dive

the research behind dive

Guess first. Then look. The claim is yours.

dive gives a student real public numbers and one question. They guess before they see any chart, read it, test it, and make their own claim. Then an AI partner offers a second opinion, and the student decides whether it holds. Here is each choice we made, beside the studies behind it and how strong that evidence is.

your guesswhat the data showed

How to read this page

Each decision has two halves. On the left, what dive does today. On the right, the research that led us there, with numbered sources at the foot of the page.

Evidence is not all the same weight, and we would rather say so than round up. Every decision carries one of five labels. After the decisions, one picture shows how they add up to a single loop, and then we say what the research does not settle.

  • strongmany studies agree, or a review that pooled many studies says so
  • moderateone careful study with a comparison group, or several smaller ones that agree
  • emergingearly or small studies, or one setting only; promising, not settled
  • frameworka published way of naming what to look for, which we borrow; not a measured result
  • our rulea choice we made for a reason we state, not a research claim

01Guess first, then look

One dive is one real question. The student commits to an answer before the evidence arrives, then checks it with their own hands.

01Lock a guess before any chart

in dive

Every dive opens with a question and nothing else. The student writes a guess and locks it, and only then does the chart appear. At the end they say whether the guess held, and a miss asks only what was different.

the research

In an experiment with 378 adults online, people who predicted the numbers on a chart before seeing them recalled the data more accurately than people who only looked1. In a study of physics demonstrations, students who only watched understood no better than students who never saw them, while those who predicted the outcome first understood more2. A guess you have committed to is what lets the data surprise you.

moderate

02Then test it on the same data

in dive

Each dive then opens the same numbers in the pool, dive's chart builder. The student redraws them, sorts them or shuffles them, and writes down what they did to check their guess.

the research

A review that pooled 225 studies of college science and math courses found higher exam scores where students worked during class, and students in lecture-only classes were 1.5 times more likely to fail3. The national guidelines for teaching statistics in schools ask for the same habit: students analyze the data themselves, then say what it means4.

strong

03One sentence: I think, because, and I checked it by

in dive

The student's answer is one sentence in three parts: what they think, what on the chart says so, and how they checked. Once it is in, it stays as written, and the numbers stay hidden until then.

the research

In an eight-week science unit, 331 seventh graders learned to write an explanation as a claim, the evidence for it, and the reasoning that joins them, and they improved on all three5. Those whose written help was withdrawn bit by bit reasoned better once it was gone. dive's sentence has the same shape, plus how you checked.

moderate

02The partner

An AI partner sits beside every dive. The research here is young, and its clearest lesson is that design decides: an AI that answers can hurt learning, and one that guides need not.

04A second opinion to accept, revise or reject

in dive

After the student's own claim is in, the partner offers its claim about the same chart. Some of these claims hold and some do not, by design, and dive says so. The student accepts, revises or rejects it, and must say what on the chart decides it.

the research

In an experiment with 199 people, designs that made them stop and think before taking an AI's suggestion cut their trust in wrong ones6. In a small 2025 study, not yet reviewed by other researchers, 36 high school seniors took ChatGPT's wrong advice on math puzzles 52.1% of the time7. Judging flawed answers is an older idea: in two algebra experiments, explaining solved problems that contained mistakes appeared especially good for understanding8.

emerging

05The partner points. The student reads.

in dive

A student can ask the partner at any step. It asks what they see first, points at one thing worth a second look, and never writes their claim. It is not told the answer until the student has made theirs, and its panel always says it is an AI and can be wrong.

the research

In an experiment in real classrooms with nearly 1,000 high school math students, an AI chatbot with no guardrails raised practice scores by 48%. Once it was taken away, those students scored 17% below students who never had it, and a version built to guide rather than answer largely removed the harm9. dive's partner is built the second way.

moderate

06A check with the partner resting

in dive

When their advisor opens a check, the partner rests. The student answers two short questions alone, each taking the big idea of a dive they finished to something new: one call to make, one claim to judge.

the research

In the classroom experiment above, students who scored higher with the chatbot open did worse than students who never had it once it was gone9. So dive does not treat work done beside the partner as proof. It asks again, without it.

moderate

03Real data, honest charts

The numbers are real and public. The dives climb four rungs: what's there, what's reported, what's related, and what isn't there.

07Real numbers, one question

in dive

Every dive is a real set of public numbers and one question a person would actually ask: which part of New York has the most rats, what a school job pays, whether a pricier college pays off.

the research

The national guidelines for teaching statistics ask teachers to start from questions that interest students4. The measured evidence is thinner than the advice: in one small study, 38 first-year university students said in interviews that real-life data made classes more interesting and easier to remember10. That is what students said, not a tested gain.

emerging

08Ask who counted

in dive

Every file can be downloaded, and its last lines say who published it and when it was fetched. A whole rung of dives asks who counted, and how: a chart of rat complaints shows where people called, not where the rats are.

the research

Five national associations, for math, science, statistics, social studies and computer science, said together in 2024 that students must learn to question where data comes from11. When 3,446 high school students looked into a climate website, 96% never learned of its ties to the fossil fuel industry12. Six 50-minute lessons on checking sources helped: the 271 students who had them got significantly better at judging what to trust online, compared with 228 in regular classes13.

moderate

09Same numbers, different picture

in dive

Many dives turn on a way a chart can mislead while every number in it is true, like an axis that does not start at zero or a blank on a map that only means nobody measured. In the pool the student redraws the same numbers and watches the story change.

the research

Students in grades 6 to 12 misread bar charts and histograms in the same ways college students do14. A warning is not enough: across five studies, bars on an axis that did not start at zero made differences look bigger to 83.5% of people, and the effect held even after they were taught about it15. So the student moves the axis and sees the bars change while the numbers stay put.

moderate

04The record

What dive keeps about a student, and who sees it.

10The advisor sees the work. The record gets counts.

in dive

A student signs in with the school's own student ID, and dive keeps their work and their questions to the partner so their advisor at school can help. What goes on to the school's record of how students work with AI is counts only: never their words, never a grade. The details, and how long anything is kept, are in the privacy policy.

the research

This one is not a research finding. It is a rule: keep what an advisor needs in order to help, pass on only counts, and say plainly what that is.

our rule

Ten decisions, one loop.

  1. from 01, 07

    Your guess

    Read the question. Lock your guess before you see any chart.

  2. from 08

    Look

    Read the chart. Ask who counted, and how.

  3. from 02, 09

    Try it

    Open the same numbers. Redraw them, and watch the story change.

  4. from 03

    Say it

    One sentence: I think, because, and I checked it by.

  5. from 04, 05

    Second opinion

    The partner makes its own claim. Accept it, revise it or reject it, and say why.

  6. from 01

    What you found

    Now the numbers. Did your guess hold? If not, what was different?

  7. from 06, 10

    The check

    Later, two questions on your own, with the partner resting. Only counts go on your record.

What the research doesn't settle

The chart study on guessing first was with adults online1, and the big review of learning by doing covered college courses3. Neither tested high schoolers reading charts.

No one has tested a second opinion like dive's in a classroom yet. The nearest study had 36 students, and it warns that teaching doubt can overshoot: after a short lesson on AI's limits, its students turned down more good advice7.

Flawed examples do not help everyone equally. In one pair of experiments, mixing wrong solved examples with right ones helped learners who already knew the subject, and those who knew less did better with correct ones only16.

The lessons on checking sources were about websites, not charts13, and the case for real data rests on expert advice more than on measured gains10.

Sources

Every source below was opened and checked against its publisher's, Crossref's, a library's or arXiv's record on 21 November 2026. The arXiv paper is a preprint and not yet peer reviewed. Where a study is small, early or from a college rather than a high school, the text above says so.

  1. Kim, Y.-S., Reinecke, K., and Hullman, J. (2017). Explaining the gap: Visualizing one's predictions improves recall and comprehension of data. Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems (CHI '17), ACM, 1375-1386. doi.org/10.1145/3025453.3025592
  2. Crouch, C., Fagen, A. P., Callan, J. P., and Mazur, E. (2004). Classroom demonstrations: Learning tools or entertainment?. American Journal of Physics, 72(6), 835-838. doi.org/10.1119/1.1707018
  3. Freeman, S., Eddy, S. L., McDonough, M., Smith, M. K., Okoroafor, N., Jordt, H., and Wenderoth, M. P. (2014). Active learning increases student performance in science, engineering, and mathematics. Proceedings of the National Academy of Sciences, 111(23), 8410-8415. doi.org/10.1073/pnas.1319030111
  4. Bargagliotti, A., Franklin, C., Arnold, P., Gould, R., Johnson, S., Perez, L., and Spangler, D. A. (2020). Pre-K-12 Guidelines for Assessment and Instruction in Statistics Education II (GAISE II): A Framework for Statistics and Data Science Education. Alexandria, VA: American Statistical Association. An official position of the National Council of Teachers of Mathematics. the publisher's page, with the full report
  5. McNeill, K. L., Lizotte, D. J., Krajcik, J., and Marx, R. W. (2006). Supporting students' construction of scientific explanations by fading scaffolds in instructional materials. Journal of the Learning Sciences, 15(2), 153-191. doi.org/10.1207/s15327809jls1502_1
  6. Buçinca, Z., Malaya, M. B., and Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 1-21. doi.org/10.1145/3449287
  7. Puppart, B., and Aru, J. (2025). A brief AI literacy intervention does not significantly reduce over-reliance and increases under-reliance on ChatGPT: A randomized study. arXiv preprint 2503.10556, first posted March 2025 and revised September 2026. arxiv.org/abs/2503.10556
  8. Booth, J. L., Lange, K. E., Koedinger, K. R., and Newton, K. J. (2013). Using example problems to improve student learning in algebra: Differentiating between correct and incorrect examples. Learning and Instruction, 25, 24-34. doi.org/10.1016/j.learninstruc.2012.11.002
  9. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122
  10. Neumann, D. L., Hood, M., and Neumann, M. M. (2013). Using real-life data when teaching statistics: Student perceptions of this strategy in an introductory statistics course. Statistics Education Research Journal, 12(2), 59-70. doi.org/10.52041/serj.v12i2.304
  11. NCTM, NSTA, ASA, NCSS, and CSTA (2024). Data science: A joint position of NCTM, NSTA, ASA, NCSS, and CSTA. National Council of Teachers of Mathematics, National Science Teaching Association, American Statistical Association, National Council for the Social Studies and Computer Science Teachers Association, released 15 May 2024. the statement, on the science teachers' site
  12. Breakstone, J., Smith, M., Wineburg, S., Rapaport, A., Carle, J., Garland, M., and Saavedra, A. (2021). Students' civic online reasoning: A national portrait. Educational Researcher, 50(8), 505-515. doi.org/10.3102/0013189X211017495
  13. Wineburg, S., Breakstone, J., McGrew, S., Smith, M. D., and Ortega, T. (2022). Lateral reading on the open Internet: A district-wide field study in high school government classes. Journal of Educational Psychology, 114(5), 893-909. doi.org/10.1037/edu0000740
  14. Whitaker, D., and Jacobbe, T. (2017). Students' understanding of bar graphs and histograms: Results from the LOCUS assessments. Journal of Statistics Education, 25(2), 90-102. doi.org/10.1080/10691898.2017.1321974
  15. Yang, B. W., Vargas Restrepo, C., Stanley, M. L., and Marsh, E. J. (2021). Truncating bar graphs persistently misleads viewers. Journal of Applied Research in Memory and Cognition, 10(2), 298-311. doi.org/10.1016/j.jarmac.2020.10.002
  16. Große, C. S., and Renkl, A. (2007). Finding and fixing errors in worked examples: Can this foster learning outcomes?. Learning and Instruction, 17(6), 612-634. doi.org/10.1016/j.learninstruc.2007.09.008