What students build
A web experiment that randomizes feedback framing and records retry, strategy-change, or help-seeking choices
Stanford · Stanford Medicine · Harvard public research, independently adapted by CIT
Questions about friendship
and attraction, tested with data.
Students test questions about friendship, attraction, and relationships with privacy-safe data and AI. They explain what the result supports and what it does not support.
- 10core research sessions
- 8optional final-project sessions
- 120 minprediction, experiment, build
- One-to-one or small grouponline and in person
Four live experiments
Change one setting and see whether the conclusion holds.
Every value below is fictional. Change the controls and check how the result and conclusion change.
Fictional learning task
Fictional puzzle PROCESS + STRATEGY FEEDBACK
One attempt at a difficult puzzle did not work. Check the feedback framing before choosing the next action.
Retry52%
Change strategy44%
Ask for help39%
Fictional classroom choice ratesExpected checks7.1
Difficulty estimating reply time0.46
This fictional program does not diagnose smartphone addiction.disappointment cues52%
pressure cues43%
AI gives no answer61%
The program shows cues in the text only. It cannot know the sender's actual feelings.Every session starts with a prediction, not a conclusion.
Students predict or touch a visual within 15 minutes and spend at least 45 percent of the session investigating or building.
- 01Question to check
- 02Prediction before results
- 03Short experiment
- 04Two possible explanations
- 05Break
- 06Build
- 07Check when the conclusion does not apply
- 08Research record
10 core sessions
10 sessions, one question answered with data each time
Open each session to see the question, main ideas, experiment, completed work, supported statement, and unsupported statement.
S01Can Growth-Mindset Feedback Change the Choice to Try Again?After the same difficult task, can feedback wording change whether students retry or change strategy?
Question to check
After one failed fictional puzzle, choose a next step after neutral feedback, ability praise, or process-and-strategy feedback.
Main ideasgrowth mindsetprocess-and-strategy feedbackrandom assignmentcontext and effect size
Completed work
A graph of retry and strategy-change rates by feedback condition, plus a result-and-limits note
Check when the conclusion does not apply
Check whether prior achievement, task difficulty, peer norms, wording, or condition order changed the result.
Supported statementIn this fictional classroom dataset, retry rates were higher in the process-and-strategy feedback condition.
Unsupported statementGrowth mindset raises every student's grades or changes every student's brain.
S02Why Does Believing We Are Watched Change Us?Can an audience cue change speed, accuracy, or confidence without real public exposure?
Question to check
Complete a short task in private, instructor-view, and fictional-audience conditions.
Main ideasbelieving that another person is watchingparticipants guessing the purpose of a studytask results and self-ratingslimits of conclusions from brain images
What students build
A safe web program that records response time, accuracy, confidence, and the order of conditions
Completed work
A comparison of three audience-cue conditions and two possible explanations
Check when the conclusion does not apply
Check whether practice, screen color, or task difficulty affected the result.
Supported statementMeasured performance changed under this fictional-audience condition.
Unsupported statementTeenage brains stop working normally when watched.
S03How Much Effort Would We Spend to See Feedback?Does expected feedback value change virtual effort and opt-out choices?
Question to check
Choose different click costs to unlock fictional peer feedback, points, or neutral information.
Main ideaschoosing what to measurewhat people say and how much effort they makecomparing age without fixed age groupsfirst looking for patterns and then testing a planned question
What students build
A record of public-data analysis and a web task using fictional feedback
Completed work
A graph from public data, a web task, a comparison of two methods, and a limitation
Check when the conclusion does not apply
Look for information that can separate curiosity, the game design, and the value of feedback as explanations for effort.
Supported statementIn this task, the number of clicks differed by the type of feedback participants expected.
Unsupported statementThis result reveals who needs approval.
S04Why Do Uncertain Replies Invite More Checking?Can uncertainty about reply timing matter more than overall reply probability?
Question to check
Run fictional messaging programs with different reply probabilities and waiting times.
Main ideasfixed and changing reply timesthe difference between an expected and actual replymaking a decision while the situation is uncertainthe cost of checking
What students build
A fictional messaging program where students can change reply probability, waiting time, notifications, and checking cost
Completed work
A graph of checks over time and a comparison of two checking methods
Check when the conclusion does not apply
Check whether the same result appears with points that have no social meaning.
Supported statementIn our fictional computer experiment, uncertainty about reply timing changed the checking method.
Unsupported statementDopamine makes students addicted to messages.
S05Should AI Use One Word for Mixed Feelings?When should AI give no answer for Korean text that people also understand differently?
Question to check
Annotate short fictional Korean messages with multiple cues before viewing any AI output.
Main ideasdistinguishing similar emotionsfinding more than one emotion in a sentencedifferences between people's answersgiving no answer when AI confidence is low
What students build
A Korean emotion-cue program that compares word rules, a simple AI method, and a more complex AI method, then gives no answer when confidence is low
Completed work
An error screen, an AI confidence table, and an explanation of when the program gives no answer
Check when the conclusion does not apply
Measure how one context line changes human and model judgments.
Supported statementThe text may contain cues of disappointment and pressure, but context is limited.
Unsupported statementAI discovered the sender's true emotion or mental health.
S06Does Everyone Else Really Think That?How far apart are personal beliefs, close-friend estimates, and general-peer estimates?
Question to check
Rate non-sensitive fictional situations from personal, close-friend, and general-peer perspectives without names.
Main ideaspersonal views and estimates of other people's viewsmistakenly assuming that everyone thinks the same waya sample that may not represent everyonea minimum group size that protects identity
What students build
An anonymous results screen showing the spread of answers, number of participants, uncertainty, and errors in estimates of peer views
Completed work
A graph comparing personal views with estimates of peer views and a caution about comparing Korean and U.S. data
Check when the conclusion does not apply
Test question order and socially desirable responding.
Supported statementIn these anonymous responses, personal averages differed from estimated peer averages.
Unsupported statementAll Korean teenagers are misled by peer pressure.
S07Do Similar Friends Choose Each Other or Become Similar?Can the same-looking friendship graph come from choosing similar friends or becoming similar later?
Question to check
Compare two moving friendship graphs with similar endings but different rules.
Main ideasshowing people and relationships as points and lineschoosing friends who are already similarfriends becoming more similar over timedifferent causes producing similar results
What students build
A fictional computer experiment that runs friend selection, friend influence, both conditions, and random connections from several starting points
Completed work
A moving friendship graph, numbers showing how groups formed, and a report comparing settings
Check when the conclusion does not apply
Change the starting relationships and random starting value to see whether the same conclusion appears.
Supported statementUnder these settings, friend selection and friend influence both produced similar group patterns.
Unsupported statementThe final friendship graph reveals why the relationships formed.
S08Can Screen Design Change Whom We Choose?Do information order, popularity scores, and rapid choice mechanics change fictional team selection?
Question to check
Choose fictional teammates on one screen that shows images first and another that shows interests, reliability, and communication first.
Main ideashow screen design affects choicesa fair comparison of two screensthe difference between recording and changing a preferencefair choices
What students build
A fictional profile experiment that records decision time, consistency, remembered information, and variety of choices
Completed work
A screen comparing the two designs and a redesign that does not pressure users
Check when the conclusion does not apply
Separate the effect of first-seen information from the choice mechanic.
Supported statementInterface conditions changed choice patterns and recall in this fictional-team task.
Unsupported statementAn app interface determines whom a student truly likes.
S09Do Fictional AI Participants Form Separate Groups, and Can We Trust the Result?How much can one instruction, memory setting, AI system, or starting condition change a fictional society?
Question to check
Run two fictional social experiments that differ by only one instruction sentence.
Main ideasgiving behavior rules to fictional participantsresults changing under the same settingsthe effect of AI instructionschecking whether the experiment represents the intended question
What students build
Compare fictional participants that follow fixed rules with AI participants that generate text across several starting conditions, instructions, and memory settings
Completed work
A comparison of friendship graphs, categories of failed results, a guide to the fictional experiment, and a cost record
Check when the conclusion does not apply
Check whether the fictional AI participants simply repeated assumptions that we supplied.
Supported statementUnder these instructions, settings, AI systems, and starting conditions, the fictional participants formed these groups.
Unsupported statementThe fictional AI society shows how real Korean teenagers form groups.
S10What Did Our Project Actually Show?Does the result show an observation, a relationship between two values, a prediction for new data, or a cause?
Question to check
Compare a 2003 study that reported shared brain regions for rejection and physical pain with a 2014 study that separated the two patterns, then mark what each result supports.
Main ideasseparating an observation from a causeerrors caused by inferring a feeling from a brain regionother possible explanationsletting another person check the same method
What students build
Use the open Neurosynth maps to calculate how strongly seeing a brain region supports identifying a feeling, then connect the data, comparison method, results, errors, and limits in one project package
Completed work
A probability notebook, GitHub project folder, research poster, three-minute explanation, and answers to review questions
Check when the conclusion does not apply
Check whether the same brain region appears in unrelated tasks, then name the result that would change the conclusion and the simplest other explanation.
Supported statementState only what the method and participant group support, and do not identify a person's feeling from a brain region alone.
Unsupported statementIt is already proven that rejection uses the same brain circuit as physical pain.
8-session extension
8 sessions that end in a poster and a talk
Choose one of three projects. Record the question, sources, data plan, simple comparison, evaluation, repeated tests, results, ethics, and limits.
Peer Feedback Data Study
Review public data and make a safe browser task with fictional feedback.
Mixed Feelings: Korean Emotion AI
Show several possible emotion cues, differences between people's answers, AI confidence, and cases where the AI gives no answer.
Unread Messages: Checking Behavior
Compare reply probability, wait time, checking cost, and two ways to estimate when a reply may arrive.
S11Choose a Final Research Question and Define What You Will MeasureWhich peer-feedback, mixed-emotion, or reply-uncertainty question can be measured responsibly?
Question to check
Use public response data to check whether thirty items that claim to measure one idea actually receive similar answers.
Main ideasclearly defining what to measurechecking whether items measure the same idea consistentlychecking how items form groupsa result that would show the prediction was wrong
What students build
A research-question worksheet, a table explaining the main terms, and an analysis of item consistency and grouping using public response data
Completed work
A research question, prediction, result that would change the prediction, project scope, item-consistency value, and factor structure
Check when the conclusion does not apply
Check that the question does not claim to reveal a real person's thoughts, health, or private relationships, and do not draw a conclusion from an unreliable set of items.
Supported statementIn this public sample, the items formed one group and the consistency value fell in this range.
Unsupported statementAI will reveal who likes whom.
S12Connect Each Claim to Its SourceHow do the participants, method, supported wording, and limits of a study shape a project claim?
Question to check
Rewrite a strong marketing claim so it matches a paper's method and number of participants.
Main ideasthe original research sourcethe participants and sample sizewhether a study can support a claim about causeconnecting claims to sources
What students build
A source comparison, reference list, and record connecting each claim to the sample size, method, supported wording, and limits
Completed work
Five core sources with sample sizes and a table of supported and unsupported wording
Check when the conclusion does not apply
Remove references that provide a university name but no useful method, and check that a small-sample result was not written as a general fact.
Supported statementEach major claim is linked to the study participants, sample size, method, supported wording, and limits.
Unsupported statementOur conclusion is correct because the research comes from a famous university.
S13Approve the Ethics and Data PlanCan the core question be tested without private personal data?
Question to check
Compare the risks and classroom value of private messages, professionally written fictional messages, public data, and fictional class data.
Main ideascollecting only the data that is neededgetting permission to use datapreventing answers from identifying a personchoosing the hypothesis and analysis method before seeing results
What students build
A data-source table, minimum group size that protects identity, a choice not to participate, an offline alternative, and a research plan saved before results are viewed
Completed work
A data and ethics plan reviewed by the instructor and a timestamped preregistration
Check when the conclusion does not apply
Explain why adding sensitive data for a small accuracy gain is risky, and check whether any part of the plan changed after the results were viewed.
Supported statementThe main question uses public data, professionally written fictional data, or fictional class data, and the analysis plan was saved before results were viewed.
Unsupported statementA friend's messages are safe to analyze after removing the name.
S14Build a Simple Comparison FirstWhat simple method can show whether advanced AI is actually needed?
Question to check
Review a small case where one simple rule nearly matches a more complex AI method.
Main ideascomparing with a simple methodreasons for using a more complex methodseparating practice data from evaluation dataletting another person run the same method
What students build
Build one rule or simple statistical method and keep practice data separate from evaluation data
Completed work
A working simple comparison, result table, and reason for using it
Check when the conclusion does not apply
Check whether evaluation answers were visible during practice and whether one person or conversation appears in both practice and evaluation data.
Supported statementA more complex AI method must show added value over the simple comparison on criteria chosen in advance.
Unsupported statementThe more complex model is automatically more scientific.
S15Run the Main Experiment and a More Complex AI MethodHow do we produce comparable experiment records instead of one good-looking result?
Question to check
Run the same AI method several times with different random starting values and inspect how much the results change.
Main ideasrecording settings and results for every runcomputer-selected random starting valuesthe AI version usedchoosing the question and evaluation method before seeing results
What students build
The final project's more complex AI method or changed-condition experiment with automatic run records
Completed work
An AI method or experiment, settings file, run record, and first result
Check when the conclusion does not apply
Check whether the question or evaluation method changed after viewing the results.
Supported statementAcross several runs with settings chosen in advance, the results fell within this range.
Unsupported statementThe best single run represents the model's ability.
S16Measure Results and Review ErrorsWhen average scores are the same, how do the failed cases differ?
Question to check
Compare two models with equal accuracy but very different failure types.
Main ideashow often a positive answer is correcthow many correct answers the method findscomparing AI confidence with actual resultsgrouping errors by type
What students build
A table comparing predicted and actual answers, an AI confidence graph, at least five failed cases, and a review of possible harm
Completed work
An evaluation report, error screen, and rule for giving no answer
Check when the conclusion does not apply
Check whether averages hide failures on short, mixed, or specific conditions.
Supported statementThis rule for giving no answer reduced one type of error but also withheld some correct answers.
Unsupported statementHigh accuracy makes the model trustworthy for every input.
S17Change the Settings and Build an Interactive ProductDo the conclusion and user experience remain similar when random starting values, data, AI instructions, or the no-answer rule change?
Question to check
Find a small setting change that alters the conclusion and turn it into a warning or control on the screen.
Main ideasdifferences caused by setting changesremoving one part for comparisona fictional comparison that changes one conditionletting users control how results are shown
What students build
A results table for changed settings and a web program that shows AI confidence, missing information, corrections, and no-answer results
Completed work
A report comparing settings, an interactive web program, and a guide to the AI or fictional experiment
Check when the conclusion does not apply
Find screens that hide uncertainty or invite users to read estimates as facts.
Supported statementThe result held within this range and changed outside it.
Unsupported statementIt worked once, so it is stable in real settings.
S18GitHub Project Folder, Poster, and Final PresentationWhat must be shared so another student can follow the same method and ask questions about the work?
Question to check
Review an incomplete GitHub project folder and find where another person would be unable to run the work.
Main ideasletting another person check the same methodguides to the data and AIlimits of the resultexplaining the method while answering questions
What students build
A GitHub project folder with a README, software setup, random starting values, sources, run commands, results, ethics, and limits
Completed work
A GitHub project folder, research poster, three-to-five-minute explanation, and answers to review questions
Check when the conclusion does not apply
Explain why the conclusion may change with another random starting value, participant sample, country, or age group.
Supported statementAnother person can check a result within the expected range using the same data and settings.
Unsupported statementPublishing the GitHub project folder makes the finding true for every person and situation.
Three options based on readiness
A starting option chosen subject by subject
A student can use the Researcher option for psychology, the Builder option for Python, and the Explorer option for statistics. The option can change as the student learns.
Change one condition on a guided screen and finish a graph and limitation statement.
Modify a simple comparison, an AI method, or a computer experiment with Python and JavaScript.
Read original papers, repeat tests after changing settings, and complete a poster and presentation.
Personal relevance without exposure
Relationship questions without personal data
Students can study friendship, attraction, and relationships without naming anyone, uploading messages, sharing histories, or accepting diagnosis.
- Prefer fictional class data, public research data, and professionally written fictional scenarios.
- Every romantic activity has a friendship, study-group, club, or project-team alternative.
- Do not build crush detectors, popularity rankings, attachment or mental-health diagnosis, or lie detection.
- AI may describe cues in text and missing information. It may not claim to know a person's actual feelings or thoughts.
Research sources and limits
So no claim outruns its evidence: who was studied, the method, the limits.
Results from U.S. or adult participants are not presented as facts about Korean teenagers. Results from fictional computer experiments are not treated as evidence about real people.
Harvard University
Peer feedback effort study
- Participants and limits
- 102 participants ages 12-23
- What this course checks
- Review public data and build a safe task with fictional feedback
Harvard University
Study of changes when participants believe a peer is watching
- Participants and limits
- Children, adolescents, and adults in a laboratory study
- What this course checks
- Design a safe audience-cue task and explain what brain images cannot prove
Harvard University
Study of changes in how people distinguish emotions by age
- Participants and limits
- 143 participants ages 5-25
- What this course checks
- Study mixed emotions, differences between people's answers, and the limits of text AI
Harvard University
Study of reward estimates in uncertain situations
- Participants and limits
- Laboratory mice and computer models
- What this course checks
- Build a reply-checking experiment without claiming that dopamine explains teen behavior
Stanford · UT Austin
National growth-mindset experiment
- Participants and limits
- 12,490 ninth-grade students in 65 U.S. public high schools
- What this course checks
- Test whether process-and-strategy feedback changes retry choices while treating effects as modest and context-dependent
Stanford Medicine
Individual brain dynamics in children
- Participants and limits
- More than 4,000 children in the Adolescent Brain Cognitive Development study
- What this course checks
- Explain why group-level brain findings cannot diagnose or predict an individual student
Harvard University
Mina Cikara research overview
- Participants and limits
- Laboratory studies of people and groups
- What this course checks
- Study groups, separate friend groups, fictional friendship experiments, and limits on conclusions
Harvard SEAS
Study of how a profile screen can change choices
- Participants and limits
- Adults using a fictional dating website
- What this course checks
- Adapt the method to fictional study-partner, club, friendship, or adult profile choices
Harvard Graduate School of Education
Youth Voice Playbook
- Participants and limits
- Guidance for research with young people
- What this course checks
- Ask students what they want to study and use their feedback to improve the course
Harvard Graduate School of Education
Report on healthy relationships and ethical guidance
- Participants and limits
- U.S. high-school students and young adults
- What this course checks
- Study beliefs about peer expectations, care, honesty, consent, and differences between countries
Boston Children's Hospital
Survey of how teenagers view relationships and social media
- Participants and limits
- 1,124 U.S. teenagers ages 16-18 in 2026
- What this course checks
- Use the findings only to form questions, then check those questions with Korean students
Seoul National University
KOTE Korean text emotion data
- Participants and limits
- 50,000 public Korean comments reviewed by five people each
- What this course checks
- Compare people's answers with text AI, check confidence, and let AI give no answer
Harvard and 4 partner institutions
Human Connectome Project in Development
- Participants and limits
- A planned sample of 1,350 participants ages 5-21
- What this course checks
- Optional advanced work on age patterns and the limits of brain-image measurements
Google DeepMind
Concordia fictional social experiment software
- Participants and limits
- Fictional AI participants, not real people
- What this course checks
- Check how instructions and starting conditions change a fictional AI experiment
UCLA
Does Rejection Hurt? An fMRI Study of Social Exclusion
- Participants and limits
- A small adult sample in a single fMRI study
- What this course checks
- The original overlap claim, read before its rebuttals
University of Colorado Boulder
Separate Neural Representations for Physical Pain and Social Rejection
- Participants and limits
- 60 adults, plus a separate resting-state sample of 91
- What this course checks
- The rebuttal read against the 2003 overlap claim
Washington University in St. Louis
Reproducible Brain-Wide Association Studies Require Thousands of Individuals
- Participants and limits
- About 50,000 participants across public neuroimaging datasets
- What this course checks
- Why the claim registry carries a sample-size column
Open meta-analysis tool
Neurosynth Automated Meta-Analysis
- Participants and limits
- Coordinates pooled from thousands of published fMRI studies
- What this course checks
- Students compute the reverse-inference probability themselves
Open data project
Open-Source Psychometrics Project Raw Data
- Participants and limits
- Large self-selected public volunteer samples
- What this course checks
- Reliability and factor structure on public data, collecting nothing from classmates
Center for Open Science
OSF Registries Preregistration
- Participants and limits
- Not a study. A public timestamped record
- What this course checks
- The capstone files a timestamped plan before it measures anything
This is not an official, certified, or jointly developed course of Stanford University, Stanford Medicine, Harvard University, or any institution shown. CIT independently adapts public research, data, and methods for education.
What students and parents ask first
Is this an official Stanford or Harvard course?
No. This is not an official, certified, or jointly developed course of Stanford University, Stanford Medicine, Harvard University, or any institution shown. CIT independently adapts public research, data, and methods.
Do students need dating experience or private messages?
No. Every activity can use friendship, study-group, club, team, or fully fictional scenarios.
Can a Python beginner start?
Yes. The Explorer option begins with guided screens and data records. Students can do more independent work in the Builder or Researcher option when ready.
Is the course only online and one-to-one?
No. It supports in-person or instructor-led online delivery in both one-to-one and small-group formats.
What will students make?
Behavioral experiments, data records, computer experiments, web programs, GitHub project folders, guides that explain the AI or computer experiment, a research poster, and a presentation.
We turn one question into something a student can actually test.
We review the student's interests and current Python, mathematics, and statistics readiness to choose the Explorer, Builder, or Researcher option.