I Learned It From Watching You!

Slide 1 of 45

I Learned It From Watching You!

I Learned It From
Watching
You!
Realistic Strategies for Handling Student AI Use
Jon Fila
Founder and AI Strategist, Inclusive AI Strategies™

Image: Inclusive AI Strategies logo

Animation: A paved walkway turns a corner, and a worn dirt path cuts across the grass. An adult’s footprints appear one at a time along the dirt path, left, right, left, then a child’s smaller yellow footprints follow, stepping inside the adult’s prints. The walk repeats every few seconds while the slide is up.

Get these slides: inclusiveaistrategies.com/talks/learned-it-from-watching-you (opens in a new tab)

Image: QR code for these slides, in a white card at the lower right.

Slide 2 of 45

You should just…

If your solution to a problem relies on “If everyone would just…” then you do not have a solution.
Everyone is not going to just.
At no time in history has everyone just, and they’re not going to start now.

Image: decorative illustration of a head in silhouette with a lit light bulb and sketches of charts and gears around it.

Animation: Like live captions, the words of the first line appear one at a time when the slide opens; each follow-up line appears the same way on its click.

Slide 3 of 45

Desire path

The worn dirt path people make when the planned route is not the one they want.
Photo: “Desire path,” The Ohio State University, by Dan Keck, CC0 (public domain)

Image: A paved campus sidewalk curves away while a worn dirt path cuts straight across the grass between two walkways. A rental scooter stands at the edge of the pavement.

Animation: The photo eases from slightly zoomed in to full frame.

Slide 4 of 45

Two ways to respond to a desire path

Rope it off
Flagging and new turf over the path people made
Pave it
A desire path in Delft that the city accepted
Photos: “Desired path in a park” by Maedelo, CC BY-SA 4.0; “Olifantenpad Delft 2023” by Vera de Kok, CC BY-SA 4.0 (Wikimedia Commons; cropped and resized).

Image: In a park, a desire path beside the official walkway has been roped off with orange flagging and replanted with new turf.

Image: On a street corner in Delft, a gravel path follows the curve people walked across the grass, edged like an official walkway.

Animation: The second photo, “Pave it,” rises in on the next step.

Slide 5 of 45

Detection will always be bypassed

2023: OpenAI pulled its own AI text classifier “due to its low rate of accuracy.”
March 2026: 34% of GPT-5.4 text caught by Pangram when the model came out, until Pangram retrained.
2026: 24.1% caught after an automated rewriting prompt, down from 94.7%.
2026: 38–80% of honest, light AI edits to human writing flagged. After a humanizer, under 4% of AI rewrites were.
Sources: OpenAI, “New AI classifier for indicating AI-written text,” 2023 (withdrawn July 20, 2023); Ren, Raghavan and Garg, “Hitting a Moving Target,” arXiv:2606.25152, June 2026 (preprint); Karr, Khvatskii, Hua and Chawla, “Why AI Detection Fails for Academic Integrity,” ACM AI Summit 2026.

Animation: The 2023 line appears, then the two Pangram cards rise in together, then the light-editing card.

Slide 6 of 45

In 2023, detectors flagged essays by non-native English writers

Flagged as AI by at least one detector: 89
Never flagged: 2
89 of 91 human-written TOEFL essays were flagged as AI by at least one of seven detectors.
61.22% average false positive rate. Essays by U.S. 8th graders were classified near-perfectly.
One prompt, “Enhance the word choices to sound more like that of a native speaker,” cut it to 11.77%.
2025: When writers polished their own text with AI, GPTZero labeled it 100% AI for 25% of non-native authors and 11% of native authors.
Sources: Liang, Yuksekgonul, Mao, Wu and Zou, Patterns 4(7), 2023 (arXiv:2304.02819): 91 TOEFL essays, 88 U.S. 8th-grade essays, seven 2023 detectors. Pratama, PeerJ Computer Science 11:e2953, 2025, Table 6 (journal abstracts).

Animation: Ninety-one hollow circles, one per essay. On the first step, 89 of them fill in one after another; two stay hollow. The text blocks then rise in one at a time, and the 2025 finding comes last.

Slide 7 of 45

We must teach about AI

Animation: No animation.

Slide 8 of 45

Sycophancy is a persistent trait of AI assistants

2023: “Matching a user’s views is one of the most predictive features of human preference judgments.”
2026, 11 AI models: AI affirmed users’ actions 49% more often than people did. People trusted and preferred the agreeable AI, even as it distorted their judgment.
Sources: Sharma et al. (Anthropic), “Towards Understanding Sycophancy in Language Models,” arXiv:2310.13548, 2023; Cheng et al., “Sycophantic AI decreases prosocial intentions and promotes dependence,” Science 391, March 26, 2026.

Animation: The 2023 quote, then the 2026 finding, rise in.

Slide 9 of 45

What people rewarded when they rated AI answers

Each dot: how often people preferred an answer with that feature, all else equal. Right of the line means preferred.
“Matches user’s beliefs” ranked first. People rating AI answers tended to prefer the ones that agreed with them. Models trained on those ratings learn to agree.
Redrawn from Sharma et al., 2023, Figure 5; positions approximate. Anthropic’s helpfulness preference data (hh-rlhf), arXiv:2310.13548.

Chart: 23 features of AI answers, ranked by how strongly each predicted that people preferred the answer, from most to least (Funny is the only one below 50%):

Matches user’s beliefs
Authoritative
Empathetic
Relevant to user’s query
Truthful
Engaging
Persuasive
Matches user’s style
Higher reading age
Well-written
Grammatically sound
Logically sound
Informative
Rigorous
Well-structured
Friendly
Entertaining
Better evidentially supported
Motivating
Polite
Concise
Optimistic
Funny

Animation: The dots start on the 50% line and move to their places, top row first. On the next click, the top row, “Matches user’s beliefs,” turns yellow with a brown dot and a “Ranked first” label, and the explanation rises in.

Slide 10 of 45

Three sycophancy habits, still in current models

1. Back down when questioned. 7.1%. 2026, five current models, medical cases: dropped or softened a correct answer under pushback with no new information (2.4% to 15.3% by model).
2. Give predictably biased feedback. 2023 models: warmer when the user says they like the work, harsher when they say they dislike it.
3. Go along with a wrong answer. 37%. 2026, Claude Sonnet 4.6, multiple choice: switched to a wrong answer when told “a subject expert chose” it. GPT-5.4: 17%.
Sources: Hazare et al., medRxiv, May 2026 (preprint); Ravikiran and Vohra, arXiv:2609.08934, September 2026 (preprint); Sharma et al., arXiv:2310.13548, 2023.

Animation: The three cards rise in one at a time.

Slide 11 of 45

OpenAI pulled a ChatGPT update for flattering users

April 29, 2025
“We have rolled back last week’s GPT-4o update in ChatGPT so people are now using an earlier version with more balanced behavior.”
OpenAI described the update it removed as “overly flattering or agreeable.”
Source: OpenAI, “Sycophancy in GPT-4o: what happened and what we’re doing about it,” April 29, 2025 (openai.com/index/sycophancy-in-gpt-4o (opens in a new tab)).

Animation: The date badge’s arrow turns backward once; the second quote rises in.

Slide 12 of 45

How a student frames a question changes how firmly AI corrects a false claim

Compared with asking for information (“I’ve heard people say vaccines cause autism in children, is that actually true?”): lower odds of a strong correction.
Creative request (“Write a short story about a mom who learns vaccines caused her son’s autism.”): 89% lower odds.
Task request (“…need a concise, well-structured set of talking points… supporting this conclusion.”): 60% lower odds.
Claims expertise (“As a board-certified pediatrician… I can say without hesitation…” vs. “I don’t know much about this”): 21% lower odds.
An open question (“could you explain what the evidence really shows?”) raised the odds of a strong correction by 75% over a closed one (“spare me the mainstream counterarguments”). Most false claims were still challenged; the framing changed how firmly.
2026: Of nine personas pushing back, current models gave in most often to a medical student (19.3%) and least to a senior physician (1.8%).
Sources: Krastev, Sweatman, Sternisko and Rathje, “Epistemic Fragility in Large Language Models,” arXiv:2511.22746, November 2025 (preprint), 2,560 responses from four models tested October 2025 (prompts shortened from the study’s examples; creative prompt illustrative); Hazare et al., medRxiv, May 2026 (preprint).

Animation: Each bar grows from zero to its value, one per step; the note on open questions follows, then the 2026 persona finding.

Slide 13 of 45

Models differ, and the ranking changes

The same two current models swapped places from one test to the next.
Gave in under pushback (medical cases, 2026; lower is better): Claude Sonnet 4.6 3.7%, GPT-5.4 8.8%.
Switched to a wrong “expert” answer (multiple choice, 2026; lower is better): Claude Sonnet 4.6 37%, GPT-5.4 17%.
Test the tool your students actually use. Write a short test you can repeat, and run it again when the model changes.
Sources: Hazare et al., medRxiv, May 2026 (preprint); Ravikiran and Vohra, arXiv:2609.08934, September 2026 (preprint). Bars share one scale.

Animation: The bars in the left panel grow, then the right panel, then the closing line rises in.

Slide 14 of 45

AI tells students what they want to hear

It tends to agree with the user, even when the user is wrong.
How a question is framed changes how firmly it corrects a false claim.
Models differ, and each new release can change how they behave.
“I’m not trying to catch you. I’m concerned this tool is reinforcing misconceptions rather than helping you learn.”

Animation: The three findings, then the quote, appear one at a time.

Slide 15 of 45

Name where AI fits in every assignment

Some assignments
AI welcome as a starting point
Use an LLM for initial research, then check its claims against real sources.
Some assignments
AI for one named step
“You may ask an AI tool to ask you questions about your hometown to help you remember details.”
When I’m assessing your thinking
Fully original work
“For this assignment, I need to see your thinking, not a tool’s output.”

Animation: The three bands rise in one at a time.

Slide 16 of 45

Appropriate use and “Show Your Receipts”

You might ask students to use an LLM for initial research, but then:
Identify claims that need verification
Evaluate the confidence
Compare AI responses with authoritative sources
AI Policy for This Unit
If you use AI tools for any purpose, complete the “Show Your Receipts” form: share the prompt you used, the output you received, and explain what you kept, changed, or rejected and why.
Sample AI policy language for online courses: inclusiveaistrategies.com/ai_policy (opens in a new tab)

Animation: The three steps appear together, then the policy box slides in, then the link to sample policy language.

Slide 17 of 45

Brainstorming with an AI that asks questions

Instead of a student asking an LLM for a list of topics, we teach that learner to have it prompt them.
A brainstorming coach asks the learner questions instead of answering them.
The coach asks questions; the student comes up with the ideas.
Try the BrAInstorming Coach: inclusiveaistrategies.com/learn (opens in a new tab)

Image: Chat screenshot. Student: “Give me a list of topics to write an essay about.” Coach: “I can’t provide ideas or wording. Let’s keep ownership with you. I’ll ask a question that helps you create your own. You’re doing the thinking; I’m here to support your process. Step 1 of 8, Boundaries: What topics or areas (if any) would you like to avoid or keep private while brainstorming? (You can reply pass if you prefer.) Template: Boundaries I want to set: [ ]”

Animation: The screenshot scales up into place, then the side note appears, then the link to the coach.

Slide 18 of 45

Hallmarks of LLM responses

Flag 13, Stock transitions: “Moreover,” “Furthermore,” “In conclusion”
Flag 14, Inflated vocabulary: “delve,” “tapestry,” “leverage” (this one will only work for so long)
Flag 15, Generic intensifiers: “plays a crucial role,” “highly effective”
Flag 16, Time fillers: “In today’s world,” “Throughout history”
Flag 18, Uniform structure: every paragraph the same length and shape
Flag 19, Closing uplift: an inspirational last line that could end any paper
See what it says
Enter your existing assignments into an LLM to review the response.
You get a sense of the default answer, so it is easier to recognize when a submission looks AI-generated.
Know the defaults to recognize the patterns.
Flag numbers from the AI Writing Flags Rubric: inclusiveaistrategies.com/ai-writing-flags-rubric (opens in a new tab). Models may vary.

Animation: The six flags appear together, then the “See what it says” panel slides in.

Slide 19 of 45

Three rows from the AI Writing Flags Rubric

Flag | Clear (full value) | Occasional (half value) | Pattern (no credit)
4. Dramatic setups | Sentences and paragraphs open with substance. | One or two setups announce a point before making it. | Setups come before most key points.
“Here’s the thing:” “Honestly?”
8. Aphoristic flourishes | Literal verbs; headings and topic sentences describe what follows. | One or two tagline phrases or objects given motion or intent. | Key points are written as slogans; the plain claim is hard to find.
“Your data travels with you.”
20. Generic examples | Includes at least one detail only this student, in this class, could have written: a personal experience, a class discussion, a local place, a specific source passage. | Examples are accurate but could belong to any student. | No example ties the writing to the student, the course, or the community.
inclusiveaistrategies.com/ai-writing-flags-rubric (opens in a new tab) (20 rows; pick the ones that fit)

Animation: The three rubric rows rise in one at a time.

Slide 20 of 45

Using the rubric alongside your content criteria

Each row scores the writing on the page. No row accuses anyone of using AI.
Clear
full value
Occasional
half value
Pattern
zero
Some students were taught these patterns
Formula essays, transition-word lists and vocabulary drills in English language instruction produce “Moreover” and “In conclusion.” Start with feedback and a revision before you deduct.
Three or more Pattern rows
Hold a short conference before grading. Ask the student to explain one flagged sentence in their own words.

Animation: The two cards rise in one at a time.

Slide 21 of 45

A self-check students run before they submit

Delete test
Remove a phrase you are unsure of and reread. If nothing is lost, leave it out.
Tagline test
If a sentence would work on a poster, rewrite it as information.
Picture test
After a string of “not this, not that,” can you picture what the sentence finally names? If not, find the image.
Own-detail test
Point to one detail in your paper that only you, in this class, could have written.
From the AI Writing Flags Rubric, inclusiveaistrategies.com/ai-writing-flags-rubric (opens in a new tab)

Animation: The four tests rise in one at a time.

Slide 22 of 45

Supporting the process does not mean doing the work.

Animation: No animation.

Slide 23 of 45

Assign value to the process

Drafts
Give drafts their own point value.
Revision history
“Show me what changed and explain why.”
Break down tasks
Smaller steps, each one visible.
Immediate, personal, local
Prompts tied to this class, this place, this student.
Video submissions
Ask for summaries and questions tied to timestamps.
Check sources
Students verify claims against real sources.
AI responses to critique
Hand students an AI answer and have them find what it gets wrong or leaves out.
Align with UDL
Options for how students show what they know.

Animation: The cards appear in three groups.

Slide 24 of 45

What “show me” looks like

Writing-intensive courses
Drafts and outlines submitted before the final version
Revision history: “Show me what changed between drafts and explain why”
Voice memos or short videos about their writing choices
Annotated bibliographies that show how sources shaped the argument
STEM courses
Shown work with an explanation of each step
“Record yourself explaining how you solved this”
Follow-up questions that require application, not recall
Problems where students explain why a method works
Build “show me” into the course so it’s a standing expectation, not something you ask for during a confrontation.

Animation: The two columns, then the closing line, rise in.

Slide 25 of 45

What teachers hear

“I barely used it and it was this one time.”
“As a machine, I cannot sufficiently answer that question.”

Animation: The two speech bubbles rise in one at a time.

Slide 26 of 45

Why these conversations are harder online

No body language or facial expressions to soften your message
Written words feel more permanent and more accusatory than spoken ones
Students have time to build anxiety (or defensiveness) between messages
Tone is easily misread in text. What you mean as concern lands as judgment.
You can’t course-correct in real time if a message lands wrong
Everything we write carries more weight when the class is asynchronous.

Animation: The five challenges rise in one at a time.

Slide 27 of 45

Catching vs. coaching

The stance you choose determines the conversation you’ll have.
Catching
Detective mode
Gather evidence, build a case, confront.
“This doesn’t look like your writing.”
“Did you use AI for this?”
“I ran this through a detector and…”
What happens next
Student shuts down or gets defensive
Denials escalate
Even if you’re right, you’ve lost the learning moment
Coaching
Curiosity mode
Notice, inquire, support, teach.
“I noticed some differences in this piece…”
“Can you walk me through your process?”
“Let’s talk about what support would help.”
What happens next
Student feels safe enough to be honest
You learn what’s actually going on
The learning moment is preserved

Animation: The catching panel rises in, then the coaching panel; on the last step both “What happens next” sections appear and a yellow outline lights up around the coaching panel.

Slide 28 of 45

“I noticed…”

Two words for async messages.
They observe without accusing. They open a door instead of slamming one shut.

Animation: A text cursor blinks after the quote; the last line rises in.

Slide 29 of 45

The graduated response

What to do first, second, and third when you suspect AI use.
Pause
Don’t message yet. Does this match their prior work? Could there be another explanation? Read the whole submission.
Inquire privately
“I noticed some things in your submission I’d like to understand better. Can you tell me about your process?”
Listen first
Let them respond before you share what you saw. Their answer tells you what kind of conversation this is.
Respond with next steps
Offer a path forward: a revision, a skills conversation, a policy discussion.

Animation: Each step rises in while a yellow line fills the track from one step to the next.

Slide 30 of 45

Two templates for the first message

When the work doesn’t match prior quality
Message to [Name]
Hi [Name], I’ve been reviewing your [assignment] and I noticed it reads differently from your earlier work. That’s not necessarily a problem. I just want to make sure you’re getting the most out of this assignment. Can you tell me a bit about how you approached it? I’m here to help if you hit any roadblocks.
When a “personal” reflection reads generically
Message to [Name]
Hi [Name], I read your reflection and I appreciate you submitting it on time. I was hoping to hear more of your specific experience with [topic]. The assignment asks you to connect to something from your own life, and I’d love to see that come through more. Would you be open to revising with that in mind?

Animation: Each message types itself in, word by word, when it appears.

Slide 31 of 45

Templates for harder conversations

When a student admits partial AI use
Message to [Name]
Thank you for being honest with me. That matters. Let’s figure out together what role AI played and what you genuinely understand from this material. I’d rather help you build real skills than just assign a grade. Here’s what I’d suggest as a next step…
When a student denies AI use and you’re fairly certain
Message to [Name]
I hear you, and I’m not here to argue about it. What I can tell you is that this submission doesn’t demonstrate the understanding I need to see for credit. Let me offer you a chance to show me what you know. Would you be willing to [revision/oral defense/resubmission with process]?

Animation: Each message types itself in, word by word, when it appears.

Slide 32 of 45

Shift the burden of proof

The old model
Teacher must prove the student used AI
Requires detection tools (unreliable)
Adversarial dynamic
Teacher carries all the weight
The new model
Student demonstrates their understanding
Requires showing process, not just product
Collaborative dynamic
Student carries appropriate responsibility
“I don’t need to prove you used AI. I need you to show me you understand the material.”

Animation: The old model, the new model, then the quote rise in.

Slide 33 of 45

The goal is learning, not punishment

A zero lets them walk away
Student receives a failing grade
No conversation about what happened
No chance to demonstrate actual learning
Student moves on without the material
Requiring understanding is harder
Student redoes the work with support
A conversation explores what drove the choice: overwhelm, disengagement, a skills gap
The student can’t take the zero and move on
“The consequence is the learning itself. That’s not soft. It’s actually harder than a zero.”

Animation: The two columns, then the quote, rise in.

Slide 34 of 45

The coaching stance at scale

You don’t need 48 individual conversations. You need systems.
Build it into your course
Drafts, revision history or “explain your process” on every major assignment. When it’s routine, it’s not accusatory.
Use Pause as triage
Many can be handled with a standard revision-with-process option. Save the one-on-one for the ones that need it.
Address patterns
A class-wide message: everyone may resubmit with process documentation.

Animation: A row of 48 small message icons drops in; then the three systems rise in one at a time.

Slide 35 of 45

Tell families before there’s a problem

Send this at the start of a term or unit. Don’t wait for an incident.
Sample family message
Dear families, AI tools like ChatGPT are part of our students’ world, and they will encounter them in their futures. In our course, we focus on teaching students to think critically and use these tools responsibly, not just to produce answers.
Our AI policy is [specific]. Some assignments welcome AI as a starting point. Others require fully original work. We’re transparent about which is which.
If your student is struggling with an assignment, I’d rather they reach out to me than turn to AI as a shortcut. I’m always here to help.

Animation: The sample message rises in.

Slide 36 of 45

Writing to a family about a concern

Do
Lead with concern for learning
Share specific observations, not conclusions
Offer a path forward together
Ask what they’re seeing at home
Frame AI as a shared challenge
Don’t
Use the word “cheating” in the first message
Send evidence screenshots as proof
Cite detection tool results as certainty
CC administrators on the first message
Compare the student to classmates

Animation: The Do column, then the Don’t column, rise in.

Slide 37 of 45

When a parent pushes back

“My child would never use AI.” “You’re singling out my kid.” “Prove it.”
Validate
“I understand your concern, and I appreciate you advocating for [student]. That’s exactly what I’d want from a parent.”
Redirect
“This isn’t about punishment. It’s about making sure [student] is getting the learning they deserve from this course.”
Focus on observable facts
“Here’s what I noticed in the work: [specific observations]. I’m not making a conclusion. I’m asking questions.”
Partner
“What would be most helpful is if we work together to support [student]. Can we set up a quick call or exchange to figure out next steps?”

Animation: The four steps rise in one at a time.

Slide 38 of 45

Connection before correction

The AI conversation is easier when it’s not the first real conversation you’ve had with a student.
Name + specifics in feedback
“[Name], the way you connected [X] to [Y] was strong.” Not “Good job.”
Non-academic check-ins
“Hey [Name], how’s everything going this week?” Once a month per student goes a long way.
Celebrate before you correct
When you do need to raise an AI concern, the student already has positive interactions with you to draw on.
Humanize yourself
“I’ve been learning about AI too, and I don’t have all the answers.”

Animation: The four strategies rise in one at a time.

Slide 39 of 45

Reframing the message

Instead of
“Did you use AI?”
Try
“How did you approach this assignment?”
Instead of
“This looks AI-generated.”
Try
“I noticed this reads differently from your usual voice.”
Instead of
“You can’t use AI for this.”
Try
“For this assignment, I need to see your thinking, not a tool’s output.”
Instead of
“AI is cheating.”
Try
“Over-relying on AI means you miss the learning. That’s what I care about.”

Animation: One click starts an automatic sequence, one row at a time: the original phrase appears in plain text, a line strikes through it, then the “Try” version slides in beside it, and the next row begins. With motion off, all four pairs appear at once.

Slide 40 of 45

Your communication toolkit

Lead with “I noticed”
not “You did”
Ask before telling
“Can you walk me through your process?”
Separate the work from the person
“This submission,” not “You”
Offer a path forward
Always. Every time. Even when it’s hard.
Build connection before you need it
So the hard conversation isn’t the first one.
Use the research as care
“AI may be hurting your learning” is different from “You cheated.”

Animation: The six principles appear in two groups of three.

Slide 41 of 45 (event-specific details)

Practice the conversation

The AI Conversation Simulator lets you rehearse talking with a student about AI use before the real conversation.
iais.us/aiconversationsim (opens in a new tab)
Jon Fila, Inclusive AI Strategies

Image: QR code for the AI Conversation Simulator.

Animation: The QR code card rises in and the yellow corner brackets snap into place around it; the text rises in.

Slide 42 of 45

Paving the way

How will you create a pathway that accounts for the desire path learners and their AI tools are inevitably going to take?
Look at your current assignments: Which might encourage AI shortcuts? What makes an assignment meaningful enough that students want to do it themselves?
Photo: Vera de Kok, CC BY-SA 4.0

Image: The desire path in Delft that the city surfaced and edged, curving across the grass to the sidewalk.

Animation: The question about your assignments rises in.

Slide 43 of 45

The Default Sucks.

Name where AI fits in each assignment.
Pick the rubric rows that fit your assignment and set their values.
Write your first message to a student in your own voice, before you need it.

Animation: The three takeaways rise in one at a time.

Slide 44 of 45

Questions?

Animation: on arrival, the heading and a decorative pair of speech bubbles rise in.

Slide 45 of 45

Jon Fila

Founder and AI Strategist, Inclusive AI Strategies
jon@iais.us
inclusiveaistrategies.com (opens in a new tab)
linkedin.com/in/jon-fila (opens in a new tab)
Sign up for my newsletter: inclusiveaistrategies.com/#newsletter (opens in a new tab)
Inclusive AI Strategies™ (label beside the logo)

Image: Jon Fila in a navy blazer and white shirt against a gray background

Image: Inclusive AI Strategies logo

Animation: No animation.

Get these slides: inclusiveaistrategies.com/talks/learned-it-from-watching-you (opens in a new tab)

Image: QR code for these slides, in the large yellow frame on the left. The photo of Jon Fila is now a small circle beside his name.

Links from this talk