professional certification, edtech

Context

Coursera piloted an AI grading feature for the open-ended, text-based peer-review assignments used in many of its Professional Certificate courses, co-branded with partners such as Google, IBM and Meta, where instructors supply a grading rubric and the AI model scores learner submissions against it, supplementing or replacing peer and teaching-assistant grading; the feature is limited to text submissions and excluded from Coursera's for-credit degree programs.

Objectives

In its own beta report published September 2024, Coursera states the system graded roughly 300,000 submissions, aiming to cut grading time and deliver faster, more detailed feedback to learners.

Activities

Independent reporting from Class Central raised specific accuracy concerns with the early beta, finding that the AI grader sometimes appeared to score based on the tone or 'sentiment' of a submission rather than strict rubric alignment, produced inconsistent scores across repeated grading of the same text, and required correction from human graders, particularly those who had completed the course themselves; Duke University's continuing-education arm ran its own opt-in pilot inviting learners to compare AI and human grading directly.

Results

Coursera's data show average grading time fell from about 15 hours for human graders to about 1 minute, the system delivered roughly 45 times more written feedback per response (averaging 326 characters), 90% of learners rated the AI feedback positively, and course completions within a day of a peer review rose 16.7% after the feature's introduction; however, first-attempt pass rates fell from 88% to 72%, AI-assigned grades averaged about 3% lower than prior peer grades, and only 7% of learners opted to switch back to human grading.

Conclusions

Even adopting institutions such as Duke treated the accuracy question as still open rather than settled, reflecting a genuinely mixed picture of faster, more abundant feedback alongside a measurable drop in first-attempt pass rates and documented consistency problems in the early beta.

Implementation
Indicative cost
Low (< €50k) — No price disclosed; framed by Coursera as a grading-time reduction from ~15 hours to ~1 minute per batch.
Time to results
Short (< 1 year) — Beta reported September 2024, continuing through 2025.
Staffing & skills
Course instructors supplying grading rubrics, Teaching assistants and peer graders in a supplementary role

Conditions for success

  • Restricted to text-only submissions and excluded from for-credit degree programs
  • Instructor-supplied rubrics anchoring the AI grading rather than open-ended scoring

Common failure modes

  • Independent reporting found the AI grader sometimes scored on submission tone rather than rubric alignment
  • First-attempt pass rates fell from 88% to 72% and AI grades ran about 3% lower than prior peer grades

Where it fits

Governance type
private ed-tech platform, self-reported beta evaluation
Scale
~300,000 submissions across Professional Certificate courses
Income level
high-income (global platform)

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful