Skip to main content

The literacy rollout: training 400 people who did not ask for this

The short answer

A rollout changes behavior; it does not deliver information

The goal is a reflex that fires at the moment of use, not a fact people can recite. Vanderbilt did not lack information; it lacked a shared reflex about when a chatbot is wrong for the task. Every design choice follows from treating the rollout as behavior change.

What you will be able to do

  • Execute an Article 4 literacy rollout across a whole workforce, turning the program you designed in Topic 5.2 from a document into installed behavior, using the role segments your workforce map already defined (see Topic 9.1).
  • Segment and sequence the rollout so each role gets content pitched to what it actually needs, in an order that reaches the highest-risk people first, rather than one identical session pushed at everyone.
  • Design delivery that lands, applying how adults actually learn (relevance, problem-centered practice, and reflexes rehearsed at the moment of use) rather than a slide deck people click through and forget.
  • Measure behavior, not attendance, distinguishing who showed up from who changed what they do, because a completion rate is not evidence that the reflex was installed.
  • Capture the evidence the rollout produces, completing the dated Article 4 literacy program evidence record that a regulator, a conformity file, and the Module 13 dossier will later ask you to produce (see Topic 5.6) (see Topic 13.1).
  • Handle the reluctance of people who did not ask for this, reading resistance as information about your rollout and your tools rather than as an obstacle to push through (see Topic 9.3).
  • Defend the rollout against the challenge that it was theater: prove, with the record and the behavior change, that your people can now catch the specific failures your systems produce.

The ideas, one by one

Segment by relationship to AI, then tailor

Operators, communicators, builders, approvers, and general staff face different failures and need different content pitched at different depth. One generic session reaches no one; relevance is the currency of adult learning, and segmentation is how you buy it.

Sequence by risk, not by scheduling ease

Reach the highest-harm, highest-exposure people first, tie a system's overseers to its go-live date, and run the first segment as a real pilot to fix delivery before scaling. Starting with the biggest easy group optimizes the metric and misses the point.

Deliver practice, not a lecture

Adults learn from relevant, hands-on work in their own domain. Make people catch a planted error in a real output from their own system; do not make them watch a slide describe verification. Short and role-specific beats long and comprehensive.

Measure behavior, not attendance

A completion rate measures exposure, which is not the goal and is the number that lets you lie to yourself. Use realistic-task checks and behavioral signals, and name the result that would mean the rollout failed for a segment.

Reach the people who are not in the room

Contractors and outsourced teams who act on your behalf are in scope; shadow-AI users are your highest hidden risk; senior people assume they are exempt. A rollout scoped to the employee directory misses exactly the populations that produce public failures.

Plan the fade

One session does not install a habit; memory of a single exposure decays fast. Build spacing, point-of-use prompts, and trigger-based refresh into the first design, tied to the inventory's heartbeat, or the behavior you installed will quietly disappear.

Read reluctance as information

People who did not ask for this are pushing back for reasons that improve the rollout: irrelevance, fear of layoffs, wasted time. Fix the cause. And never deliver literacy inside an unspoken lie; the honest change narrative comes first, or the training is theater the workforce sees through (see Topic 9.6).

The rollout is finished at the evidence, not the last session

The obligation is to prove your people are literate, not to have run sessions. Complete a dated, inventory-wired evidence record, because a conformity file, a regulator's letter, and the Module 13 audit will each ask you to produce it (see Topic 5.6) (see Topic 5.7) (see Topic 13.1).

Listen to it

The same lesson, talked through by two hosts as episode 71 of the podcast.

Real cases

These are documented cases used to illustrate the rollout, not to predict your organization. Each is cited and used for a specific point.

Example 1: Vanderbilt's Peabody EDI office and the ChatGPT condolence email (this topic's anchor). In February 2023, following the Michigan State University shooting, Vanderbilt's Peabody College EDI office sent a consolation email written with ChatGPT and left the AI citation in the text; the message had skipped the college's normal review layers, the dean had not seen it, and students reacted with anger. An associate dean apologized for "poor judgment" and she and an assistant dean stepped back from EDI duties during the review (CNN Business, 22 February 2023). The value for a rollout is precise: this was not a knowledge failure by one person but an organization-wide literacy gap, made visible in public at a moment of grief. It shows exactly the segment (communicators producing sensitive public content), the failure mode (using AI for a task where human judgment and appropriateness were the whole point), and the missing reflex (stop, question whether this is even a task for AI, and keep the human review that was bypassed). A rollout that had installed that reflex in that segment is the difference the topic is about.

Example 2: The reflex that would have caught a fabricated output. The verification reflex that literacy installs, "a fluent AI output is a draft, not a fact, confirm it against a real source before it goes out under your name," is the same reflex missing in the newsroom cases owned elsewhere, such as the Chicago Sun-Times summer reading list that paired real authors with AI-invented book titles (see Topic 5.2) and the CNET AI-written finance articles that had to be corrected (see Topic 3.7). Those are owned by other topics; here they are the pattern that shows a rollout's target. In every case, a trained reflex at the point of use was cheap and its absence was expensive. The rollout exists to make that reflex the default across the specific people who could reproduce the failure in your context.

Example 3: Technical skill is not literacy. Builders (engineers, data scientists) are frequently exempted from literacy rollouts on the assumption that technical people already understand AI. Article 4 literacy is not technical proficiency; it centers on awareness of risks and possible harm, including harm to an affected person, which a strong engineer may never have considered (see Topic 5.2). The rollout point: a segment can be highly capable with the technology and still lack the literacy the obligation requires, so the rollout tailors depth to the builders (bias, evaluation, drift, downstream harm) rather than skipping them.

Example 4: A public-sector framework for what "literate" means. The US Department of Labor issued its AI Literacy Framework to the public workforce system via TEN 07-25 (13 February 2026), defining AI literacy across content areas and delivery principles as voluntary, non-binding guidance (US DOL Employment and Training Administration). It is a useful reference for a rollout designing per-segment content, and a reminder that literacy frameworks are proliferating globally; the rollout borrows the substance without importing any single instrument's politics, and remembers that the binding obligation in scope here is the EU AI Act's Article 4, not the DOL guidance.

Example 5: A management-system standard's competence requirement. ISO/IEC 42001:2023, the first certifiable AI management system standard, includes clauses on awareness and competence for people whose work affects AI performance (ISO/IEC 42001:2023; established as a voluntary, certifiable standard). For an organization pursuing that certification, the literacy rollout is part of how the competence requirement is met and evidenced. The point for the rollout: the same well-run rollout that satisfies Article 4 also feeds a management-system standard's competence clause, so the evidence record does double duty.

Example 6: Where the sensitive-topic reflex matters most. The Vanderbilt case is specifically a failure of appropriateness, not just accuracy: the problem was not only that ChatGPT wrote imperfectly, but that a chatbot was used at all for a message whose entire value was human presence in the face of tragedy. The rollout lesson that generalizes: for communicators, literacy must teach not only "verify the output" but "decide whether this is a task for AI at all," because some tasks (condolence, apology, discipline, anything where the human relationship is the point) fail the moment a machine is known to have written them, regardless of how good the text is. That judgment is a taught reflex, and the rollout is where it is installed.

Example 7: The shadow-AI disclosure the rollout should welcome. A well-documented pattern across organizations in 2023 to 2024 was staff pasting confidential material into consumer AI tools. The most cited case is Samsung engineers who reportedly entered sensitive internal code and notes into ChatGPT, prompting the company to restrict such use (owned as a data-leak case by a sibling program, referenced here only for the rollout point). The literacy-rollout lesson is about the general-awareness track and the disclosure culture around it: the danger was not that people are malicious but that they did not know the tool was not a private, contained workspace. A rollout that makes it safe to say "I have been using a consumer tool for work" surfaces exactly this risk, converts it into a literacy moment (the universal reflex about not feeding confidential data into tools you do not control), and adds the tool to the inventory rather than driving it underground (see Topic 0.2). The pattern shows why the general-awareness segment is not filler: it is where the largest, quietest exposure lives.

Example 8 (a general pattern, not a documented case): a rollout that measured the wrong thing. Unlike the seven cited cases above, this is not a specific, sourced incident; it is a structural pattern named here because it is worth naming plainly. In corporate training generally, well before AI, organizations have long reported near-total completion of a mandatory course and then been blindsided by the very behavior the course was meant to prevent. No single named source is cited for this because the point is the structure, not an instance: completion is a lagging comfort, not a leading indicator, and an organization that tracks only completion learns nothing about behavior until a failure teaches it the hard way. This is the generic version of what makes the Vanderbilt-class failure so instructive: somewhere there was almost certainly a record that people had been told to use AI carefully, and it did not matter, because telling is not installing. A rollout that measures behavior catches the gap before the failure does; one that measures attendance finds out afterward.

Where people go wrong

  • Treating the rollout as information transfer. The deepest error. A literacy rollout is a behavior-change program; the goal is an installed reflex that fires at the moment of use, not a fact the person can recite. Vanderbilt did not lack information; it lacked a shared reflex. Design for behavior, or you will produce a workforce that has been exposed to literacy and behaves exactly as before.
  • One identical session for everyone. The 400 people do not have the same relationship to AI. Pushing one generic session at an operator, a communicator, an engineer, and an executive wastes some, terrifies others, and reaches none with the reflex their actual work needs. Segment by relationship to AI and tailor the content, because relevance is the currency of adult learning.
  • Sequencing by scheduling ease instead of by risk. Starting with the biggest, easiest group to post a fast completion number leaves your highest-risk operators and communicators untrained. Reach the highest-harm, highest-exposure people first, and tie the sequence to your deployment roadmap so a system's overseers are literate before it ships.
  • Measuring attendance and calling it evidence. "97 percent completed" measures exposure, not behavior. It is precisely the sentence Vanderbilt could have said the week before its email. Measure at the behavior level (realistic-task checks, escalations of suspect outputs, sampled real work), because a completion rate is the number that lets you lie to yourself.
  • Delivering a passive lecture. The recorded webinar with a quiz is cheap to make and easy to track and installs almost nothing. Adults learn from relevant, problem-centered, hands-on practice in their own domain. Make people do the reflex on a real output from their own system, do not make them watch a slide describe it.
  • Assuming one session is enough. Memory of a one-time exposure decays fast. A reflex needed under pressure months later will not survive on a single launch. Build spacing, point-of-use prompts, and trigger-based refresh into the first design, or the behavior you installed will quietly fade.
  • Scoping the rollout to the employee directory. Contractors, agency staff, and outsourced teams who act on your behalf are in Article 4 scope and are often the likeliest to use AI outside your controls. A rollout that reaches only badged employees silently excludes exactly the populations that produce public failures. Reach them through contracts and onboarding, and record it.
  • Exempting the senior people. Executives who skip training or assume seniority equals literacy are exactly the approvers who sign a high-risk deployment without the right question, and the leaders whose sensitive public statements become the biggest-blast-radius Vanderbilt. Build a short, respectful version for them, but do not exempt them.
  • Stopping the rollout one level below the boardroom. A board that approves AI spend and owns AI risk but never sits a literacy session is overseeing systems it cannot interrogate, and appointing one AI-expert director does not close the gap: by 30 June 2025 only 14 percent of boards had effectively integrated AI expertise while 25 percent had recruited an expert (MSCI Institute, 2026). Collective literacy across the whole table is what lets every director challenge management instead of deferring to one colleague. Put the board in the segment map, keep its session short and free of tool demonstrations, and repeat it on a fixed cadence, because a director's qualification at appointment decays as the technology and the law move.
  • Punishing shadow-AI disclosure. If admitting "I have been using an unapproved tool" gets a person in trouble, the rollout drives shadow AI underground. Treat every disclosure as a literacy opportunity and a new inventory entry, because a workforce that feels safe to disclose becomes your sensor network for the systems you did not know you ran.
  • Delivering literacy inside an unspoken lie. A rollout that is really a prelude to layoffs, presented as neutral upskilling, is a betrayal the workforce detects and never forgives. If roles will change or end, the honest change narrative and the hard conversations come first; literacy delivered on top of hidden fear is theater the audience can see through (see Topic 9.6).
  • Teaching only accuracy, never appropriateness. For communicators especially, verifying the output is not enough. Some tasks (condolence, apology, discipline, anything where human presence is the point) fail the moment a machine is known to have done them, however good the text. The rollout must install the prior judgment: is this a task for AI at all, which is exactly the reflex Vanderbilt lacked.
  • Finishing at the last session instead of at the evidence. The obligation is not "we trained people," it is "we can prove our people are literate about the systems they touch." A rollout that ends without a complete, dated, inventory-wired evidence record has performed literacy without being able to prove it, which fails the moment a regulator, a conformity file, or an auditor asks (see Topic 5.6) (see Topic 5.7) (see Topic 13.1).

Questions people ask

What is literacy rollout?
The delivery of an Article 4 literacy program across a real workforce so that it becomes installed behavior, distinct from the design of that program (Topic 5.2). The crossing from a document to a workforce in which the verification and appropriateness reflexes actually fire under pressure.
What is article 4 (EU AI Act)?
The provision of the European Union's Artificial Intelligence Act (Regulation (EU) 2024/1689) requiring providers and deployers to take measures to support the development of AI literacy among their staff and others operating AI on their behalf; in force since 2 February 2025, one of the first two obligations in the Act to bind, alongside the Article 5 prohibited-practices ban, and rewritten by Regulation (EU) 2026/1744 (in force 27 July 2026) so that it expressly guarantees no individual's level. More on Article 4 (EU AI Act)
What is article 4 literacy program evidence?
The named, chained artifact that is designed in Topic 5.2 and completed by the rollout in this topic: a dated, inventory-wired record of who was covered, what content tied to which systems, when, at what sufficient-level bar, and what behavior checks showed. Cited by a conformity file (5.6), produced for a regulator (5.7), and inspected in the Module 13 dossier.
What is behavior change (versus information transfer)?
The correct frame for a rollout. The goal is an installed reflex that fires at the moment of use, not a fact the person can recite. Information transfer produces a workforce that has been exposed to literacy and behaves as before; behavior change produces one that catches bad AI outputs.
What is literacy segment?
A group of people defined by their relationship to AI (operators, creators/communicators, builders, approvers/overseers, general awareness), not by department, so that content and depth can be tailored to the failure modes and risk each group actually faces.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →