Fable 5's Cost Gap, AI "Cheating" at Scale, and a 70-Page Handbook in 20 Minutes

Fable 5's Cost Gap, AI "Cheating" at Scale, and a 70-Page Handbook in 20 Minutes

WHEN BETTER MODELS WIDEN THE GAP: AI'S COST DIVIDE IN HIGHER ED (AI GOES TO COLLEGE, EP. 36) What happens to students when the best AI models cost ten times more than the basic ones? That is the question Craig and Rob keep circling in this episode, prompted by Anthropic's brief and strange release of Fable 5. Fable 5 arrived as a guardrailed version of Mythos, a model so good at exposing software vulnerabilities that Anthropic had restricted it to a small set of secure organizations. For about a week it was freely available to paid users; then federal import controls landed and Anthropic pulled it, with no clear word on when, or whether, it returns. The hosts use that whiplash to get at the questions that actually matter for higher ed: who can afford the most capable tools, what that does to learning, and why none of it changes the deeper problem with how we assess students. They also dig into a large new study on student AI use, the agents Rob is building for faculty this summer, and a 70-page course handbook Craig generated in an afternoon. WHAT YOU'LL HEAR The cost gap, in real numbers. Craig walks through Anthropic's tiers (Haiku, Sonnet, Opus, Fable) and what they cost to run: a task that runs free under his Opus subscription would have cost roughly $50 in Fable 5, while Haiku sits around $5. His worry is that this turns into an SAT-prep dynamic on steroids, where score gaps come from resource access rather than ability. Rob's counterintuitive flip. Rob raises the possibility that students stuck on weaker models might actually learn more, because they cannot offload as much of the cognitive work and have to stay involved in it. Neither host claims to know; they treat it as a real open question. A large study on student AI use. The hosts dig into a Science paper covering more than 95,000 students across 20 major U.S. public research universities. About two-thirds reported using generative AI in the prior year; roughly 9% of those users said they turned in AI-generated work knowing it wasn't allowed. The inappropriate-use rates run higher in non-STEM fields even though adoption there is lower. Faculty tools built over the summer. Rob describes agents his student interns are building: a syllabus-comparison tool that flags where a faculty member's syllabus diverges from the new template, an active-learning brainstorming assistant, and an AI-resilience checker for assignments and assessments. A textbook-grade handbook in an afternoon. Craig recounts handing OpenAI's Codex a couple of syllabi and one-shotting a 70-page course handbook for a freshman business course, then refining the activities. He pledges to release the finished version under a Creative Commons license. WHY THE GAP IS THE REAL STORY The Fable 5 saga is good copy, but the hosts keep pulling it back toward something more durable. When the most capable models cost an order of magnitude more than the entry-level ones, the divide isn't only between rich and poor institutions; it reaches into a single classroom, where one student on a free model and another paying for the frontier model are turning in work that no longer means the same thing. Craig's answer isn't to chase the frontier. It's to teach students to match the model to the task; you don't pay for the expensive employee to do routine work, and you don't burn Fable 5 on something Haiku can handle. Rob extends the point to policy: banning AI outright is folly, both because it's nearly impossible to detect without introducing bias and because it leaves you with a classroom where you have no idea who learned what. Craig demonstrates the detection problem directly, running lightly edited AI text through Pangram and getting a "100% human" verdict. The shared conclusion is one they've made before and make again here: the urgent work is assessment reform, because a graded artifact is no longer a trustworthy signal of what a student actually knows. EPISODE HIGHLIGHTS * (12:11) Rob on weaker models and learning: "I wonder if the people who aren't using these highly capable models might actually learn more, because they're not going to be able to cognitively offload as much of the things that they're doing, and they'll need to be more involved in it." * (15:51) Craig on Anthropic's rollout: "Anthropic really came off like a drug dealer that gives you a little taste before they try to get you hooked." * (20:01) Craig on the study's central finding: "About 9% of those users turned in AI-generated work knowing it wasn't allowed." * (26:31) Craig on why assessment has to change: "It's no longer a trustworthy signal of what they know if we keep doing things the way we've been doing them." * (37:21) Craig on the AI-built handbook: "It was a 70-page handbook with learning activities; good, not great, in about 20 minutes." * (45:31) Craig's top tip: "Handoff documents and memos will make your life so much easier when it comes to AI." REFERENCES MENTIONED * Anthropic's Fable 5 and Mythos models. Discussed as a guardrailed public release (Fable 5) built on the restricted Mythos family, later pulled following federal import controls. (Link to Anthropic announcement / status page: see "What I need from you.") * Study on student generative-AI use in Science (data collected 2024). (Full citation and DOI to be provided: see "What I need from you.") * Pangram: AI-text detector Craig used to test a lightly edited, skill-generated draft. (Link: see "What I need from you.") * OpenAI Codex: GPT-based coding agent Craig used to generate the course handbook. * Skills in Claude / Cowork and Microsoft Copilot: discussed as reusable, callable tools for tasks like generating active-learning activities. * Van Slyke & Crossler textbook (with co-author Franz Boulanger): referenced as a point of comparison for AI-assisted content creation. (Title / link if you want to include it: see "What I need from you.") QUESTIONS TO CONSIDER * If two students submit work produced with very different models, are you assessing their learning or their access? * Where in your own courses is a graded artifact still a trustworthy signal of what a student knows? * What's one tedious faculty task this summer that an agent could take off your plate? AI Goes to College is a podcast for higher education professionals trying to make sense of artificial intelligence in their classrooms, their research, and their institutions. Co-hosted by Craig Van Slyke and Rob Crossler, the show focuses on practical, evidence-based perspectives on AI in higher education without the hype. Subscribe and follow: https://www.aigoestocollege.com/follow [https://www.aigoestocollege.com/follow] · Newsletter: https://aigoestocollege.substack.com/ [https://aigoestocollege.substack.com/] LINKS Science article Chirikov, I., Smirnov, I., & Kizilcec, R. F. (2026). Generative AI use and misuse call for assessment reform in higher education. Science, 392(6800), 818-820. https://www.science.org/doi/10.1126/science.aec5115 [https://www.science.org/doi/10.1126/science.aec5115] Anthropic Fable 5/Mythos announcement https://www.anthropic.com/news/fable-mythos-access Pangram (AI "detector") https://www.pangram.com/ [https://www.pangram.com/] Information Systems for Business: An Experiential Approach (Belanger, Van Slyke & Crossler) https://www.prospectpressvt.com/textbooks/b%C3%A9langer-information-systems-for-business-an-experiential-approach-5-0 [https://www.prospectpressvt.com/textbooks/b%C3%A9langer-information-systems-for-business-an-experiential-approach-5-0] Mentioned in this episode: AI Goes to College Newsletter

AI's Underused Capabilities and Hidden Risks

EPISODE 35: AI'S UNDERUSED CAPABILITIES AND HIDDEN RISKS What happens when a university scrapes faculty lectures from its LMS, feeds them into an AI course builder, and sells the result for five dollars a month without telling the professors whose faces appear in the videos? Craig and Rob cover a packed news cycle in this episode, anchored by two stories about institutional vulnerability. The Canvas ransomware attack that disrupted final exams at thousands of schools opens a conversation about single points of failure; ASU Atomic, Arizona State University's new AI-powered course builder, raises harder questions about who controls faculty content and what happens when AI strips the context out of teaching. The episode also features Craig's deep dive into what coding agents like Codex and Claude Code can actually do for faculty (spoiler: it goes well beyond writing code), and a cautionary tale about Gemini failing spectacularly on a home networking problem. WHAT YOU'LL HEAR The Canvas ransomware attack and what it reveals about AI dependency. The attack took down learning management systems at roughly 8,800 institutions during final exam season. Rob connects this to the broader security landscape for AI tools, arguing that the same single-point-of-failure problem applies to the AI agents and workflows faculty are starting to build. Craig's own Claude outage, which wiped out one of his custom skills mid-edit, underscores the point. ASU Atomic and the faculty backlash nobody saw coming. ASU's new platform uses an AI system called Atom to pull faculty lectures, assignments, and slide decks from Canvas, chop them into short clips, and reassemble them into personalized learning modules. Faculty weren't consulted. Rob immediately draws a parallel to NCAA name, image, and likeness rights. Craig argues the program will push faculty to pull their materials off the LMS entirely, hurting the most vulnerable students who depend on recorded lectures and posted materials. A practical showcase of coding agents for non-coders. Craig walks through a series of tasks he completed using Codex and Claude Code: de-identifying and structuring messy focus group transcripts, running text analysis algorithms, auditing and reorganizing doctoral seminar materials, and renaming over 130 PDFs with no coherent naming scheme. None of it required writing a single line of code. Rob pushes back on trust and sandboxing, and the two discuss the "middle ground" between AI slop and untouched human work. When AI hits a wall. Craig recounts an hour-and-a-half failure trying to use Gemini to troubleshoot a mesh network failover setup. The AI kept providing outdated instructions because the ISP had changed default settings without documenting the changes. The fix required a human tech support agent who could reset the modem remotely. The lesson: AI tools are great until they encounter the kind of hidden institutional knowledge that every organization has. THE CHILLING EFFECT ON ACCESSIBILITY The ASU Atomic discussion surfaces a consequence that hasn't gotten enough attention in the broader coverage. Craig argues that the predictable faculty response to programs like Atomic is to minimize what they post to the LMS. No more recorded lectures, fewer slide decks, assignments handed out in person rather than uploaded. This is a rational defensive move for faculty, but it disproportionately harms students who depend on those digital materials: working students, parents, students with disabilities. The lifelong learning mission that ASU Atomic claims to serve gets undermined by the very mechanism used to pursue it. Rob extends this to the tension between financial incentives and student interests at land-grant institutions, noting that the populations these universities were built to serve may not be well-served by this model. EPISODE HIGHLIGHTS * (09:42) Craig on ASU Atomic: "They started up ASU Atomic, which uses something called ASU Atom, which is an AI course builder that goes out into the learning management system, pulls content from all these different courses, and repackages them into something that is going to be a $5 a month consumer-facing web app." * (11:22) Rob on the NIL parallel: "I can totally see where faculty feel that they own their name, image, likeness, right? Much like our athletes deal with." * (13:22) Craig on the chilling effect: "If you're worried about this, okay, I'm just not gonna have my lectures recorded. I'm gonna minimize what I put on the LMS... that's gonna have a detrimental effect on the most vulnerable students." * (17:03) Craig on deepfakes and harassment: "You throw that in with deepfakes and forget about harassment. You could have considerable misinformation and disinformation campaigns built around legitimate faculty members." * (30:22) Craig on the middle ground for AI in research: "There's this huge middle ground that we're gonna have to figure out where we're using AI to let us do better research and produce knowledge more effectively and more efficiently. But it's not AI slop. It's still something that was done with human oversight, kind of like we've been doing with GAs for a long time." * (40:22) Craig on AI limitations: "These AI tools are great until they're not." REFERENCES MENTIONED * ASU Atomic (also called ASU Atom, internal codename "Project Atomizer"): Arizona State University's AI-powered course builder, launched as a pilot in April 2026 * Canvas ransomware attack (May 2026): attack on Instructure's Canvas LMS affecting approximately 8,800 institutions during final exam season * OpenAI Codex: OpenAI's autonomous coding agent * Claude Code: Anthropic's coding agent (Craig's primary tool for the tasks described in the episode) * Google Antigravity: Google's coding agent (mentioned but not tested for the tasks Craig describes) * Gemini: Google's AI assistant (used in the networking troubleshooting story) * NCAA Name, Image, and Likeness (NIL) rights: invoked by Rob as a parallel to faculty IP concerns * Arizona Board of Regents intellectual property policy: the work-for-hire framework under which ASU claims ownership of faculty-created course materials * Eero 7 mesh network devices (Amazon): the hardware in Craig's networking troubleshooting story AI Goes to College is a podcast for higher education professionals trying to make sense of artificial intelligence in their classrooms, their research, and their institutions. Co-hosted by Craig Van Slyke and Rob Crossler, the show focuses on practical, evidence-based perspectives on AI in higher education without the hype. Subscribe and listen: [link to platforms] | Read more: [link to AIGTC Substack] Takeaways: * The reliance on learning management systems like Canvas exposes institutions to vulnerabilities during outages, especially during critical academic periods. * AI tools, while enhancing productivity, present significant risks if not managed with proper oversight and backup strategies. * The recent developments at Arizona State University demonstrate a growing trend of institutions utilizing AI in ways that may undermine faculty autonomy and intellectual property rights. * The integration of AI into educational settings necessitates a shift in teaching methodologies towards experiential learning and greater student engagement. * Backup protocols are essential when utilizing AI tools to prevent loss of critical data and ensure continuity of work in educational environments. * The evolving landscape of AI requires educators to actively engage with these technologies to better understand their implications and guide student use effectively. Mentioned in this episode: AI Goes to College Newsletter

When You Bring AI to the Party Matters More Than Whether You Bring It

When in your thinking process should AI show up? A new study suggests the timing matters more than the access. In this episode, Craig and Rob work through a recent CHI (Computer Human Interaction) conference paper that found a counterintuitive pattern: participants who had AI access from the start of a 30-minute task wrote weaker reports than those who got AI late or had no access at all. Same tool, same task, opposite result. The hosts connect the finding to Herbert Simon's satisficing concept and ask what it means for how faculty should teach AI use in their classrooms. The conversation also covers entry-level hiring trends in tech (which look better than the headlines suggest), Microsoft Office Agent's strange refusal to generate slides on a textbook chapter about AI, and why Rob worries the floor in higher education is rising while the ceiling may be coming down. WHAT YOU'LL HEAR A four-tool slide deck experiment. Craig made the same presentation in Microsoft Office Agent, Claude Cowork, ChatGPT, and Gemini. The differences in output quality, refusal behavior, and editability are larger than the marketing suggests. The CHI satisficing study. Researchers from Chicago and Toronto ran almost 400 participants through a civic decision-making task. With ten minutes, early-AI access helped. With thirty minutes, it hurt. The hosts unpack why and what it means for any knowledge work that requires actual thinking. Why "good enough" is now a problem. When AI can produce a serviceable draft in seconds, the differentiator shifts to what happens after the first draft. Craig and Rob discuss why the floor is rising for entry-level work and why the ceiling may not be rising with it. Entry-level hiring data. Recent IEEE Spectrum reporting suggests entry-level tech roles are growing in some categories, contradicting the prevailing narrative. The hosts walk through which roles and what the trend means for university programs preparing students for those jobs. AI sycophancy in the wild. Rob shares why the tools' tendency to agree with the user's framing is more dangerous in high-stakes situations than in low-stakes ones, and what that means for how we should be using them. WHY TIMING MATTERS MORE THAN ACCESS The dominant question in higher education has been whether students should use AI. The CHI study suggests that's the wrong fight. The better question is when AI should appear in a student's thinking process. Participants with late AI access in the study produced the same number of arguments as those without AI, but with more balanced pro-and-con reasoning. The tool became a counterweight to their own thinking rather than a substitute for it. That's a different mental model than the one most faculty (and most knowledge workers) default to, and it has practical implications for course design, assignment structure, and how we coach students to work with these tools. EPISODE HIGHLIGHTS * (approx. 5:41) Craig on the four-tool slide experiment: "A lot of times with AI, 50% is better than 100%, because you can get the 50% really quickly." * (approx. 17:10) Craig on what entry-level workers need now: "The solution to helping our students get jobs is to show them how to lean into their humanity." * (approx. 19:55) Rob on the floor-and-ceiling tension: "The floor is going up, but the fear is that our ceiling is coming down." * (approx. 31:15) Craig on the satisficing finding: "It's not literally where you stopped — it's where your engagement stopped." * (approx. 37:15) Rob's end-of-semester challenge to faculty: "Pick one thing. One thing that you're going to engage with over the summer." LINKS AND REFERENCES * Computer Human Interaction (CHI) conference paper on AI access timing and decision quality (researchers from Chicago and Toronto): https://dl.acm.org/doi/pdf/10.1145/3772318.3791796 * IEEE Spectrum reporting on entry-level technology hiring trends: https://spectrum.ieee.org/ai-effect-entry-level-jobs [https://spectrum.ieee.org/ai-effect-entry-level-jobs] * Herbert Simon's concept of satisficing (1956) * Microsoft Office Agent, Claude Cowork (Design feature), ChatGPT, Gemini FOR FACULTY: QUESTIONS WORTH SITTING WITH * Where in your course design does AI currently show up, and would your students be better off if it appeared later in their process? * How would you redesign one assignment so that students engage with the problem cold before the AI shows up? * What does excellence look like in your discipline now that "good enough" is trivially achievable? How will you recognize it, and how will you teach students to reach for it? ABOUT THE SHOW AI Goes to College is a podcast for higher education professionals trying to make sense of artificial intelligence in their classrooms, their research, and their institutions. Co-hosted by Craig Van Slyke and Rob Crossler, the show focuses on practical, evidence-based perspectives on AI in higher education without the hype. Subscribe and listen: https://www.aigoestocollege.com/ [https://www.aigoestocollege.com/] Newsletter: https://aigoestocollege.substack.com/ [https://aigoestocollege.substack.com/] Mentioned in this episode: AI Goes to College Newsletter

Accessibility Hacks, 81,000 Interviews, and the Choppy Waters of Academic AI

AI GOES TO COLLEGE, EPISODE 33: ACCESSIBILITY HACKS, 81,000 INTERVIEWS, AND THE CHOPPY WATERS OF ACADEMIC AI Higher education is drowning in accessibility deadlines, grappling with what 81,000 AI interviews reveal about how people actually use these tools, and watching the academic publishing system creak under new pressures. In this episode, Craig and Rob dig into all three, with practical advice, a few uncomfortable truths, and their usual mix of optimism and healthy skepticism. THE ACCESSIBILITY CRUNCH IS HERE (AND AI CAN HELP) The episode opens with a problem that's top of mind for faculty everywhere: the April 24 federal deadline requiring public-facing digital content to meet WCAG accessibility guidelines. Universities have been scrambling, and many of the contracted tools designed to help have been, as Craig diplomatically puts it, hit and miss. Craig shares a concrete example from his own workflow. He took three image-heavy slide decks from his Principles of Information Systems course and handed them to Claude Cowork with a simple instruction: add alt text for all the images. Within about 30 minutes, the job was done. The accuracy? Roughly 75 to 80 percent. A handful of images needed corrections, but instead of writing alt text for 40 or 50 images from scratch, he only had to fix six or eight. Rob tried something similar with Microsoft Copilot on a keynote presentation he gave at the SAIS conference in Asheville; two images, 30 seconds, done. Rob makes the important point that accessibility isn't just a PowerPoint problem. It extends to whiteboard files, videos, and essentially everything faculty communicate digitally. The burden is real, and it lands on faculty who are already overwhelmed by the changes AI is bringing to their professional lives. Craig adds a note of personal sensitivity here; his wife has a profound hearing disability, which makes these issues more than abstract compliance for him. The larger takeaway? When you hit one of these friction points in your work, try AI. It won't always solve the problem, but it often will, and the time savings can be substantial. WHAT 81,000 INTERVIEWS TELL US ABOUT HOW PEOPLE ACTUALLY USE AI Link: https://www.anthropic.com/features/81k-interviews [https://www.anthropic.com/features/81k-interviews] Craig's article: https://open.substack.com/pub/aigoestocollege/p/what-81000-people-told-anthropic [https://open.substack.com/pub/aigoestocollege/p/what-81000-people-told-anthropic] The conversation shifts to Anthropic's large-scale qualitative study, where Claude was used to conduct and analyze 81,000 interviews about how people use AI tools. Rob, who has spent considerable time doing qualitative research the traditional way (36 interview transcripts with families, a labor-intensive process), finds the scale almost hard to believe. Craig wrote a separate article about this study for the AI Goes to College newsletter. The phrase that catches both hosts' attention is one from the report: "the light and the shade are tangled together." It captures the tension between excitement about AI's possibilities and anxiety about what those possibilities mean for how people work, learn, and think. Craig connects this to a concept from technology studies: this is not technological determinism. The outcomes aren't dictated by the tools themselves. They emerge from the sociotechnical space where human choices and technological capabilities intersect. Rob observes that most current AI use cases still amount to doing what we've always done, just faster. The real transformation will come when people start imagining entirely new approaches (he draws an analogy to cloud computing, which started as a backup solution and eventually reshaped how people interact with technology in ways nobody initially anticipated). One quote from the Anthropic study lands hard. A freelance software engineer in Pakistan says: "I want to learn skills, but learning deeply is of no use. Ultimately I can just use AI." Craig points out that if a working professional thinks this way, the implications for students who may not yet appreciate the long-term value of deep learning are sobering. Rob agrees but pushes back slightly: people who lean too far into this mindset will eventually hit a wall where they lack the critical thinking skills to know when or why AI has gotten something wrong. The hosts converge on what's becoming a running theme for the podcast: higher education's central task is helping students understand the long-term value of cognitive engagement, because without that understanding, the default will always be to let AI handle it. ACADEMICS NEED TO WAKE UP: 10 THESES ON A SHIFTING LANDSCAPE Link: https://substack.com/home/post/p-189705626 The second major discussion centers on Alexander Kustoff's Substack article, "Academics Need to Wake Up on AI: 10 Theses for Folks Who Haven't Noticed the Ground Shifting Under Their Feet." Rob sees it as a useful prompt for conversations the research community needs to have. Craig appreciates the ambition but pushes back on some of the claims. Take thesis number one: AI can already do social science research better than most professors. Craig's reaction is nuanced. The claim is probably technically true if "most" is read literally, since many professors don't publish much (Rob notes the median number of publications for business school professors may be as low as one). But the implication that AI can replace skilled researchers? Not yet. Craig estimates that a knowledgeable researcher can use AI to cut research production time by about three-quarters, but that knowledge is the key ingredient; without research skill, you'll just produce publishable garbage faster. Rob raises something interesting: colleagues who are brilliant thinkers but never thrived in research because they didn't enjoy writing may now have a path to contribute. AI could genuinely democratize parts of the research process. Craig extends this point to data analysis; tools like Cowork can run Python and R analyses without expensive specialized software, which matters enormously for under-resourced institutions and researchers in developing countries. The conversation turns to the strain AI is putting on the peer review system. More submissions (many of them better written thanks to AI) are flooding journals, but finding reviewers was already difficult. Craig, speaking from his role as a journal editor, argues that well-trained AI could do a better job reviewing than roughly half of current human reviewers. Rob agrees but emphasizes that journal leaders need to come together and define norms for what's acceptable. Right now, the rules are either nonexistent or unrealistically restrictive ("just don't use AI for anything"), which creates the same kind of confusion faculty have imposed on students with inconsistent classroom policies. One of the most provocative moments comes when Craig reads a quote from the Kustoff article: "I don't envision a research assistant role in my workflow anymore. What I want from collaborators is original thinking, domain expertise, and intellectual challenge. This is a genuine loss for the traditional apprenticeship model, and I don't have a clean answer for how to replace it." Both hosts take this seriously. Craig argues that senior scholars will need to accept some suboptimal results in the short term to continue mentoring the next generation. Rob suggests the apprenticeship model isn't dying; it's transforming. The mentorship shifts from teaching students how to do tasks to teaching them how to direct AI tools and critically evaluate what those tools produce. Craig closes with a characteristically honest observation: senior scholars get stuck in their ways of thinking, and one of the real values of working with early-career doctoral students is the occasional moment when their unformed, messy thinking reveals a perspective that nobody in the room had considered. That's worth protecting. AI-GENERATED LESSON PLANS AND THE BLOOM'S TAXONOMY PROBLEM Link: https://citejournal.org/volume-25/issue-3-25/social-studies/civic-education-in-the-age-of-ai-should-we-trust-ai-generated-lesson-plans/ [https://citejournal.org/volume-25/issue-3-25/social-studies/civic-education-in-the-age-of-ai-should-we-trust-ai-generated-lesson-plans/] The final segment covers a paper by four researchers from UMass Amherst, "Civic Education in the Age of AI: Should We Trust AI-Generated Lesson Plans?" The study found that roughly 90 percent of AI-generated lesson plans hit only the lower levels of Bloom's taxonomy (remembering, understanding) rather than the higher-order thinking skills like analyzing, evaluating, and creating. Craig's first reaction was that the prompts used in the study were terrible. But he acknowledges the researchers had a reason: they were mimicking how most teachers would actually prompt. And that's the real finding. The problem isn't that AI can't produce sophisticated lesson plans; the problem is that untrained users produce unsophisticated prompts, and the output reflects the input. Rob agrees and broadens the point: if even a fraction of teachers are prompting this way, that's affecting a lot of students. Craig shares a personal anecdote from his one year as a high school teacher. He diligently wrote lesson plans; a veteran teacher (whom he describes as one of the best he'd ever seen) simply copied his plans to satisfy an administrative checkbox. The experienced teacher didn't need detailed plans because she could read the room and adapt in real time. Some lesson planning, Craig suggests, falls into a compliance category where the quality of the plan matters less than the quality of the teaching. But the bigger message is one both hosts keep returning to: we have to teach people how to use these tools well. Craig suspects that even a slightly more complex prompt ("address this level of Bloom's taxonomy and make sure you include demographic diversity") would produce dramatically better lesson plans. Rob makes a final observation that resonates beyond lesson planning. People who spend a lot of time thinking about AI (like Rob and Craig) can easily forget that most people don't. Understanding what AI use looks like for someone without deep expertise, and then helping to lift them up, is the real work ahead. Craig's response? Maybe the strategy should be seeding the field with AI evangelists, a small number of engaged opinion leaders who help others one conversation at a time, rather than trying to train everyone through top-down institutional programs. That's how innovations actually spread. A META-MOMENT: WHO WROTE THIS, REALLY? In a brief but revealing aside, Craig mentions that his Substack article about the Anthropic study was entirely generated and posted by an agentic AI workflow using Claude Code and Opus 4.6, built on his custom "write like Craig" skill. He asks Rob to guess the accuracy. Rob says 75 percent. Craig confirms. The question lingers: if AI can write in your voice with 75 percent accuracy and post it autonomously, who's really the author? Craig leaves that for the listener to decide. KEY TAKEAWAYS AI is a practical solution for the accessibility crunch. With the April 24 WCAG deadline looming, tools like Claude Cowork and Microsoft Copilot can generate alt text for images at roughly 75 to 80 percent accuracy, dramatically reducing the manual burden on faculty. "The light and the shade are tangled together." Anthropic's 81,000-interview study reinforces that AI's benefits and risks aren't separable. Higher education's job is to help students navigate both, not pretend one side doesn't exist. AI adoption follows a predictable pattern. First we use new technology to do old things faster. The real transformation comes when we start imagining fundamentally new approaches. Higher ed is still mostly in phase one. The prompt is the bottleneck, not the tool. AI-generated lesson plans that hit only lower-order Bloom's taxonomy levels aren't evidence that AI can't do better. They're evidence that untrained users produce unsophisticated prompts. Academic publishing is under real strain. More submissions, better surface-level writing, reviewer shortages, and undefined norms for AI use are all converging. Journal leaders need to establish clear, workable standards. The apprenticeship model is transforming, not dying. Mentoring doctoral students shifts from teaching them to do tasks toward teaching them to direct AI tools and critically evaluate the output. Senior scholars need to stay open to messy, unexpected thinking from early-career researchers. Seed the field with opinion leaders. Rather than top-down institutional training programs, Craig argues for cultivating AI evangelists who spread knowledge one conversation at a time; that's how innovations actually diffuse. Links Anthropic's 81,000 interviews: https://www.anthropic.com/features/81k-interviews [https://www.anthropic.com/features/81k-interviews] Craig's article: https://open.substack.com/pub/aigoestocollege/p/what-81000-people-told-anthropic [https://open.substack.com/pub/aigoestocollege/p/what-81000-people-told-anthropic] Academics need to wake up on AI: https://substack.com/home/post/p-189705626 AI generated lesson plans: https://citejournal.org/volume-25/issue-3-25/social-studies/civic-education-in-the-age-of-ai-should-we-trust-ai-generated-lesson-plans/ [https://citejournal.org/volume-25/issue-3-25/social-studies/civic-education-in-the-age-of-ai-should-we-trust-ai-generated-lesson-plans/] Companies/Products mentioned in this episode: 1. Claude Cowork 2. Microsoft Copilot 3. Anthropic 4. University of Central Oklahoma 5. UMass Amherst Mentioned in this episode: AI Goes to College Newsletter

We're On Our Own: Academic Integrity through AI Resilience

Craig and Rob kick off this episode with a deep dive into Claude's Constitution — the 84-page document Anthropic released to explain how Claude is governed. The document lays out a four-part hierarchy of priorities: be broadly safe, be broadly ethical, follow Anthropic's guidelines, and be genuinely helpful — in that order. Craig walks through the key language, and both hosts zero in on the uncomfortable questions it raises. Who gets to define "broadly ethical"? Whose values count? Craig points out that collectivist and individualist cultures would answer those questions very differently, and Rob raises the example of how privacy has historically carried different social weight in China versus the United States. They give Anthropic credit for the transparency. Rob notes that he has no idea what governs ChatGPT by comparison, and Craig argues the openness could become a real differentiator for universities evaluating which AI tools to bring in-house. But the Constitution also includes some curious language — the phrase "during the current phase of development" gives Anthropic significant room to evolve these guardrails over time, and a section on emotional support states that Claude should "show that it cares," which both hosts flag as a strikingly anthropomorphic choice of words. Craig shares a fun aside: he used Claude Code to build a clone of the classic Colossal Cave Adventure game — reframed around understanding large language models — using just a few sentences as a prompt. The game was up and running in about an hour. That kind of capability would have been unthinkable a couple of years ago, and it underscores why the Constitution's language about the "current phase" matters so much. The big takeaway from the Constitution discussion lands hard: higher ed is on its own when it comes to academic integrity. Anthropic — arguably the most transparent of the major AI companies — has no interest in blocking students from misusing its tools. Rob mentions a new product called Einstein that will watch your Canvas videos, write your discussion posts, reply to classmates, and complete your assignments. All you have to do is hand over your login credentials. That sets up the episode's second major topic: AI resilience. Rob explains the concept as designing learning outcomes that hold up regardless of what AI can do. If a major portion of a student's grade depends on writing an essay that AI could produce in seconds, that assignment has very little resilience. The shift Rob advocates moves evaluation toward process — asking students for the prompts they used, reflections on how they refined their approach, and demonstrations that they understand what was produced. He shares the example of a colleague whose programming class now requires students to record videos explaining their code rather than just submitting it. Craig raises the scaling problem. He regularly teaches 90 to 100 undergraduates. Rob suggests that AI itself can help with formative feedback on scaffolding assignments, freeing faculty to focus their grading energy on fewer, higher-stakes assessments. Craig uses an analogy from music: scaffolding assignments are like playing scales — you do them to build toward performance, and they don't need to carry grade weight. Both hosts agree this represents a move away from the grade economy, where students rationally minimize effort because every small assignment is a transaction. Craig pushes the conversation further by proposing live client projects — or AI-simulated client projects — as a way to create the messiness and ambiguity that real work demands. Rob's initial reaction is skepticism (live client projects are logistically brutal), but he warms to the idea of using AI to simulate clients with realistic fuzziness and scope creep. The broader point: AI could be the lever higher ed needs to fix problems that have been accumulating for decades. The episode wraps with an update on NotebookLM. Craig walks through the recent changes — more user control over reports, slide decks, flashcards, quizzes, and other outputs in the Studio panel. You can now specify the structure and focus of custom reports rather than relying solely on canned formats. Slide decks can be exported (though editing remains clunky since each slide is essentially an image). Craig's recommendation: if you have a Google account and you work with knowledge in any form, you should be using NotebookLM. Rob notes that Microsoft Copilot has added a similar notebook feature worth exploring, and they float the idea of a future head-to-head comparison episode. Links referenced in this episode: 1. notebooklm [https://notebooklm] 2. anthropic [https://anthropic] 3. claude [https://claude] 4. google [https://google] 5. canvas [https://canvas] 6. einstein [https://einstein] Mentioned in this episode: AI Goes to College Newsletter