[2407.12687] Towards Responsible Development of Generative AI for Education: An Evaluation-Driven Approach

1 2 minutes read

240712687 Towards Responsible Development of Generative AI for Education An.png

[Submitted on 21 May 2024 (v1), last revised 2 Dec 2025 (this version, v4)]

Authors:Irina Gorenka, Markus Konesh, Kevin R. Mackey, Danielle Gillick, Shaojian Zhu, Sarah Wiltberger, Shubham Milind Fall, Katherine Herman, Daniel Kasenberg, Avishkar Bhubchand, Ankit Anand, Mirona Beslar, Stephanie Chan, Lisa Wang, Jennifer Shi, Parsa Mahmoudia, Alia Riesbeck, Wei-Jin Kuo, Andrea Huber, Brett Wiltshire, Gal Illidan, Ronnie Rabin, Jasmine Rubinowitz, Amit Pitaro, MAC McAllister, Julia Wilkowski, David Choi, Roy Engelberg, Lidan Hackmon, Adva Levin, Rachel Griffin, Michael Sears, Philip Barr, Mia Mazar, Mana Jabbour, Arsalan Chowdhury, James Cohan, Sridhar Thiagarajan, Nir Levin, Ben Brown, Dylan Goror, Svetlana Grant, Rachel Hashemshoni, Laura Weidinger, Jiro Ho, Don Chen, Kuba Dolecki, Kanfer Akbulut, Maxwell Belshy, Laura Culp, Wen Shin Dong, Nahima Marshall, Kelsey Van Deman, Hema Bajaj Misra, Michael Duah, Moran Ambar, Avi Casciolaro, Sandra Livdahl, Chris Summerfield, James Ahn, Pierre-Alexandre Kamini, Abhinit Mehdi, Theophilus Strinopoulos, Annie Hill, Wayne Anderson, Louis C. Cobo, Nev Ephron, Muktha Ananda, Shakir Mohammed, Maureen Hemans, Zubin Ghahramani, Yossi Matias, Ben Gomez, Laila Ibrahim.

View PDF of the article Towards Responsible Development of Generative AI for Education: An Evaluation-Based Approach, by Irina Gorenka and 73 other authors

View PDF HTML (beta)

a summary:A major challenge facing the world is providing equitable and universal access to quality education. Recent advances in generative artificial intelligence (gen AI) have created excitement about the potential of new technologies to deliver a personal tutor for every learner and a teaching assistant for every teacher. But the full extent of this dream has not yet been realized. We argue that this is primarily due to difficulties in verbalizing pedagogical intuitions in AGI claims and a lack of good assessment practices, reinforced by challenges in identifying excellent teaching methods. Here we present our work in collaboration with learners and teachers to translate high-level principles of science learning into a practical set of seven diverse learning standards, including quantitative, qualitative, automated, and human assessments; And to develop a new set of fine-grained datasets to improve Gemini’s pedagogical capabilities, by introducing LearnLM-Tutor. Our ratings show that LearnLM-Tutor is consistently preferred over the fast-tuning Gemini by teachers and learners on a number of pedagogical dimensions. We hope that this work will serve as a first step toward developing a comprehensive educational evaluation framework, and that this will enable rapid progress within the AI and EdTech communities toward maximizing the positive impact of the AI generation in education.