Georgia Tech has published its first Academic AI Playbook, a free 278-page volume in which faculty from all seven of its colleges document 67 AI initiatives from real courses: the setting, what the instructor tried, what happened and what they would change. Instead of another forecast about AI in education, it offers designs that have already met students, and its closing chapter condenses them into 20 lessons that run from tutor architecture to grading policy.
How the book came together
The editors, David Joyner, interim vice provost for AI in Education, and Nimisha Roy, director of online undergraduate initiatives in the College of Computing, explain the selection in their opening letter. A first internal call led to 73 invitations for full articles, and a second call brought another 38 proposals. Of those 111 candidates, 67 became articles, so by our count roughly six in ten made it into print.
The editors are open about their own tooling. AI-generated code compiled the volume, sent drafts to authors and inserted cross-links between articles, and AI also proposed candidate lessons and supporting quotes for the final chapter. The decisions about what to include and which lessons to draw were made by the authors and editors, they write.
The lessons that go against the default
Several of the 20 lessons push back on how commercial AI tools behave out of the box. These stood out in our reading:
- Grounding beats scale. Every team that built a tutor or study tool reached the same conclusion: students notice whether answers are anchored in their own course materials, not how powerful the model is. The StudyBuddy team advises spending the engineering budget on retrieval and document ingestion “before chasing a larger model.”
- Design the tool to resist. Several instructors had to put friction back in, because assistants tuned to be agreeable remove the struggle that learning needs. One team’s logs showed that higher-performing students asked concept questions while lower-performing students more often asked for direct answers.
- Decide the data question on day one. Teams that treated data residency as a founding constraint, such as the StudyBuddy team with its locally run models, could still make credible privacy claims. Commitments made late were, in the book’s words, impossible.
- Do not build on a capability gap. Rules based on what AI cannot yet do expire quickly. Banerjee, writing about a physics course, puts it as a “shelf life measured in model generations” and argues for designing around what a discipline requires instead.
- Assess the process, not the artifact. Weekly chat-log submissions, progress reports, debates and peer review gave instructors a view of the reasoning behind a final submission.
- Protect human work deliberately. Some courses moved core work into electronics-free workshops or group formats, which helped both learning and assessment integrity.
Where it fits in the assessment debate
The timing is useful. On November 11, England’s Office for Students warned that AI has made some coursework close to impossible to assess in the old way and pointed universities toward process-based formats, as we reported. The regulator described the problem; lessons 11 to 13 in the Georgia Tech book describe what process-based assessment looks like in individual courses, from chat-log portfolios to in-class workshops.
It also differs from the campus deals that dominate the headlines. Agreements such as the University of California’s expanded OpenAI deal are about access to tools. The playbook starts where those deals end: what instructors do once students already have the tools.
What educators can reuse, and the license catch
Georgia Tech’s announcement only says the book is under a Creative Commons license. The PDF names the exact one: CC BY-NC-SA 4.0. Schools and universities may adapt the plays with attribution and must share their adaptations under the same terms, but commercial use is excluded, so edtech vendors cannot fold the material into paid products without separate permission. Readers do not need the full volume: the 20 lessons fill pages 265 to 278, and each article is also available as a separate PDF on the playbook site. Our AI in Education Policy Tracker tracks how institutions and governments set rules for the same questions.
What is still open
- The results are reported by the instructors who ran the courses, not by independent evaluators, and most come from a single course and semester.
- The 2027 edition is meant to focus on faculty who are only starting with AI, a group this edition covers less.
i10x Perspective
The most valuable part of the book is its admissions, not its successes. Teams write about tools they had to rebuild, rules that aged within a semester and assistants that were too helpful. That makes it more useful to a department head than most vendor guides, and the non-commercial license keeps it in the hands of educators. We expect lesson 12, building on what a discipline needs rather than on what AI cannot do yet, to become the line other universities quote when they rewrite their AI policies.



