Is AI transforming instructional design, or just speeding up content?

AI is transforming how instructional content gets produced, not yet how people learn. That distinction is the whole story of AI in instructional design right now, and most of the industry is still talking past it.

The field’s biggest gathering, ATD’s international conference, ran in Los Angeles from May 17 to 20. The interesting part was the reframe that followed it. dominKnow’s recap, titled “AI Got Quiet, Content Ops Got Loud” (June 1), reported that almost nobody was still asking “what AI do you have.” The new worry is governing the flood of AI-produced content. Their own research puts 38% of L&D leaders naming content management and governance as the biggest blocker to business value, against only 17% who say creating new content is the harder problem (dominKnow, vendor research). The same fortnight, Google shipped Gemini 3.5 Flash at its I/O conference. The tools got cheaper and faster to use at the exact moment the field admitted that producing more was no longer the point.

The honest inventory is production. In Synthesia’s 2026 L&D survey (a video-AI vendor, so weigh the numbers accordingly), the leading uses were voiceover at 63%, content and quiz drafting at 60%, and video at 52%, and 84% named speed as the main draw. The adoption itself is not in dispute; AI has topped Donald Taylor’s L&D Global Sentiment Survey two years running. And the 2026 launches all point one way: Docebo’s AgentHub now assembles a whole course, outline and assessments included, straight from your existing documents, and Workday’s relaunched Sana shipped its own learning agents in the same wave. Synthesia raised 200 million dollars in January to build “agents for learning.” The bet is plain. AI eats authoring.

Here is where the story turns. In that same Synthesia survey, 88% reported time saved on content creation, but only 41% said AI was contributing to business impact. That gap is the tell, and it sits on top of an older problem: by the field’s own accounting, only about half of L&D functions even have a way to measure learning impact (CIPD). We have run this experiment before. The MOOC era flooded the internet with free, professionally produced courses, and Reich and Ruiperez-Valiente, writing in Science in 2019, tracking 5.6 million learners across MIT and HarvardX, found completion fell to around 3% by 2017 to 2018. Unprecedented supply, and the outcomes never followed. Content volume was never the constraint. Call it the content-speed trap: mistaking faster production for better learning. AI makes it cheaper than ever to fall into. What drives durable learning is unglamorous and old news: people get better by doing the real thing, repeatedly. The evidence is settled, from retrieval practice (Roediger and Karpicke, 2006) to spacing study out over time (Cepeda and colleagues, 2006); in Dunlosky’s 2013 review of ten common study techniques, only practice testing and distributed practice earned the top “high utility” rating. Producing the content faster does not touch any of that.

The most important AI-and-learning result of the past two years is a warning. Bastani and colleagues, in PNAS in 2025, gave roughly 1,000 students access to AI during practice. Plain ChatGPT lifted their performance while they used it, then left them about 17% worse on a later test with no AI. A version built with guardrails, made to coach rather than hand over answers, brought those students back in line with classmates who never touched AI. Same model, opposite outcome, decided entirely by design. The encouraging mirror image: when An and colleagues (2025) used a model to generate retrieval-practice questions instead of more reading, retention went up. And Kestin’s 2025 Harvard study found a well-designed interactive AI tutor beat an active-learning classroom. The pattern is consistent. AI helps learning when it is built to create practice, and tends to hurt it when it just delivers content faster. It is the same line between making people more capable and making them dependent that shows up everywhere else AI touches work. As a Brookings review put it plainly this year, the claims about generative AI’s educational benefits have outpaced the high-quality evidence (Burns, 2026). For AI-authored static courseware specifically, the outcome evidence is simply not in yet.

Not extinction. The more careful voices, Philippa Hardman and the Learning Guild among them, argue the role is specialising rather than dying, and that is the better-supported read. But the displacement is real at the edges: Skillsoft cut its entire Codecademy curriculum team back in 2025. The way to stay on the right side of that line is to own the half of the job AI cannot do. AI now drafts. The designer’s value moves to deciding what is worth practising and engineering the reps that make it stick. If you hold the budget, that is where it should go too: fund less of what only makes content faster, and more of what turns content into practice. So two questions are worth asking in your own organisation: are you measuring production speed or learning outcomes, and is your AI generating more content or more practice? If the answer is “speed” and “content” on both, you have fallen for the content-speed trap. You are not closing the gap between what your people can do and what the work now demands; you are filling it faster with capability debt, and the interest still comes due.

• Reich & Ruiperez-Valiente, “The MOOC pivot,” Science (2019): https://doi.org/10.1126/science.aav7958
• Roediger & Karpicke, “Test-Enhanced Learning,” Psychological Science (2006): https://doi.org/10.1111/j.1467-9280.2006.01693.x
• Cepeda et al., distributed-practice meta-analysis, Psychological Bulletin (2006): https://doi.org/10.1037/0033-2909.132.3.354
• Dunlosky et al., “Improving Students’ Learning,” Psychological Science in the Public Interest (2013): https://doi.org/10.1177/1529100612453266
• Bastani et al., generative AI and learning, PNAS (2025): https://www.pnas.org/doi/10.1073/pnas.2422633122
• An et al., LLM-generated retrieval practice, arXiv (2025): https://arxiv.org/abs/2507.05629
• Kestin et al., AI tutoring vs active learning, Scientific Reports (2025): https://www.nature.com/articles/s41598-025-97652-6
• Burns, generative AI in tutoring, Brookings (2026): https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/
• Synthesia, AI in Learning & Development Report 2026: https://www.synthesia.io/reports/ai-in-learning-and-development-report-2026
• Taylor, L&D Global Sentiment Survey 2026: https://donaldhtaylor.co.uk/research_base/global-sentiment-survey-2026/

This is the first AI and Learning Field Report, a monthly read on what is actually happening in AI and learning, told straight and sourced. See the whole series.