AI Fluency Framework for Educators

Summary

A classroom that lets students use AI without a shared standard for what "good" looks like is training them to settle for good enough.

That was the argument Rick Dakin, head of creative technologies and AI coordinator at Ringling College of Art and Design, made in a session on the AI fluency framework he co-developed with research partner Joseph Feller at University College Cork. The framework, later expanded into a set of free courses with Anthropic that has logged over a million enrollments, organizes AI use into four competencies: delegation, description, discernment, and diligence. Dakin walked through each one, then applied them to three modes of working with AI (automation, augmentation, and agency) and four distinct roles AI can play in a finished piece of work, from quiet consultant to full creative medium. He closed with a case for redesigning assignments around process instead of product, and fielded questions on cognitive dependence and what employers actually look for in new hires. The throughline: fluency isn't about knowing which button to press. It's a discipline-specific judgment call about when AI helps, when it doesn't, and how to prove the difference.

Key Takeaways

• The AI fluency framework rests on four competencies (delegation, description, discernment, and diligence), each of which splits into three more specific sub-skills.
• Unguided AI access lowers test scores, while a well-structured AI tutor tied to sound teaching methods raises them; the difference comes down to how the tool is used, not whether it's used.
• AI use falls into three modes: automation (AI executes a defined task), augmentation (human and AI trade ideas back and forth), and agency (AI pursues a goal across multiple steps on its own).
• Within finished work, AI can occupy four distinct roles: consultant (no AI output ships in the final piece), tool builder (AI builds the infrastructure, a human makes the work), cocreator (AI-generated material ships, credited and edited by a human), and medium (the piece could not exist without AI running inside it).
• Redesigning assignments around process rather than final product (asking students to submit drafts, prompts, and edits alongside the finished piece) makes AI use visible without relying on unreliable AI-detection tools.
• Employers report they'd rather hire curious, adaptable learners than people trained on one specific AI tool, since the tools in use on a given job will likely change before a new hire's first day.

Deep Dives

Four Competencies, Not One Skill

Dakin opened by separating literacy from fluency. Literacy is knowing the basics of how a tool works. Fluency is the next level: using AI systems in ways that are effective, efficient, ethical, and safe, all at once. He built that definition into four competencies, each subdivided into three narrower skills, twelve in all. Delegation covers setting goals and deciding whether AI belongs in a task at all. Description covers communicating those goals to the AI clearly enough to get useful output. Discernment covers judging whether what comes back is actually good. Diligence covers taking responsibility for the result, including disclosing how AI was used.

The framework isn't a checklist to run once. Dakin described it as two nested loops: an outer loop that sets the terms before any AI tool opens, and an inner loop of description and discernment that runs moment to moment during the work itself. "Only once you've answered the sort of big picture questions do you then go into sort of moment to moment, hour to hour usage of that description and discernment," he said.

He built the framework with Joseph Feller at University College Cork more than two years ago and has used it in his own teaching since. All of the underlying material sits under a Creative Commons license, free to copy and adapt for teaching. The best-known extension is a set of courses co-created with Anthropic; Dakin noted his research group holds joint copyright and was paid to help produce the videos, but received no ongoing payment from the company. "I received just zero dollars from Anthropic. I'm not an employee of Anthropic," he said, framing the partnership as collaborative rather than sponsored. As of last month, that core course had logged over a million enrollments and more than 550,000 completions worldwide.

Dakin was direct about why any of this matters. Studies now show both outcomes at once: a pedagogically sound, custom-built AI tutor can produce substantially higher test scores, while unsupervised access to a general AI tool lowers them. "Success in teaching and learning and really any use of these tools depends on how we use AI, how we teach others to use AI," he said. The framework exists to make that "how" explicit instead of leaving it to chance.

Three Ways to Hand Work to AI

Inside the delegation competency, Dakin split AI use into three modes, each requiring a different kind of oversight. Automation is the simplest: the AI performs a task the human has already defined, whether that's generating an image, summarizing a spreadsheet, or reading a paper. Success here depends on stating the goal precisely and recognizing quality output on sight, which loops straight back to discernment.

Augmentation is more open-ended. Human and AI move back and forth, building on each other's prompts and ideas, often without a fixed target in mind at the start. Dakin placed this mode closest to genuine collaboration: exploring possibilities together rather than handing off a finished brief.

Agency is the mode Dakin flagged as having shifted fastest. "Agency is what's really come on strong in the last year, year and a half, which is where you configure the AI to go perform multiple tasks independently," he said. Give the system a goal, and it works through the steps on its own: a coding agent, a research assistant, or any chatbot configured to complete a multi-step job without a human checking in after each move.

The three modes aren't ranked by sophistication. Dakin's point was that each one demands a different amount of upfront description and ongoing discernment, and that mismatching the mode to the task is where fluency breaks down. Handing an ill-defined creative brief to pure automation produces flat, generic output; handing a well-specified routine task to an unsupervised agent invites errors nobody catches until it's too late. The skill is matching the mode to the work, not defaulting to whichever one is fastest to set up.

Four Roles AI Can Play Inside a Finished Piece of Work

Beyond the three modes of delegation, Dakin described four distinct roles AI can occupy in a piece of finished work, a distinction he said matters most in creative fields, where disclosure and audience expectations carry real weight.

As a consultant, AI never appears in the final product. It's used for brainstorming, critique, or research, but "the actual work output remains entirely human made," Dakin said. He gives his own students an exercise built on this role: have a long conversation with AI about something you already know well, specifically to find where the tool's answers fall apart. Knowing a tool's failure points, he argued, is what makes its praise trustworthy elsewhere.

As a tool builder, AI writes the scripts and utilities that support the work rather than the work itself: renaming files, moving assets between programs, generating throwaway prototype code to test an idea before committing real time to building it.

As a cocreator, AI-generated text, images, or code ships directly in the final piece, credited and edited by a human. Dakin called this the most contested role in design education, since standards on disclosure vary by institution and industry, and using AI-generated work against a stated policy can trigger academic sanctions.

The fourth role, AI as a medium, is the one Dakin was most animated about: work that could not exist without AI running inside it, not as a production shortcut but as a constituent part of the experience. His examples ranged from AI tutors students converse with directly to generative art installations that respond to a room in real time. "The artifact, whatever it is you're making, could not exist without AI running," he said, distinguishing it clearly from cocreation, where a human could in theory have produced the same output alone.

Grading the Process, Not Just the Paper

Asked how to redesign assessments for an AI-saturated classroom, Dakin didn't reach for detection software. "I don't recommend using AI detector tools. I don't think they're good enough to be reliable," he said, adding that he doesn't want to be in an adversarial standoff with his own students over the question.

Instead, he restructured the assignment itself. In his AI course, students don't submit a finished story. They submit six documents: their initial prompt, the AI's response, their edits to that response, an AI-generated draft built from the edited outline, their edits to that draft, and finally the finished piece. "I'm gonna look at the changes you made," he tells them upfront, rather than reading every word of every draft. The grade tracks the decisions made along the way, not just the polish of the final page.

He connected the approach to what employers have told him they actually want. Hiring managers, he said, aren't primarily looking for fluency in one specific AI product, since those tools will likely have changed again by the time a new hire's first day arrives. "They want to hire people who are curious, who are lifelong learners, who are willing to engage with the tools," Dakin said, which made demonstrated adaptability, not tool-specific expertise, the trait worth grading for.

He was candid that process-based grading doesn't fit every assignment. A finished illustration or film still has to exist as a finished illustration or film. But wherever the format allows it, he pushes toward capturing the decisions behind the output and having students critically reflect on both the AI's work and each other's, a shift he sees less as a workaround for AI than as overdue attention to how learning actually happens.

The Cognitive Cost, and What Students Do With AI When It's Their Choice

A question about whether frequent AI use makes people sharper or duller got a direct answer: it depends entirely on how the tool gets used, but the risk is real. "Cognitive decline, to use the fancy word, or dependence on these tools is a real, real danger," Dakin said, comparing it to how few people can read a paper map anymore now that phones handle navigation. Some tradeoffs like that one feel fine. The concern is that AI now handles thinking work that sits closer to the center of who a person is. "These tools are now doing thinking and cognitive work that is really important to who we are," he said, calling it something to actively guard against rather than accept by default.

He offered his own use as a counterexample of dependence done well. Working through a statistics course as part of a PhD program, he uses AI as what he called the best statistics tutor available for walking through chi-squares and standard deviations, a subject he doesn't find intuitive. The same logic applied to a robot he and colleagues built with a student mascot project, and to a student in his class who used AI to teach herself crochet mid-semester and brought a finished piece to class. What made those examples work, in his telling, was that the students chose the subject themselves. Citing the philosopher John Dewey's century-old argument that learning sticks when it's tied to what a student already cares about, Dakin said the same principle now decides whether AI supercharges a student's ability or quietly replaces it. "If there's something I wanna do, I'm just pretty sure an AI can help me figure out how to do it," he said, framing it as an opportunity that only pays off once a learner has something they actually want to build.