October 2, 2026

AI for the Over 40: Week 35 - Where AI Literacy Meets AI Fluency

In the last series blog, I wrote about the humbling experience of auditing my own AI skills, discovering that the most important techniques I had developed over 33 weeks had never been named, and that the simplest practical habits were the ones I still skipped. I walked away with a four-skill framework: Set the Stage, Read It Like a Skeptic, Work the Conversation, Think Bigger / Think Better.

Three days later, Anthropic published a research report that measured many of the same behaviors across nearly 10,000 real conversations.

My first reaction was not validation. It was curiosity. I have been preaching literacy before agency for months, so I wanted to know what this study revealed about fluency, and how it overlapped or challenged what I had been saying. But as I read deeper, what excited me most was not the findings themselves. It was the idea that fluency could be measured. I had spent 34 weeks advocating for AI literacy without ever thinking seriously about how you would know if someone had it. What would improvement look like? How would you track it? And more importantly, could you build a system that actively coached you toward it?

Within a few hours of finishing the report, I had built exactly that. But I am getting ahead of myself.

What the 4D AI Fluency Framework measures

The study is built on the 4D AI Fluency Framework, developed by professors Rick Dakan (Ringling College of Art and Design) and Joseph Feller (University College Cork) in collaboration with Anthropic. It defines four core competencies for effective human-AI collaboration.

Delegation is deciding whether, when, and how to engage AI. Description is communicating effectively with AI. Discernment is critically evaluating AI outputs. Diligence is taking responsibility for what you do with AI.

Each competency breaks down into subcategories, 12 specific behaviors in total. The researchers used Anthropic’s privacy-preserving analysis tool to study which of these behaviors appeared in 9,830 Claude conversations during a one-week window in January 2026.

Here is the critical limitation: they could only observe half of the behaviors, the ones that happen inside the chat window. The ones that cannot be observed inside the chat window, including things like being honest about AI’s role in your work and considering the consequences of sharing AI-generated output, happen after you close the conversation. Those are arguably the most consequential dimensions of fluency, and they are invisible to any platform that is watching.

I will come back to that gap. It matters more than the findings themselves.

Three AI fluency findings every business leader should know

Finding 1: Iteration is everything, and most people are not doing it.

The single strongest predictor of every other fluency behavior was iteration and refinement, treating AI’s first response as a starting point rather than a final answer. Conversations that included iteration showed roughly double the rate of all other fluency behaviors. They were 5.6 times more likely to involve users questioning AI’s reasoning, and 4 times more likely to see users identifying missing context.

Last week I called iterative refinement “the most important unnamed skill in the entire series.” I had been demonstrating it in every article for 33 weeks without once calling it out as something to learn. Now there is data showing it is the single strongest correlate of AI fluency across nearly 10,000 conversations. Not just important. Foundational.

If you are using AI and accepting the first response, you are not collaborating. You are placing an order.

Finding 2: The better AI looks, the less people think.

This is the finding that should genuinely give leaders pause. When AI produces polished artifacts, code, documents, interactive tools, things that look finished, users become more directive upfront but less critical of what comes back. They are less likely to question the reasoning, less likely to check facts, less likely to identify what is missing.

Read that again. The more competent the output appears, the more people let their guard down.

But I think the research is pointing at something deeper than people failing to review polished output. The real issue predates AI entirely: confirmation bias.

We have always done just enough research to justify our point of view. The difference is that without AI, I had an external gauge of how seriously I had challenged my own assumptions. If I spent no time researching, I could not pretend to have overcome my own preferences. If I spent hours digging through sources, I could at least make a case that I had tried. The effort itself was a signal, imperfect, but visible.

AI eliminates that signal. It produces well-formed, authoritative-looking output almost instantaneously. I can review it carefully, confirm that AI agrees with me, and present it as if it were the product of hours of painstaking research, because I have no external gauge of whether AI is telling me what I want to hear or whether the analysis is truly objective. And it is just amazing how often AI confirms what I believed all along.

This is the abdication problem I have been writing about, now measured. It is not that AI produces bad output. It is that AI produces output that looks good enough to stop thinking, and that feels like independent validation of conclusions you had already reached.

Finding 3: 70% of people never set the terms of the collaboration.

Only 30% of users in the study told Claude how they wanted it to interact with them. Things like “push back if my assumptions are wrong” or “walk me through your reasoning before giving me the answer” or “tell me what you’re uncertain about.”

Seventy percent of people are showing up to a collaboration without establishing how the collaboration should work. That is like hiring a consultant and never telling them what you actually need, then being disappointed when they give you something generic.

Week 34 Skill 1 was “Set the Stage,” everything before AI’s first response. Context, roles, constraints, format. The research says most people skip all of it.

How the 4Ds compare to the four AI skills framework

Here is where it gets interesting, and where I want to be honest about what this research revealed about my own blind spots.

The 4D Framework breaks AI fluency into Delegation, Description, Discernment, and Diligence. Last week I broke my skills into Set the Stage, Read It Like a Skeptic, Work the Conversation, and Think Bigger / Think Better. The overlap is significant but not perfect, and the gaps are revealing.

My “Set the Stage” maps roughly to their Description, how you communicate with AI. My “Read It Like a Skeptic” is their Discernment, critically evaluating what comes back. My “Work the Conversation,” iteration, context management, knowing when to start over, is what their research showed as the single most important behavior.

But their framework has two areas where my self-taught approach has gaps.

Delegation, the decision about whether to use AI at all, and which modality to use, is something I have mostly learned by instinct rather than by design. I have written about automation versus augmentation, about the progression from literacy to agency. But the deliberate, conscious choice of “should AI be involved in this task, and if so, how?” is a muscle I developed by trial and error, not by intention. The framework names it as a discrete competency. I treated it as obvious. It is not obvious to someone just starting.

Diligence, taking responsibility for AI-assisted work, is the one that hits hardest. Not because I have been irresponsible, but because this is the competency that matters most for organizational transformation and it is almost entirely invisible to measurement. Anthropic could not observe half of the behaviors because those behaviors happen after the conversation ends: verifying before publishing, being transparent about AI involvement, considering stakeholder impact, maintaining ethical standards.

Those invisible behaviors? That is where organizational AI transformation lives or dies. And it is exactly the territory I am moving into as this series shifts from personal literacy to leadership.

Building an AI fluency coaching system from research

Most people will read that research report, nod thoughtfully, and change nothing. I know this because I have watched it happen with every AI insight I have shared over 34 weeks. Reading about fluency does not build fluency. So I decided to build something.

The question was simple: could I take the 4D Framework and embed it directly into my AI project instructions, not as a reference document, but as an active coaching system? Instead of just knowing what fluency looks like, could I get AI to flag when I was falling short in real time?

I used the research itself as my starting point and collaborated with AI to design the system. The first design decision was what not to do. A static self-rating, “I’m a 3 out of 5 on Discernment,” would go stale immediately. Instead, we designed behavioral triggers mapped to each of the 4Ds. Specific patterns that, when AI observed them in our actual work, would prompt a coaching intervention.

For Delegation gaps: if I ask AI to make a judgment call that benefits from my domain expertise, it pushes back and asks for my instinct first. If I am doing something manually that is a strong candidate for AI augmentation, it flags the missed opportunity.

For Description gaps: if my prompt is vague when the task demands precision, it asks for specifics rather than guessing my intent. If I am not providing enough context for quality output, it tells me what is missing rather than working around it.

For Discernment gaps: if I accept a complex output without questioning it, it flags the weak point and asks if I want to stress-test it. If I am treating AI output as authoritative in areas that require verification, it names the risk.

For Diligence gaps: if I am preparing to share AI-assisted work without discussing verification, it asks about my review process. If I am not considering stakeholder impact or transparency, it raises what I am missing.

The hardest design problem was timing. When I first tested the concept, I realized that AI would read the fluency instructions and immediately start coaching, before it had any signal to coach on. Getting a Discernment nudge on your opening message would be insufferable. So we added a calibration rule: fluency coaching is observational, not preemptive. Never coach in the first response. Wait for actual behavioral evidence. And if the conversation pace indicates rapid task execution, those moments when you are just trying to get something done quickly, hold observations for a natural pause rather than interrupting flow.

That calibration detail is small, but it is the kind of thing you only figure out by building and testing. The research can tell you that iteration matters. It cannot tell you that your coaching system needs a timing delay to avoid becoming noise.

I have since embedded these instructions across my active projects. It is early, but I am already noticing something: just knowing the coaching system is there changes my behavior. I catch myself being more specific in my prompts because I know vagueness will get flagged. I pause before accepting polished output because I know the Discernment trigger is watching. The measurement itself is changing what it measures.

AI literacy vs. AI fluency: what’s the difference

There is a language choice in this research worth addressing. I have been using “literacy” throughout this series. Dakan and Feller chose “fluency.” Those are not synonyms, and the distinction matters.

Literacy is foundational understanding, can you read the language? Fluency is comfortable, natural application, can you think in the language? What I have been documenting over 34 weeks is actually the progression from one to the other. Early weeks were literacy: understanding what AI is, how it works, where it breaks. Later weeks, building solutions, diagnosing organizational problems, designing workshops, embedding coaching systems into project instructions, that is fluency.

The distinction matters because most organizations are trying to achieve fluency while skipping literacy. They want their teams to think naturally with AI without first ensuring those teams understand what AI actually is. The research shows what that looks like in practice: people who are directive but not evaluative, who specify what they want but do not question what they get. That is not fluency. That is ordering confidently from a menu in a language you do not speak.

What this means for AI governance and adoption

If you are responsible for an organization’s AI adoption, this research gives you three actionable insights.

First, stop measuring adoption and start measuring fluency. How many people have AI accounts is meaningless. How many are iterating, questioning, setting collaboration terms, that tells you whether your investment is producing capability or just consumption.

Second, the biggest risk is not AI failure, it is AI success. The finding that polished outputs reduce critical thinking should reshape how you think about AI governance. Your team will not get burned by obviously bad AI output. They will get burned by output that looks professional, reads well, and is subtly wrong in ways that matter. Especially when it confirms what they already believed.

Third, the invisible behaviors are your responsibility to build. Anthropic cannot measure whether your team is being transparent about AI use, verifying before publishing, or considering stakeholder impact. No platform can. Those are organizational culture decisions, and they require leadership that has experienced the full journey from literacy to fluency, not leadership that skipped to the end.

Your Week 35 challenge: run your own fluency audit

1. Score yourself against the 4Ds. Delegation: Are you making conscious decisions about when to use AI and which modality to use? Description: Are you setting context, specifying format, providing examples? Discernment: Are you questioning AI’s output, catching errors, evaluating whether the collaboration is working? Diligence: Are you verifying before sharing, being transparent about AI involvement, considering stakeholder impact? Be honest about where you are strong and where you are coasting.
2. Check your iteration habit. Look at your last five AI conversations. How many times did you accept the first response and move on? How many times did you push back, refine, redirect? If you are mostly accepting first drafts, you are leaving the most important fluency behavior on the table.
3. Set the terms once. In your next AI conversation, start by telling it how you want the collaboration to work. “Push back on my assumptions. Tell me what you’re uncertain about. Don’t just agree with me.” See how the conversation changes. You are joining the 30% who actually do this.
4. Examine your Diligence. The next time you are about to share, publish, or act on AI-assisted work, pause. What is your verification process? Who knows AI was involved? What is your accountability if something is wrong? These questions do not happen inside the chat window. They happen inside you.

The bottom line

Thirty-four weeks of documenting my own AI transformation, and a research team just published data validating the core patterns I discovered through trial and error. Iteration matters most. Polished outputs are the most dangerous, especially when they confirm what you already believed. Most people never set the terms of the collaboration. And the behaviors that matter most for organizations cannot be measured by any platform, they require human judgment, human accountability, and human leadership.

But validation is not the point. Action is. Within hours of reading this research, I had built a coaching system that uses the 4D Framework to actively improve my own fluency in every AI conversation. Not because reading the research made me fluent. Because 34 weeks of building things taught me that the distance between insight and capability is always the same: you have to do something with what you learn.

I have been calling this “literacy before agency.” The academic framework calls it developing fluency across Delegation, Description, Discernment, and Diligence. The vocabulary is different. The destination is the same: you have to do the work yourself before you can lead others through it.

Nobody develops AI fluency by reading a research report. Just like nobody developed it by reading this series. You develop it by opening a conversation, setting the terms, doing the work, questioning the output, and showing up again tomorrow.

Thirty-five weeks in, I am still showing up. The research just confirmed why it is working.

Sources referenced:

• Swanson, K., Bent, D., Ludwig, Z., Dakan, R., & Feller, J. (2026). “Anthropic Education Report: The AI Fluency Index.” Anthropic, February 23, 2026.
• Dakan, R. & Feller, J. “The 4D Framework for AI Fluency.” AI Fluency: Framework & Foundation.” Anthropic Academy. Framework documentation.

This post is part of my “AI Over 40” series. It first appeared on LinkedIn: Where AI Literacy Meets AI Fluency

Read more AI and Copilot blogs.

Trending Posts