AECT 2026 · Emerging Learning Technologies

Don’t Write It
For Me

Moving AI from writing learning objectives to judging them, and what that exposed.

Hongming (Chip) Li¹Shan Zhang¹Seiyon M. Lee¹Mark Trexler²Anthony F. Botelho¹

1 VIABLE Lab, University of Florida

2 Johns Hopkins University

Scan to follow along

The slides, the demo, and the data

Start with a draft

AI can write a strong objective in seconds.

A teacher candidate writes

“Students will understand how to compare fractions.”

AI rewrites it

“Given pairs of fractions with unlike denominators, compare them using benchmarks and justify the result with a visual model.”

The objective got better. Did the writer?

Last year at AECT

LOGEN AI: the AI suggests, the educator decides.

Built on our ARCHED framework. AI generates, refines, and reviews. The educator keeps the call.

For someone with a trained eye, that’s real help, with judgment left where it belongs.

The catch

That design assumes the person in the loop can already judge.

Experienced educatorHas the judgment

Reads the draft, spots what’s off, keeps what fits. AI saves time.

but

Teacher candidateStill building it

Can’t yet tell strong from polished. AI’s call becomes theirs.

A human in the loop protects judgment only when the human already has it.

How can AI build judgment without replacing it?

Same judge. Different policy.

LOGEN StudioClass-Ready

AI writingGenerates and refinesNever in class work

ReviewRewrites, one click to applyNames the gap, asks one question

EvidenceThe final objectivesThe process behind them

Keep the judge. Take away the pen.

The learning loop

The learner judges first. AI asks second.

01Write

In their own words, with no AI.

02Check yourself

Level and S·M·A·R·T, before any feedback.

03Read feedback

A gap and one question.

04Revise

And say what changed, then go again.

05Submit

With an explanation and a receipt.

Then the instructor reads the process and grades. Every grade is the instructor’s.

Since our proposal

Three rounds with experts changed what we built.

In the proposalLive today

A stand-alone prototypeA mode inside LOGEN, same judge

AI suggests revisionsAI names the gap and asks

Submit, then AI evaluatesThe learner judges first, every round

Learner and AI onlyThe instructor sees the process and grades

Try it with me

Scan this, or open the link.

logen.viablelab.org/class-ready

Choose Try the open demo class
Sign in with Google
Write one objective, then get feedback

[5:30] Scan this or type the link. Click the link on the slide to open it in a new tab. Give people 30 seconds.

On screen, as a learner: Try the open demo class. Read the join screen aloud: “You write, AI asks.” Open the assignment and show the brief and the five-step tracker on top.

Type “Students will understand fractions.” The hint catches “understand” while you type, and that’s not AI. Save and mark it ready. Click to get feedback: the self-check appears first. Pick a level and tick letters, and say why this comes first. Get feedback, and show your level beside the review’s, the ringed letters, and “Ask yourself.” Point out there’s nothing to copy. Open Revise, change it, and go one more round if time allows.

Then: show Classes in the left sidebar. Anyone can create a class, set an assignment, and share a code, or join someone else’s. Show Try it yourself. The next two slides cover the rest.

If the wifi fails, the next slide has the same screens.

What a learner goes through

A hint, a self-check, then a question.

Objective being typed with a hint about the word understand — Local hints as you write. No AI.

Self-check dialog — Judge it yourself first.

Feedback card with Ask yourself — Your call beside the review’s, then a question.

The instructor’s side

See the process, not only the product.

Each objective from first version to latest, with every round’s judgment beside it.

An AI read can summarize a record and suggest levels. Nothing counts until the instructor saves it.

One objective from first version to latest as a word diff, with each round’s letters and question

Before a larger classroom study

We ran the whole workflow with three recruited participants.

3

Participants

A teacher candidate, a university designer, and a designer who is also a PhD student.

33

Tasks, end to end

Think-aloud, ratings, and an interview. 25 succeeded, and none failed.

26

Real AI reviews

A real browser and the real model. We checked the findings against code and data.

It shows what breaks. It does not establish learning outcomes.

What held

The AI asked. In 26 reviews, it never wrote.

Round 2

“Students will use the symbols >, =, and < to show which fraction is bigger.”

Ask yourself

“What evidence would show that learners can use all three symbols accurately, including when the fractions are equal?”

Round 3

“…on 10 fraction problems, some of them equal, and get at least 8 right.”

That’s the first time anybody explained why ‘do a worksheet’ isn’t an objective, and it did it by asking me a question.Jordan, pre-service teacher

Then we looked at the numbers

The writing improved.
The score crashed.

Round 2 The participant replaced “understand” and “do a worksheet.” The header said: your checks agreed 13%.

That 13% was the one moment I wanted to quit, and it wasn’t even measuring my writing.Jordan, pre-service teacher

Agreement with the review, the number learners seeSpecific and Measurable the review granted

Agreement and criteria granted by round
Round	Agreement	Specific and Measurable granted
1	58%	0%
2	13%	25%
3	83%	63%

So what does it measure?

Three ticks and no new thinking move it from 13% to 63%.

#1

You Analyze · Review Apply

YouReviewSMART

#2

You Apply · Review Apply

YouReviewSMART

#3

You Understand · Review Apply

YouReviewSMART

#4

You Remember · Review Apply

YouReviewSMART

A, R and T were granted 77 of 78 times. The default rubric then pays for agreement: “Self-checks come to agree with the review.”

Second surprise

The AI read marked down the learner who kept thinking.

6.3AI suggested

9.7Participant rating, after reading the record

What was submitted“…They will get at least 4 out of 5 right.”

What the AI read said“#1–#3 lack clear success criteria.”

The record shows me a student who answered their own question, and the AI summary marks that student down for not having done it.Renee, instructional designer

Putting those together

We took AI out of the writing. It came back through the measurement.

Are we measuring a learner’s judgment, or how closely they agree with AI?

In the participant interviews

The participants trusted what they could trace.

Jordan

Teacher candidate · SUS 62.5I finally get what “measurable” means… but some of the numbers didn’t help.

Next term: would recommend it for a class that assigns it.

Renee

Designer · SUS 75…the first dashboard I’ve reviewed this year with an action behind it…

Next term: would pilot again, not ready for 40 faculty.

Min-seo

Designer and PhD student · SUS 70The log is good data and the tiles are bad labels; I’d trust what I can trace…

Next term: maybe, low stakes. Not for grading yet.

What we take from this

Three commitments, and what each one changes.

Scaffold thinking, not answers

Keep the question on screen while the learner revises.

formative assessment

Make evidence verifiable

Tell AI reads what changed, and put the record beside every claim.

feedback literacy

Keep disagreement meaningful

“I disagree because…” earns credit. Agreement stops being the goal.

self-regulated learning

What comes next

Larger classroom studies, then beyond objectives.

NowFix

AI reads, the visible rubric, and the agreement figure.

NextStudy

Does judgment grow, and does it transfer to work without the tool?

ThenExtend

Assessment and lesson plans, once this one skill works.

Thank you

Come try it.

The goal isn’t a better AI judge of students. It’s students who judge their own work better.

Designing instruction? LOGEN Studio supports your judgment.

Learning to? Class-Ready helps you build it.

Hongming (Chip) Li · hli3@ufl.eduAnthony F. Botelho · abotelho@coe.ufl.edu

Try Class-Ready

This talk

Don’t Write ItFor Me

AI can write a strong objective in seconds.

LOGEN AI: the AI suggests, the educator decides.

That design assumes the person in the loop can already judge.

Same judge. Different policy.

The learner judges first. AI asks second.

Three rounds with experts changed what we built.

Scan this, or open the link.

A hint, a self-check, then a question.

See the process, not only the product.

We ran the whole workflow with three recruited participants.

Participants

Tasks, end to end

Real AI reviews

The AI asked. In 26 reviews, it never wrote.

The writing improved.The score crashed.

Three ticks and no new thinking move it from 13% to 63%.

The AI read marked down the learner who kept thinking.

We took AI out of the writing. It came back through the measurement.

The participants trusted what they could trace.

Jordan

Renee

Min-seo

Three commitments, and what each one changes.

Scaffold thinking, not answers

Make evidence verifiable

Keep disagreement meaningful

Larger classroom studies, then beyond objectives.

Come try it.

Don’t Write It
For Me

The writing improved.
The score crashed.