How Do I Design Faculty Development for AI That Actually Changes Practice?

It is February. Your provost has announced an AI initiative with a line item, and you have one semester to make it show up in classrooms. Or you ran six workshops last fall, the rooms were full, and the teaching you sit in on has not changed. Or a dean asks the plain question at budget time: what did the money buy?

All three come from the same design fault. A workshop shows people what a tool does. Practice changes when a teacher spends supported, structured time on a problem in their own course, long enough for a new habit to displace an old one, with the institution saying yes to the experiment rather than merely tolerating it.

The bottom line for the meeting: standardize a session cycle where every participant brings a real course artifact and leaves with a revised one, put verification, disclosure and risk control inside the working document instead of a closing ethics talk, stretch the program past a single day and pair it with mentoring, run discipline-specific cohorts, and report the artifacts people produced at a follow-up rather than how confident they felt on the last afternoon.

The short version

1. Anchor every session in an artifact from the participant's own course, and end the session with that artifact revised.

2. Put verification, disclosure and risk control in the working template, where they get used, so a nervous participant has somewhere to put the worry.

3. Spend the hours on assessment and rubrics, the weakest area at baseline and the one with the largest measured gain.

4. Run discipline-specific cohorts, then differentiate inside them by experience.

5. Secure the conditions: stated permission to experiment, protected time, materials in the languages your faculty work in.

6. Evaluate revised artifacts at a delay, plus one relational indicator, and treat satisfaction scores as noise.

Run every session on the participant's own course

The cleanest comparison here is Bi et al. (2026), who surveyed 568 faculty across five universities and ran a quasi-experiment with 160 of them. Both arms got four 90-minute workshops over consecutive weeks, with identical contact time, platform and facilitator support. The content differed. One arm got conventional GenAI faculty development, the other a structured cycle of prompt formulation, output review, course-material revision and written reflection against a seven-component template covering objectives, task constraints, assessment criteria, verification, disclosure and risk control.

The structured arm scored higher on all seven readiness dimensions. Its largest advantage sat in prompt design, at roughly two standard deviations, and it was the only arm whose prompt-design gain survived delayed testing; the conventional arm's smaller gain had fallen back below baseline by follow-up.

Two more results explain why demonstrations do not travel. In a within-subjects study with 13 higher-education teachers, chatbot access raised the share of activities reaching two or more higher-order Bloom tasks from 41.7% to 91.7% (p = .01) and cut perceived cognitive effort from a median of 6.33 to 3.33 (p = .004). The pedagogical prompting training layered on top did not improve design quality further, and it slightly raised effort. Meanwhile, in a workshop with 60 middle-school science teachers, the productive element was collaborative review of ChatGPT-generated assessment questions, which is what surfaced conceptual-precision problems and the risk of reinforcing Misconceptions about AI.

What those studies share: the learning came from critiquing and revising material. People improve the thing in front of them, and they improve it faster when a colleague is reading it too.

Stop treating skill as the missing piece

The tempting diagnosis is a capability gap, closed by better instruction. Bai and Hsieh (2026) surveyed 898 university teachers on AI-TPACK scales and found that teachers who understood how generative tools work, or could operate them competently, were no more likely to report inventing and implementing new teaching practices than those who could not. The integrated AI-TPACK construct had no direct path to innovative behavior at all. Its influence ran entirely through AI Literacy, teaching Self-Efficacy and professional identity.

That changes what a session is for. A tool walkthrough addresses none of those three. A session where a teacher rewrites their own assignment, defends the change and hears a colleague agree it is better addresses all three at once, which is why the artifact cycle outperforms the demo even when the content is thinner.

The design implication is uncomfortable for anyone with a slide deck: you cannot lecture a teacher into seeing themselves as the kind of teacher who works this way. You can put them in a situation where they act that way once, with support, and let the evidence accumulate.

Put the hard parts in the template they use

Dogan's (2026) i-TPACK framework comes from a PRISMA review of AI professional development plus a search that surfaced 17 further studies. It aligns five knowledge domains (intelligent technological knowledge, technological content knowledge, technological pedagogical knowledge, integrated i-TPACK, and AI ethics) with four evidence-based pathways: Active Learning, models and examples, coaching and expert support, and Feedback and reflection, governed by five principles running from Collective Synergy Over Silos to Flexibility within Structure.

The framework's sharpest criticism of the field is that ethics is handled superficially or absent entirely in many existing programs, and the fix is placement. Verification, disclosure and risk control belong in the document participants use all week, not in the last 20 minutes of the last session. A template with a risk-control column gets filled in; a closing segment on responsible use gets nodded at.

The framework's limit is that it is conceptual, with no program-level outcome data. Use it as a design checklist rather than as evidence that the design works. The outcome evidence is in the studies above and below.

Spend the hours on assessment

If your program has room for one deep topic, choose assessment. In Talebzadeh's (2026) eight-hour program, four two-hour sessions with 163 teachers and pre-service teachers, total AI-PCK rose from 151.65 to 220.14 (d = 2.36, p < .001). The largest component effect was Assessment Rubrics (d = 2.19), which was also the weakest area at pretest. The gain was narrower in teaching method than in assessment.

Rubrics are also where the anxiety in the room can be put to work. A participant who is uneasy about students leaning on a model can write a rubric line that makes the reasoning visible, which converts a worry into a design decision with a date on it.

Budget past the single day. In that same program, pre-service teachers gained significantly more than experienced ones (p = .033), mostly in scenario-based task design and rubric design. The experienced teachers said eight hours was too short to shift deep-seated habits, and asked for long-term, mentor-based support.

Build cohorts, then differentiate inside them

Run one universal workshop and you will spend most of it serving the median participant, who does not exist. Sutedjo, Chowdhury and Liu (2026) found among 127 respondents that self-perceived pedagogical content knowledge was high (M = 4.70) and content knowledge near ceiling (M = 5.15), while technological pedagogical knowledge (M = 2.62), technological content knowledge (M = 2.75) and holistic TPACK (M = 2.55) were markedly lower. Content knowledge showed no significant correlation with any technology-integrated domain (r = .11-.15), so subject-matter mastery does not carry over.

Those three low domains correlated r = .81-.91 with each other, which plausibly means one capacity for teaching with these tools rather than three separate skills. Teach it as a shared foundation and you save two sessions.

Then split by experience. Bai and Hsieh found professional identity predicted innovative behavior nearly twice as strongly among infrequent AI users as among daily users, a borderline result but a useful pointer: identity work pays off most with people who barely touch the tools, while frequent users get more from interdisciplinary projects than from foundational guidance. Laidlaw (2026) argues for discipline-specific cohorts as the setting for exactly that conversation, on the grounds that a physicist and a historian are not solving the same problem.

Open with the question the skeptics are asking

Laidlaw's (2026) autoethnographic account makes the case that resistance is not a skills deficit. Faculty asking what the point of teaching is in a GenAI world are not asking how to use a tool safely; the question is ontological. She reads GenAI as a threshold concept (transformative, troublesome, irreversible, integrative, bounded) and concludes that anxiety, resistance and confusion are necessary parts of crossing it rather than faults to correct. Her recommendations invert the standard workshop: open with identity questions rather than demonstrations, run discipline-specific cohorts, allow different timelines, and treat principled non-adoption grounded in disciplinary values as legitimate. She also warns that communicating rules can slide into an "enforcement illusion" that substitutes compliance messaging for support.

The guilt is measurable, and it is not evenly distributed. Vassallo (2026) surveyed 109 academics and identified an "AI guilt complex": 35% worried AI use undermines their credibility and 26% reported feeling they are "cheating", with anticipatory guilt about credibility exceeding remorse after use, so the distress is socio-professional rather than private. Four profiles emerged, Comfortable Adopters (27%), Guilty Non-Users (29%), Cautious Users (28%) and Morally Distressed Avoiders (16%). One program is addressing four different problems, and the 29% who feel guilt while not using the tools need something other than a demonstration.

The size of the task is visible in the AAC&U/Elon survey of 1,057 faculty (Watson and Rainie, 2026): 68% said their schools have not prepared faculty to use generative AI for teaching and mentoring, with a similar share for scholarship. 26% of respondents do not use the tools at all, including 40% of arts and humanities faculty, and 82% name colleagues' resistance as a barrier to departmental adoption.

Fix the conditions the curriculum cannot fix

This is the finding to bring to a leadership meeting. In the 568-faculty survey behind the comparison above, Kazakhstani faculty reported higher baseline readiness than their Chinese counterparts on all seven dimensions, with the widest gaps in disciplinary transfer, prompt design and basic understanding. In hierarchical regression the Kazakhstan coefficient, clearly significant on its own, fell to non-significance once prior GenAI use, recent AI training, institutional support, perceived permission, multilingual resource access, policy clarity and risk sensitivity were entered. That is a reduction of about four fifths. Institutional support and perceived permission carried the most positive weight, and risk sensitivity was the only predictor working against readiness. More than twice as many Kazakhstani as Chinese faculty reported AI-related training in the previous six months.

So the levers available to a program lead are local and proximate: hands-on practice, explicit permission to experiment, materials in the languages faculty actually work in, visible backing from the unit. National policy framing does not carry readiness on its own. Where risk sensitivity bites, the answer is to give it somewhere to be exercised rather than to talk caution down, and rubric design and verification routines are that somewhere.

Two practical frictions belong in the same conversation. Continued-use intention was the highest-rated satisfaction dimension in Bi et al. (2026) while tool usability was the lowest, so some of what reads as reluctance is platform friction. And Aponte et al. (2026) locate the hard limit: AI cannot compensate for chronic overload, punitive accountability cultures, weak leadership, or the absence of protected time. They argue it contributes to teachers' socio-emotional development only as relational infrastructure (protected time for mentoring, peer dialogue, communities of practice) and propose relational densification as the standard, counting trust, continuity of support and co-regulation rather than contact hours. Two claims translate into program design: freed time gets reabsorbed as extra demand unless the institution formally requires it to be reinvested, and conversational agents can quietly substitute for the difficult human negotiations that produce growth. Their framework is a heuristic, not a tested causal claim.

If those conditions are yours, a better workshop will not fix them, and saying so is more useful than promising otherwise.

Report artifacts at a delay

The institutional scaffold worth aligning to is Crompton, Burke and Nickel (2026), whose design-based research across two macro cycles and 114 participants (faculty from 28 U.S. institutions and 11 countries, plus centers-for-teaching directors and accreditation coordinators) produced six standards spanning the whole academic role: Instructor, Coordinator, Leader, Researcher, Learner and Contributor. Existing frameworks and educator standards target K-12 teachers or address only teaching, so a center that maps its offer onto these can be mapped onto appraisal and accreditation, which is also the argument that survives a budget review.

Then be careful how you evaluate, because most of this literature measures self-reported impressions. Bi et al. measured self-reported readiness rather than enacted teaching, and its groups formed by voluntary sign-up rather than randomization, so the results are associations. Bai and Hsieh is cross-sectional, self-selected, drawn from eight universities in one country, and measured every construct from the same respondents at one time. Sutedjo and colleagues cannot establish direction of causation, which matters because their strongest claim, that technological knowledge acts as a gateway, is what a cross-sectional design cannot settle. Talebzadeh reports very large effects without randomization, follow-up or discipline-level breakdown, so d = 2.36 reads as short-term. Laidlaw is autoethnographic, and Aponte et al. state that their framework is conceptual.

A defensible evaluation therefore measures the course materials a participant revised, the rubrics written, the tasks designed. It tests at a delay, because the gap in Bi et al. appeared only at follow-up, where the conventional group had decayed and the structured group had not. It treats satisfaction and confidence as weak signals, since Pishtari and colleagues (2026) note that quality gains alone do not prove internalization: rising design quality can equally reflect delegating routine work or accepting AI-generated structure. And it adds at least one relational indicator, following Aponte et al., because improving artifacts while isolating participants is not what the program claimed to do.

What program leads get wrong

A better demo is a better workshop. Contact time was held constant at four 90-minute sessions in the Bi et al. comparison, and the arm that revised artifacts beat the arm that learned about tools on every readiness dimension, with only one of the two gains still visible at follow-up.

Confidence is the outcome that matters. It is the easiest thing to move and the least predictive. In the 13-teacher study, perceived effort fell while the prompting layer added no measured design quality, and end-of-day ratings would have registered both as success.

Ethics is a closing segment. Dogan's review found ethics handled superficially or absent in many offerings, and the difference between a template column and a final slide is whether anyone uses it in week three.

One program can serve the whole faculty. The 127-faculty survey describes content knowledge near ceiling beside holistic TPACK at 2.55, and Vassallo's four profiles include 29% who feel guilt while not using the tools. A single session treats four audiences as one.

The skeptics are the obstacle to manage. Laidlaw's account puts resistance at the level of identity, and Bai and Hsieh found identity predicting innovation nearly twice as strongly among infrequent users. The people you most want in the room are the ones a demonstration confirms in their suspicion.

An institutional mandate will carry the program. The Kazakhstan coefficient collapsed by about four fifths once institutional support and perceived permission entered the model, and those two carried the most positive weight. A mandate is not the same signal as permission.

What will come back at you

"We do not have the hours." Then spend them differently rather than adding them. Six hours arranged as an artifact cycle beat six hours of demonstrations on identical contact time, and a shorter format spent on assessment rubrics targets the weakest area in Talebzadeh's pretest data.

"Our faculty will not come, or the skeptics will derail it." Recruiting the skeptics into a demonstration confirms the suspicion that sent them there. Open with identity questions, run discipline-specific cohorts, allow different timelines, and say out loud that principled non-adoption is a legitimate position. That is Laidlaw's recommendation, and it takes the compliance argument off the table. Expect identity work to matter most with infrequent users, per Bai and Hsieh.

"We cannot make people act on it." You mostly cannot, and you can change what the institution signals. Perceived permission and institutional support carried the most positive weight after the national-background effect collapsed, so the useful moves are the cheap ones: state that experimentation is sanctioned, protect the time formally so it is not reabsorbed, provide materials in faculty's working languages, and remove platform friction, which was the lowest-rated dimension in that same study.

"Our exit survey is strong." Satisfaction is the measure most likely to rise while nothing else does. Ask instead for the artifacts from last cohort and re-contact those participants a term later; the Bi et al. gap existed only at follow-up.

What to do this week

1. Rewrite the invitation so attendance requires bringing one course artifact: an objective, a rubric or an assessment task.

2. Rewrite the agenda so the first session starts on that artifact, and add verification, disclosure and risk control as columns in the template participants carry all week.

3. Move one session to rubric and assessment design, and have participants leave with the rubric written.

4. Get one sentence in writing from the dean or provost that experimentation is sanctioned, then protect the hours formally so they are not reabsorbed.

5. Set the evaluation to collect revised artifacts and to contact participants again a term later, with one question about who they now work with.

Where to go next

What Competencies Do Faculty Need in Regard to AI? covers the competency target itself, and Teacher AI Competency and Faculty Development carry the framework material behind it. For structuring AI into a course rather than into a workshop, see How Should AI Be Designed Into the Learning Experience?; for the wider institutional picture, What Are the Top 10 Findings from AI in Education Research That Instructors Should Know About? and How Do We Write and Implement an Institutional AI Policy?.