What Happened When 56 Iowa Teachers Spent Three Weeks Learning AI

AI for High School Teachers · Summer 2026 · The AI Program, Department of Computer Science, Iowa State University

A public summary of the first offering of this course. Teacher names, individual grades and private reflections have been withheld; quotations are used with the wording intact but the author unnamed. Every figure below is an aggregate.

Where the cohort taught
Center PointEldoraTiffinWheatlandKeokukWaylandClarksvilleCedar RapidsLamoniMarionMarshalltownPerryColoLeGrandJackson JunctionRock RapidsWaterlooCalmarDenisonSpencerAlbiaCamancheNew LondonTruroGlenwoodOsceolaAnamosaDes MoinesWoodbineWest UnionHumboldtHamptonGrimesRiversideLake CityUrbandaleAlburnettCouncil BluffsIda GroveRuthvenPocahontas4433222222222Town the cohort taught inCircle size and number = teachers from that townSchematic state outline; towns at approximate centres. Teachers are placed by their school's town.

Sixty teachers from 41 Iowa towns, from Rock Rapids in the north-west corner to Keokuk at the southern tip. Circle size is the number of teachers from that town. This was a statewide cohort, not a metropolitan one.

Source: course registration, school town.

The short version

Fifty-six Iowa secondary teachers enrolled in a free, three-week, project-based course on artificial intelligence. It carried no licensure credit and no stipend. It asked for six substantial projects, finishing with a capstone: a three-lesson AI-integrated unit for their own classroom.

Twenty-seven earned the certificate. That is 48% of everyone who took a seat, and 59% of everyone who ever submitted a piece of work. Both numbers appear throughout this report, because quoting only the second one would flatter us.

Among the teachers who answered the survey at both ends of the course, self-rated confidence in using AI in the classroom rose one and a half points on a five-point scale, and the teachers who reported no prior coding experience gained exactly as much as the one advanced programmer.

What we did not expect: when asked what was most valuable, eleven of sixteen respondents named a thing they had built and intended to teach with, not a topic they had learned.

The cohort at a glance
27certificates earnedof 46 who started3Data Track Endorsementsall six data activities27/27capstones passedaverage 183 / 200+1.50confidence gainpaired, 5-point scale+69Net Promoter Scorezero detractors

Source: production database, read-only, 2026-08-11.

Outcomes across the 56 enrolled
48%of enrolledEarned the certificate27 · 48%Started, did not finish19 · 34%Never started10 · 18%

Counting only the teachers who ever submitted work, 59% finished. Both numbers are given throughout this report; neither is the flattering one on its own.

Source: production database, read-only, 2026-08-11.

Who came

This was a statewide cohort. Teachers came from 41 towns, and only nine of the sixty registrants taught in or beside the largest cities. Several are the only computer science teacher in their district.

They arrived with mixed backgrounds. A quarter of those who answered the entry survey had never written a line of code. One had never used an AI tool at all; six were already using them regularly. They taught Spanish, art, special education, social studies, science, business, library media and computer science.

What they said they wanted, before it started, clustered tightly: help teaching students to use AI responsibly, and enough understanding to stop feeling behind.

"I hope to learn what AI is and isn't, how to help my students understand the limitations of AI, and how to use it effectively as a teacher."
"I want to be able to use AI confidently as a valuable timesaver and exciting classroom tool for my students."

Their worries clustered just as tightly, and they turned out to be predictive. Time, and coding.

"My coding skills are not great. I have my CS endorsement but I relied a lot on my students to help me understand python. So, it is summer and I am not sure I can do any coding without my kids."
Towns that sent more than one teacher
01234Center Point42 earned the certificateEldora40 earned the certificateTiffin32 earned the certificateWheatland33 earned the certificateCedar Rapids22 earned the certificateClarksville20 earned the certificateColo21 earned the certificateKeokuk21 earned the certificateLamoni20 earned the certificateMarion22 earned the certificateMarshalltown20 earned the certificatePerry20 earned the certificateWayland20 earned the certificate

Source: registration form, school town.

AI use before the course
23pre-surveyNever used AI tools1 · 4%Tried a few times8 · 35%Use occasionally8 · 35%Use regularly6 · 26%

Source: pre and post course surveys.

Prior coding experience, at registration
024686None8A little7Comfortable2AdvancedTeachersteachers

A quarter of the pre-survey respondents had never written code. All three of them who answered both surveys gained two full confidence points.

Source: pre and post course surveys.

What they finished

The certificate required five core projects and a capstone, each at 70% or better: a custom AI assistant built for their own subject, a prompt library, a Teachable Machine model, a bias audit paired with a classroom AI-use policy, a redesigned writing assignment, and the capstone unit.

Two things stand out in the completion data. The first is that almost nobody who submitted a core project failed it — attrition happened between projects, not inside them. People ran out of time, not ability. The second is that every capstone submitted passed, with the distribution pressed against the top of the scale.

That last point deserves an honest caveat rather than a victory lap: grading was explicitly a coaching instrument. Good-faith work started at the top of the range, feedback led with strengths before numbered next steps, resubmission was unlimited, and the best score counted. High scores here mean "this teacher produced usable classroom material after revision", not "this teacher was ranked against peers".

Nobody finished quickly. In a course advertised as three weeks, the fastest completed run took eleven days and the slowest nearly nineteen.

From enrolment to certificate
Enrolled56100% of enrolledStarted4682% of enrolled−101+ core project3461% of enrolled−123+ core projects2952% of enrolled−5All 5 core projects2748% of enrolled−2Capstone submitted2748% of enrolledCertificate2748% of enrolled

Ten teachers never submitted anything. Of the 46 who started, 27 finished — and the steepest single drop is the first one, before any coursework was attempted.

Source: production database, read-only, 2026-08-11.

Each required item: who submitted, who passed
01020303434M4 AI Gem3131M6 Prompt library2929M10 Teachable Machine2827M12 Bias audit + policy2727Spotlight — Writing redesign2727M15 Capstone unitSubmittedPassed at 70% or betterteachers

Almost nobody who submitted a core project failed it. Attrition happened between items, not inside them — the course lost people to time, not to grading.

Source: production database, read-only, 2026-08-11.

Capstone results, as a share of the available points
051070–75%75–80%280–85%685–90%1290–95%795–100%teachers

The capstone is a three-lesson AI-integrated unit, marked against a published rubric with a 70% pass mark. Every capstone submitted passed, and the distribution sits hard against the top of the scale.

Source: production database, read-only, 2026-08-11.

Final grade distribution
01020303090-1001180-89270-79160-6920-59teachers

Among the 46 teachers who submitted work, 30 finished at 90% or above. The three lowest scores belong to accounts with one or two submissions.

Source: production database, read-only, 2026-08-11.

Days from first submission to finishing the certificate
051015310–12 days412–14214–161516–18318–20teachers

Nobody finished quickly. In a course advertised as three weeks, the fastest run took eleven days and the slowest nearly nineteen. This is the strongest argument in the data for either lengthening the course or stating the workload plainly up front.

Source: production database, read-only, 2026-08-11.

Did it work

The most defensible evidence is a matched comparison: ten teachers answered the identical confidence question, with identical wording and anchors, at both the start and the end.

The mean rose from 2.70 to 4.20 on a five-point scale, a gain of 1.50 points. Nine improved, one held steady, none declined (paired t-test p < 0.001, 95% confidence interval +0.99 to +2.01, Cohen's d 2.12).

Two details make this more interesting than the headline. The ten paired respondents started with a mean identical to the full entry sample, so they were not an unusually confident group. And all three teachers who reported no prior coding experience gained two full points, the same gain as the one advanced programmer. Whatever moved, it did not only move for the technically prepared.

The exit survey was similarly positive: satisfaction 4.63 out of 5, a Net Promoter Score of +69 with no detractors, and all sixteen respondents saying they would use what they learned this fall.

What these numbers cannot support

This section is here because a summary without it is advertising.

Sixteen of fifty-six teachers answered the exit survey, and every one of them earned a certificate. No exit data exists for the ten who never started or the nineteen who started and did not finish. This report documents the experience of the teachers who finished. It does not measure the effect of the course on Iowa teachers generally, and we would ask anyone citing it not to say that it does.

Measured against their final standing, exit respondents were statistically indistinguishable from certificate earners as a group, so they are a fair sample of finishers. Measured against the whole cohort they sit far above it. Both things are true.

Other limits worth stating plainly:

  • No control group. Some of the confidence gain reflects a summer of general AI exposure rather than this course.
  • Confidence is not competence. We report the confidence shift next to the project results, never folded together.
  • The survey ran inside the platform, immediately after the certificate was earned, with the instructor identifiable. That invites warm answers.
  • Two response scales were asymmetric, offering more positive options than negative ones. The perfect scores on those two items are partly a property of the instrument, not only of the experience.
Confidence using AI in the classroom, before and after
1122334455Before the courseAfterTeacher A+2Teacher B+2Teacher C+1Teacher D+2Teacher E0Teacher F+2Teacher G+1Teacher H+2Teacher I+2Teacher J+1mean 2.7 → 4.2

Each line is one teacher who answered the identical question at both ends of the course; names are withheld. Nine of ten improved, one held steady, none declined. The mean rose 1.50 points on a 5-point scale (paired t-test p < 0.001, 95% CI +0.99 to +2.01).

Source: pre and post course surveys.

The whole distribution moved, not just the average
051031 — not at all728134114145 — veryBefore (n=23)After (n=16)responses

Before the course, half the cohort sat at 1 or 2. Afterwards nobody did. Note the different sample sizes: these are all respondents at each end, not the matched pairs.

Source: pre and post course surveys.

Post-course responses (n = 16)
Overall satisfaction1411mean 4.63Confidence using AI1114mean 4.19Met expectations231111 exceededWill use it this fall31313 definitely12345 / highest

Read these with Section 3.3 in hand: two of the four scales are asymmetric, offering more positive options than negative, so the clean sweep at the top is partly a property of the instrument.

Source: pre and post course surveys.

Net Promoter Score: +69
Likelihood to recommend511mean 9.25 / 10Detractors (0–6)Passives (7–8)Promoters (9–10)

Eleven promoters, five passives, and no detractors among the sixteen respondents.

Source: pre and post course surveys.

Survey response, against a cohort of 56
0510152023Pre-course survey16Post-course survey10Answered bothResponsesteachers

Every one of the sixteen post-survey respondents earned a certificate. That is the central limitation of the evaluation: it documents the experience of finishers.

Source: pre and post course surveys.

Who answered the survey, measured by their grades
020406080100Post-survey respondents93.67%All certificate earners93.77%Everyone who submitted work89.96%Whole cohort (56)73.89%Non-respondents65.98%

Respondents are indistinguishable from certificate earners (a 0.1-point gap) but sit 28 points above non-respondents. They are a fair sample of finishers and a poor sample of the cohort.

Source: production database, read-only, 2026-08-11. Source: pre and post course surveys.

What the teachers said

The richer record is not the survey. It is 81 private weekly reflections written by 36 teachers during the course, seen only by the instructor. They are more candid than an exit survey ever is, and they trace the shape of the three weeks: 36 wrote in week one, 25 in week two, 20 in week three.

Two themes dominate, and they pull against each other. Building a custom AI assistant for their own subject was the standout win, named by 25 of the 36. Coding was the hard part, named by 19.

The single arc that best represents the course belongs to one teacher across two entries. In week one:

"Really don't like it. It is way too 'math' for me. … I worked very hard on the cocoa assignment — and received 0/20! So Frustrating."

In week three:

"I am so glad I took this course — even though it about killed me. I learned so much and my biggest takeaway is the three-day lesson I prepared that I will actually use in a month. … Don't get discouraged — I almost dropped out and am so glad I didn't. Great class."

Others, all unnamed and quoted as written:

"My single biggest takeaway from this course is integrating AI into a non-CS course is easier than I thought it might be!"
"It has helped me understand, at least to some extent, how artificial intelligence works, and that understanding it does not depend on whether you know how to program."
"Being an old dog teacher close to retirement this AI stuff is challenging for me, but I pushed through it."
"I don't want to sit on the sidelines and watch as AI transforms the way teaching and learning are done. This course gave me some practical ways I can start to do that, beginning this year."
"As much as I have been a proponent of AI, I had oversimplified its potential and primarily used it like a search engine without even realizing it. … Students are using AI with or without permission."

And the criticism, because a report that quotes only praise is not evidence:

"I rely heavily on AI which feels inauthentic and I feel like I'm not learning the stuff I'm supposed to."
"I am truly not sure I have taken anything from this course yet that I can use in my classroom."
"I am familiar with code but many teachers are not, so I believe seeing some of the code in the first few Modules would be very daunting and cause some teachers to 'drop out' before they even get to the best parts of how to use AI in the classroom."
Weekly reflections submitted
010203036W125W220W3teachers

36 in week one, 25 in week two, 20 in week three — the attrition curve, drawn by the teachers themselves.

Source: production database, read-only, 2026-08-11.

Themes across all 81 weekly reflections
0510152025Building a Gem was the standout win25 / 36Coding and Python were the hard part19 / 36Prompt engineering / RCTF16 / 36Falling behind on time15 / 36Built something for this fall15 / 36Responsible use and guardrails10 / 36Surprised by their own capability10 / 36The capstone was the payoff9 / 36Teachable Machine9 / 36Required vs optional confusion8 / 36District device blocks5 / 36The rubric-resubmit loop5 / 36Wanting access after the course4 / 36Instructor responsiveness, unprompted4 / 36

Counted by person, not by mention. This is the larger and more candid corpus: 36 authors against the survey's 16, written during the course rather than after it.

Source: production database, read-only, 2026-08-11.

Most valuable part of the course
0510Something they built and will teach with11 / 16Conceptual understanding of how AI works4 / 16Prompt engineering / the RCTF framework4 / 16Teachable Machine4 / 16The rubric-feedback-revise loop2 / 16Policy and state standards1 / 16

Coded from the free-text answers; a response can carry more than one theme.

Source: pre and post course surveys.

What is going into Iowa classrooms this fall

This is the part that matters more than any score. Thirteen teachers described specific plans, and they are concrete:

  • A sustainability unit in 9th grade Earth Science, built as the capstone
  • A week-three lesson in dual-credit Composition I
  • Reading and writing coaching tools for English 9 intervention students
  • A lesson sequence pairing Micro:bit hardware with machine learning
  • A chatbot conversation partner for world-language students
  • A rubric-feedback assistant to use in staff professional development
  • An entirely new high school elective on AI, created because of this course
"This fall, I plan to use what I learned to build and teach my AI course. Students will explore what AI is, how it learns from data, and how they can use AI tools to create projects of their own."

Three teachers also said they intend to teach what they learned to colleagues, which is the multiplier this kind of program is really after.

What they plan to do this fall
0246Deploy the capstone unit itself7 / 13A custom Gem as a student-facing tool7 / 13Teach AI literacy explicitly5 / 13Spread the practice to colleagues3 / 13Teacher productivity3 / 13Teachable Machine with students3 / 13

Source: pre and post course surveys.

What we are changing

The teachers were specific about what did not work, and we would rather publish it than not.

  1. Week one has to be gentler on the non-coder. Nineteen of 36 reflection authors named coding as the hard part, and one predicted outright that early code exposure would make teachers quit before reaching the useful material. The certificate never required coding. The opening modules failed to make that obvious.
  2. Required versus optional must be unmistakable. One teacher finished at the top of the class and still wrote that she was worried she had missed something. That is a design failure, not a reading failure.
  3. Three weeks is tight, and it collided with the start of the school year. Fifteen of 36 reported falling behind. Either the course lengthens or the real weekly hours get published up front.
  4. Say at the start that access continues. Four teachers asked unprompted whether they would lose the material. They do not, and saying so early costs nothing.
  5. Fewer places to check. Between the platform, the discussion board, the reflections and the chat channel, teachers had four surfaces to monitor.
  6. A managed-device path. Five teachers were blocked outright by district Chromebook restrictions on tools the course assumed they could reach.

One structural change is already decided: the weekly reflections drew more than twice the response of the exit survey, so the exit questions will move into the final reflection where teachers are already writing.

What teachers said to fix
012345Required vs optional was unclear5 / 16Early code is a drop-out risk4 / 16Front-loading and timing3 / 16Too many separate surfaces to check3 / 16Nothing to improve3 / 16Wanted video walkthroughs1 / 16Quiz questions not aligned to readings1 / 16Capstone expectations felt opaque1 / 16Python editor too small1 / 16

Source: pre and post course surveys.

How the course ran

The course was delivered on a platform built in-house for it rather than a commercial LMS, which is the reason this report can be as specific as it is.

Every project was scored against a published per-criterion rubric by a language model, with instructor review and override on any piece of work. The grading policy was explicitly coaching rather than gatekeeping. Teachers could resubmit without limit and the best score stood, which turned marking into a feedback loop instead of a verdict: the first real project absorbed the most revision, and scores climbed with it.

An AI tutor with a six-step hint ladder was available throughout and was used by 39 of the 56 teachers. Teachers worked in two distinct shifts, weekday mid-mornings and again between eight and ten at night, which is what summer professional development actually looks like for people with families.

The part worth reporting is the responsiveness. Every substantive complaint raised during the course was fixed during the course, usually within a day: a request-human-review control proposed by a teacher in her week-two reflection was live for the whole cohort the next day; an autosave was built the night a teacher lost work to a closed browser tab; a grading rule was changed so that a practice attempt could never lower an earned score. Teachers noticed, and four named the responsiveness unprompted.

"I truly appreciate the time, honest feedback, and communication put into the course. I have never felt so supported (especially in an online class)."
Submissions per day
020406080deadlineregrade + policy change7/207/278/38/10SubmissionsDistinct teachers activecount

Work bunched at both ends: a strong opening day, a mid-course trough, then a deadline surge. The final week carried more submissions than the first two combined.

Source: production database, read-only, 2026-08-11.

When teachers worked
0246810121416182022SunMonTueWedThuFriSat01,016 eventsHour of day, America/Chicago

Summer professional development happens in the morning and after supper. The heaviest hours are weekday mid-mornings, with a clear second shift between 8 and 10 at night.

Source: production database, read-only, 2026-08-11.

AI tutor use
0204060807/207/278/38/10Tutor messagesDistinct userscount

429 tutor messages from 39 of the 56 teachers. Use rose through the middle week, when the projects got hard.

Source: production database, read-only, 2026-08-11.

Method note

Figures come from the course's production database and its pre and post surveys, compiled at the close of the course in August 2026. Certificate eligibility was computed twice by independent methods, which agreed exactly.

Individual scores, names and the private reflections themselves are not published. Where a quotation could identify its author through detail rather than name, the detail has been generalised. The map shows the towns the cohort taught in, sized by how many teachers came from each; it does not indicate who completed the course.

Questions about the course, or interest in the next offering, are welcome at theaiprogram.org.

Published by The AI Program at Iowa State University's Department of Computer Science. Course designed and taught by Adisak Sukul, Ph.D.

This is a de-identified summary. Teacher names, individual grades and the private reflections themselves are not published.