EdTechnical

In this EdTechnical short, Owen is joined by guest co-host Alyssa Van Camp, research lead at EdTechnical, to discuss a recent paper on the learning penalty from generative AI. Drawing on survey data from around 27,000 secondary school students in China, the study looks at how using generative AI may affect homework productivity and learning.

The headline finding is that students who use AI score higher on their homework, but they spend less time on it and their exam results drop. The findings fit with research on how effort shapes learning, though the study design has some limitations. Students self-reported their AI use, time spent on homework was measured indirectly through the school's learning platform, and the paper gives little insight into the quality of the exams used to assess learning.

Surprisingly, around one in five students used AI without a drop in their exam scores. These students spent as much time on their homework as their peers, suggesting they used AI as a tool to help, rather than offload to. What does this mean for designing AI tools that support learning? And how might homework need to change?

Links

Research paper discussed in this episode: The Generative AI Learning Penalty: Evidence from Chinese Secondary Education

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this episode, Libby and Owen speak with Joanna Cannon, senior fellow at the Walton Family Foundation, where she supports a growing body of research on how teachers and students are using AI.

Measuring how teachers adopt AI is not straightforward. Joanna says the figures shift depending on what is asked and how AI is defined. The widely reported time savings of using AI come with significant caveats. And although AI can remove certain biases from grading, it introduces fresh risks of its own. When is reaching for AI appropriate, and when does it feel misplaced?

Bio
Joanna Cannon is a senior fellow with the Walton Family Foundation, where she works with the K-12 education team, including on a portfolio of research on how AI is being adopted by teachers and students. Joanna holds a PhD in developmental psychology from Columbia University and completed a postdoctoral fellowship at the University of Chicago. Previously, she was chief operating officer at the education non-profit PENCIL and chief strategic officer for the Division of Talent, Labor and Innovation at the New York City Department of Education.

Links
Walton Family Foundation, Education Program

Gallup and Walton Family Foundation research on teacher AI use and time savings

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this episode of EdTechnical, Owen reports back from the Festival of Learning in South Korea, the annual gathering that brings together academics in AI, education, learning sciences, and human-computer interaction. Owen takes the pulse of the field through four conversations at the Festival. He speaks with a Cornell professor, a funder at the Gates Foundation, an early-stage startup founder, and a researcher at Google DeepMind.

The community has stopped asking simply whether generative AI is good or bad for learning, and started asking where to use it, where not to, and how to build around it.

Guests

Rene Kizilcec is a Professor at the Department of Computing and Information Science at Cornell University and a general chair of the 2026 Festival of Learning. His work looks at how to use technology effectively and responsibly to support student learning.

Bryan Richardson works on the K12 education program at the Gates Foundation, where he focuses on helping AI models get better at education. He is part of a newly launched, pooled AI infrastructure fund backing open source tools and benchmarks for education.

Maya Bialik is founder of QuestionWell, an AI tool that helps teachers create and refine instructional materials. She is also a PhD student at Boston University, and author of Irreplaceable: How AI Changes Everything (and Nothing) in Teaching and Learning. 

Kevin McKee is a research scientist on the AI for Education team at Google DeepMind, where he leads assessments and evaluations of Gemini and other large language models in real world educational settings.

Links

The Festival of Learning 2026

Owen’s research on LLMs’ ability to interpret and grade handwritten student work 

Further reading from the conversations:

Catching The Correct Answer Trap by Moiz Imran and Sahan Bulathwela

From Untamed Black Box to Interpretable Pedagogical Orchestration by Nizam Kadir

Research by QuestionWell

Eedi and Google DeepMind math tutoring research in the UK

Gemini Guided Learning study with Fab AI in Sierra Leone

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this episode of EdTechnical, Libby and Owen speak with Rebecca Winthrop, senior fellow at the Brookings Institution, where she directs the Center for Universal Education. Rebecca and her team recently ran a ‘pre-mortem’ on generative AI in education: rather than waiting years for the evidence to settle, they conducted a yearlong study on the risks it poses to students, and what we can do now to prevent them while making the most of the benefits.

Drawing on focus groups, a thorough review of studies, and a forecasting panel, the report paints a striking picture: as things stand, the risks of generative AI for children's learning and development outweigh the benefits. Rebecca explains the risks are foundational, undermining the trust, independent thinking, and resilience that children need in order to benefit from AI. A key distinction runs through the conversation: risks tend to come from wide AI use, where children have open-ended, unscaffolded conversations with tools not designed for them, while benefits come from narrow, intentional use integrated into teaching with vetted content.

Bio

Rebecca Winthrop is a senior fellow at the Brookings Institution and director of its Center for Universal Education, where she leads the Brookings Global Task Force on AI in Education. She advises governments, policymakers and international institutions on education and previously led education work for displaced and conflict-affected communities. She is a co-author of the report A New Direction for Students in an AI World: Prosper, Prepare, Protect.

Links

Research Report by the Brookings Institution’s Center for Universal Education:
A New Direction for Students in an AI World: Prosper, Prepare, Protect

Brookings Institution Center for Universal Education: https://www.brookings.edu/topics/education-2/

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

Libby is on maternity leave, so Hermione Thompson steps in to guest co-host this episode alongside Owen. Hermione is AI and EdTech lead at Purposeful Ventures and previously advised on education policy inside the UK Prime Minister’s Office at 10 Downing Street.

Owen and Hermione sit down with Minister Olivia Bailey, the UK Department for Education minister responsible for AI in schools, for a conversation on where AI in education is headed. Bailey believes government should actively shape this space rather than sit back. Through the government's AI Tutoring Tools Pioneer Programme, up to eight companies will be selected to co-design curriculum-aligned AI tutoring tools directly with schools. Trials start this summer, and a national rollout is targeted for 2027, potentially reaching 450,000 disadvantaged pupils a year. 

Bailey says these coming tools are fundamentally different from a generic chatbot that gives children the answers without engagement or challenge. She explains how they are thinking about the tension between personalized AI and managing a classroom of 30 kids who need to study the same topic. The government's approach is to test and learn different approaches to see how children get the learning benefits normally associated with a real-life personal tutor.

Bio

Olivia Bailey is a junior minister at the UK Department for Education, responsible for early years, school food, and digital/AI policy. Before Parliament, she worked in public policy for the Labour Party, advised Keir Starmer as Leader of the Opposition, and worked across think tanks and public opinion research. She now leads the department's AI Tutoring Tools Pioneer Programme and its AI safety standards for schools.

Links

DfE announcement: Edtech and AI companies invited to help build safe AI tutoring tools https://www.gov.uk/government/news/edtech-and-ai-companies-invited-to-help-build-safe-ai-tutoring-tools-for-disadvantaged-pupils

⁠Programme detail via AI.gov.uk: AI Infrastructure for Education https://ai.gov.uk/our-work/education/ 

Schools Week coverage, including union reaction: DfE invites bids from AI tutoring pilot partners https://schoolsweek.co.uk/dfe-invites-bids-from-ai-tutoring-pilot-partners/ 

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

This season EdTech founder Libby Hills and AI researcher Owen Henkel continue to speak with leading researchers, practitioners and educators on the EdTechnical podcast series about the cutting edge of AI in education. They will break down complex AI concepts into non-technical insights to better understand what the research says and help educators sift the useful insights from the AI hype.

In the first episode of a new season of EdTechnical, Libby and Owen speak with Sam Gilbert, Professor of Cognitive Neuroscience at University College London. Sam is one of the leading researchers studying ‘cognitive offloading’, the use of tools and technologies to support thinking and memory.

Are tools like ChatGPT making us cognitively weaker?  Humans have been offloading  for thousands of years in ever-advancing ways: supporting memory through written notes, calculators, GPS, and large language models. There is a difference between losing specific information we've offloaded and losing underlying cognitive abilities. The distinction matters for thinking about AI's impact on learning and education. “We need to build up foundational abilities, and then offloading is very useful after we’ve acquired them”, Sam says.

Sam Gilbert

Sam Gilbert is a Professor of Cognitive Neuroscience at University College London (UCL), where he leads the Metacognition and Executive Functions Group. His research focuses on memory, cognitive control, metacognition, and cognitive offloading. He is widely recognised for his work on how people use external tools and technologies to support thinking and memory.

Links

Cognitive Offloading (Risko & Gilbert, 2016): Cognitive Offloading - PubMed 

UCL Metacognition & Executive Functions Group: Metacognition & Executive Functions | UCL Faculty of Brain Sciences

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this EdTechnical short, Libby and Owen unpack ‘AI agents’ and what they mean for education. Agents are large language models connected to tools and workflows that are allowed to take actions like searching, summarising, and completing multi-step tasks. Recent progress comes from the combination of stronger models and better systems for connecting agents to external tools, enabling more complex and autonomous outputs.

Applying agents to education brings a tension between flexibility and reliability. Agentic systems can be useful for teachers, who operate across varied contexts and need adaptable support. For students, especially in structured learning, too much flexibility can reduce clarity and introduce inconsistency.

This matters because effective learning depends on structure and progression. The value of agents in education depends on how well they are applied to the specific task and learning goal.

Links:

• OpenClaw & Moltbook: The viral story of AI agents building their own Reddit-like social network https://techcrunch.com/2026/01/30/openclaws-ai-assistants-are-now-building-their-own-social-network/ 
• Claude Research Mode: Anthropic's explainer on deep research https://www.anthropic.com/news/research 

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

At this year's SXSW EDU, Owen joined a panel on what it takes to make voice AI for assessment work in classrooms.

In this live recording of the session, the panelists untangle how voice AI works, and what testing this technology with kindergartners looks like in rural Georgia. They explain why the distinction between capturing what a student said versus what they meant matters enormously for literacy assessment and why questions of privacy, equity and model bias are not afterthoughts but design requirements.

Where does voice AI genuinely open up new possibilities in education, and where is the evidence still thin? 

The other panelists were Patti Ura, Director of Learning Technology Research at Digital Promise, Amelia Kelly, VP of Data Science at Curriculum Associates and former CTO of Soapbox Labs, and Kristen Hoff, Head of Measurement at Curriculum Associates.

Links:

• Soapbox Labs, now part of Curriculum Associates 
• Digital Promise 
• OpenAI Whisper, the open source speech-to-text model

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this EdTechnical short, Libby and Owen test a conversational plush toy to understand more about AI-powered toys designed for young children. Recent research from Cambridge shows that preschool-aged children can form rapid emotional connections with social robots like these, even when the responses from the robot are inconsistent.

Children’s experiences with AI toys are shaped by voice and real-time interaction. Could highly responsive, frictionless AI systems in toys influence children’s expectations of human relationships?

Libby and Owen discuss the difference between shared, supervised play and extended solo interaction with the toy, which may be less advisable. As the technology continues to improve, the key challenge becomes how these tools are introduced and used in early childhood environments.

Links:

• BBC Article: AI toys for children misread emotions and respond inappropriately, researchers warn
• Cambridge study on AI toys in early childhood
• AI chatbots and the “empathy gap” in children

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production. 

In this episode of EdTechnical, Libby and Owen speak with Panos Ipeirotis, Professor at NYU Stern School of Business, about his experiment using AI to run oral exams in university courses. As generative AI makes it easier for students to outsource written assignments, educators are asking whether traditional take-home assessments still measure real understanding.

Panos introduced AI-mediated oral assessments after noticing a mismatch between high-quality written submissions and weak classroom discussion. In the new system, students answer questions from a voice agent that probes their understanding of the material and their own work.

Panos tells Libby and Owen how the exams work, including an AI “council” of language models that evaluates transcripts and produces detailed feedback. What does this approach reveal about the future of assessment? Could AI make oral exams scalable in higher education, and even improve fairness and grading consistency?

Links:

• Panos Ipeirotis – NYU Stern Faculty Profile
• NYU Professor Uses AI-Run Oral Exams to “Fight Fire with Fire”
• Article: The case for oral assessment in the age of AI

Guest Bio

Panos Ipeirotis is a Professor of Information, Operations and Management Sciences at NYU Stern School of Business. His research focuses on data science, AI, and human-AI collaboration. In addition to his academic work, he experiments with practical applications of AI in education, including new models of assessment that combine oral exams with AI-based evaluation.

Join us on social media: 

• BOLD (@BOLD_insights), Libby Hills (@Libbylhhills) and Owen Henkel (@owen_henkel)
• Listen to all episodes of EdTechnical here: https://boldscience.org/podcasts/edtechnical/
• Subscribe to BOLD’s newsletter: https://bold.expert/newsletter
• Stay up to date with all the latest research on child development and learning: https://bold.expert

Credits: Sarah Myles for production support; Josie Hills for graphic design; Pia Schmoeckel for content production.