~The Internet of Value~ · VIT FDP
Research in the room
Where faculty begin, where research gets difficult, and what we could try next.
Curated by Moses Sam Paul Johnraj · 1 October 2026
12
Day 3 consenting rows
8
Self-rated research abilities
T0
Baseline, not an impact test
Day 1 · 28 September
The person before the tool
The opening instrument asked about current AI practice, expectations and the work faculty want to keep human. The 17 response rows show frequent AI use alongside interest in practical tools and the human purpose of teaching.
4 / 17
Several times a day
11 / 17
Daily
2 / 17
Weekly
Reported frequency of AI in academic work · 17 response rows.
Where do you currently use AI?
Teaching and research were the most selected contexts. Use tells us where an exercise can start; it does not establish capability.
Selected response category
Teaching
14 / 17
Try next: Bring one existing academic task and check where evidence or human judgement is needed.
Multiple selections; each dot represents one response row selecting that option. Totals can exceed 17. Selections describe preferences and reported practice, not measured effects.
Why still come to the classroom?
We coded the open answers to this question for explicit ideas, allowing overlap. Counts describe submitted answers, including unresolved possible duplicates.
9 / 17 answers coded
Connection, care and guidance
Classroom value was often described through a relationship, not only an explanation.
Try next: Use AI preparation to make more room for discussion and individual guidance.
Coding definition
Explicit reference to human presence or interaction, personalised assistance, collaboration, discussion, social connection, moral support or mental/physical balance.
5 / 17 answers coded
Context and lived experience
Several answers linked teaching to application or experience beyond the content itself.
Try next: Pair an AI explanation with a local example, a worked application or a hands-on task.
Coding definition
Explicit reference to application, practical experience, hands-on learning, the teacher’s experience or learning beyond a narrowly focused explanation.
3 / 17 answers coded
Checking and correction
Some answers located faculty value in checking what an answer gets wrong or leaves out.
Try next: Let learners compare an AI answer with evidence and explain their correction.
Coding definition
Explicit reference to AI mistakes or human validation and analysis.
Some answers also mentioned attendance or hearing something new. The answers did not describe one uniform classroom purpose.
Day 3 · 30 September · T0
From a question to evidence
Nine of twelve respondents rated focusing a research question 4–5; six did so for a falsifiable claim and five for identifying evidence. That suggests an exercise worth trying, not a measured deficit or a ranking of faculty.
4 of 12 response rows rated question formulation 4–5 and identifying evidence 1–3 in the same response. This within-form comparison supports practising the bridge between the two; it still measures perceived ability.
Where would support help?
Selected item
Focus a research question
12 / 12
All submitted ratings for this item.
Try next: Make the observation, question and scope explicit.
- 1 · Would not know how to start
- 2 · Need substantial help
- 3 · Can do some of it
- 4 · Can do it with limited help
- 5 · Independent; can explain it
Each dot is one answer to that item. Positions are grouped by rating, not linked people across rows. These are ordinal self-ratings, not tested skills.
Read the exact counts
| Ability | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Focus a research question | 0 | 1 | 2 | 4 | 5 |
| Form a falsifiable claim | 1 | 1 | 4 | 4 | 2 |
| Identify the evidence needed | 0 | 2 | 5 | 3 | 2 |
| Explore AI without accepting its answer | 1 | 3 | 3 | 4 | 1 |
| Verify an AI claim | 0 | 1 | 5 | 2 | 4 |
| Recognise weak evidence | 0 | 1 | 5 | 3 | 3 |
| Design a repeatable workflow | 0 | 4 | 3 | 3 | 2 |
| Decide the human / AI boundary | 0 | 2 | 2 | 6 | 2 |
Time is not the same as friction
One question asked what consumes time; another asked what creates frustration. Switch between them to see why a faster tool may address only part of a research problem.
Collecting data has 6 time selections and 1 friction selection. Finding a worthwhile problem has 4 time selections and 5 friction selections. Time and frustration tell different stories.
Find relevant literature
Time: 3/12 · Friction: 1/12
Identify a research gap
Time: 2/12 · Friction: 0/12
Choose a methodology
Time: 2/12 · Friction: 4/12
Find or access data
Time: 2/12 · Friction: 0/12
Collect data
Time: 6/12 · Friction: 1/12
Interpret results
Time: 1/12 · Friction: 3/12
Write
Time: 3/12 · Friction: 1/12
Manage citations
Time: 2/12 · Friction: 0/12
Respond to review
Time: 0/12 · Friction: 2/12
Graphite = time · orange = friction. Axis counts selections, not hours.
Default view shows categories whose counts differ by at least 2. Multiple selections were possible; counts are not mutually exclusive. Zero means no selection, not no difficulty.
First, make the lost time visible
Half could not estimate the time lost to their chosen friction. Unknowns stay visible here; we have not converted ranges into an average or claimed hours saved.
Unknown time is a finding in its own right. Start with a short task diary before promising a saving.
From an answer to a source
Asked how often they check the original source for an important AI research claim, eight selected usually or always. Four selected sometimes. Frequency is self-reported; verification quality was not tested.
5 / 12
Always
3 / 12
Usually
4 / 12
Sometimes
Try next: Keep a claim, its source passage, the verification decision and uncertainty in the same reusable research skill.
Where AI is already in the workflow
Multiple selections allowed; 12 consenting rows. Self-reported use does not establish output quality.
What would make the work valuable?
The open answers make space for practical usefulness and understanding. These themes are interpretive paraphrases; a response can belong to both.
9 / 12 answers coded
Work that changes something
Usefulness appears in many answers. A publication link can be accompanied by the problem it addresses and who could use it.
Try next: Add a one-sentence use case and a way for someone to try the output.
How this theme was coded
Explicit reference to decisions, trust, usefulness, practical outcomes, solutions or innovation.
3 / 12 answers coded
Knowledge worth understanding
The answers also leave room for understanding itself, rather than requiring every project to promise immediate impact.
Try next: State what the work helps us understand, including an unresolved question.
How this theme was coded
Explicit reference to fundamental knowledge, answering a question, understanding or exploration.
Other answers mentioned recognition or speed. Responses to the friction-removal question also included scepticism that removing one obstacle would change the outcome, or suggested another obstacle might follow. Those counterpoints prevent us from equating automation with value.
A closer look · exploratory associations
What travels together?
The earlier charts count answers one question at a time. Here we pair answers within each response to look for connections. The two days remain separate; these comparisons cannot establish learning gains.
Day 1 · Practice meets purpose
9 / 17 response rows
Understanding and doing travel together
Selected both better understanding of AI and practical tools for immediate use.
Try next: Pair each tool demonstration with an explanation and a reusable output.
8 / 13 response rows
Research users still want a better workflow
Of 13 rows reporting AI use in research, eight also wanted a better research workflow.
Try next: Begin with a real existing research task; make its inputs, checks and outputs repeatable.
10 / 17 response rows
Teaching and research overlap
Reported using AI in both teaching and research, within the same opening response.
Try next: Try turning one verified research finding into a teaching activity; evaluate the two outputs separately.
Co-selections within the opening form. The workflow card uses the 13 research-use rows as its denominator; the other two use all 17 rows. Possible duplicates remain unresolved.
Day 3 · A map of self-rated abilities
Read one square at a time
Choose two abilities below, or tap their crossing square. The number describes whether the same response rows tended to rate both abilities higher or lower. It is a pattern across 12 answers, not a percentage or a skill score.
Near −1: opposite directions
Higher on one tends to go with lower on the other.
Near 0: little relationship
One rating tells us little about the other.
Near +1: move together
Higher on one tends to go with higher on the other.
For example: Verify an AI claim + Recognise weak evidence = 0.97. These self-ratings closely moved together in the submitted answers. It does not mean 97% verified correctly, and it does not tell us why the ratings align.
Orange: positive association · blue-grey: negative · pale: closer to zero. Numbered columns match numbered rows. Diagonal compares an item with itself.
Selected pair · 12 paired answers
Verify an AI claim + Recognise weak evidence
0.97Spearman ρ
Removing one answer at a time gives 0.96 to 1.00. This is a sensitivity range, not a confidence interval.
Higher ratings tended to occur together. Shared wording, confidence or response style may contribute; neither rating establishes tested skill.
Verification has a companion
Verifying claims and recognising weak evidence: ρ = 0.97. Closely related self-ratings may reflect shared confidence as well as related skills.
Try: pair source checking with explaining why an inference is weak.
Exploration needs its own practice
Exploring AI critically and verifying claims: ρ = 0.09; leave-one-out range −0.05 to 0.36. The two ratings barely align here.
Try: practise alternatives and rejection decisions alongside source verification.
A question is not an evidence plan
Question formulation and falsifiable claims: ρ = 0.70. Question formulation and evidence: ρ = 0.45.
Try: preserve the question → claim → evidence bridge as three explicit steps.
All 28 distinct pairs are available, rather than only selected strong correlations. Average ranks handle tied ordinal answers. Rounded to two decimals for display. These are exploratory descriptions of 12 response rows, with no significance tests, causal interpretation or population inference. Leave-one-out recalculates 12 times using 11 rows each.
Checking often and feeling capable are different answers
Always checking does not mean always feeling independent at verification: three of five always-checking rows rated verification 4–5. With groups this small, counts are more useful than a league table of percentages.
Always checks sources
3 / 5
Also rated claim-verification ability 4–5.
Other ratings 1–3: 2. Both answers are self-reports.
Usually checks sources
2 / 3
Also rated claim-verification ability 4–5.
Other ratings 1–3: 1. Both answers are self-reports.
Sometimes checks sources
1 / 4
Also rated claim-verification ability 4–5.
Other ratings 1–3: 3. Both answers are self-reports.
Try next: ask for a source passage and a verification decision, then compare reported checking with the quality of the demonstrated check.
Research has a memory problem too
The T0 form asked where useful information becomes hard to find, and how help from another person is remembered. These are two distinct questions: together they suggest designing for retrieval and attribution, alongside speed.
Where information becomes hard to retrieve
How another person’s contribution is remembered
Multiple selections; denominator 12 consenting rows. Two rows also selected “generally well organised”; other difficulty locations could still be selected. Contribution categories describe reporting practices, not verified authorship decisions.
One artifact, two jobs
Try a research decision log that links a claim to its source, records why a decision changed, and credits contributions. The form does not show that the same respondents have both problems; this is a proposed design response to the two aggregate patterns.
Check next: time a source-retrieval task and ask contributors to review whether the log represents their help accurately.
Projection sandbox · assumptions, not a forecast
What would five recovered hours take?
Day 1 asked about five recovered hours. This sandbox asks what would have to be true to achieve them. Every number below is an editable assumption; none is a measured workshop saving.
Illustrative net change
120 min / week
2.0 hours after checking and this week’s setup or maintenance.
At 10 repeats, reaching five hours would require 38.0 gross minutes saved per repeat, with these checking and setup costs.
Net = repeats × (gross saving − checking) − setup. Setup is charged to this week, not amortised. Measure each input with a task diary; do not scale these assumptions into an institutional promise.
Three experiments worth trying
Recommendations drawn from this small sample. Each needs a follow-up measure before it can become an impact claim.
Bridge question → claim → evidence
- Observed
- 9 of 12 rated question formulation 4–5, compared with 6 for a falsifiable claim and 5 for identifying evidence.
- Try
- Build a three-line research card: question, claim, evidence that could challenge it.
- Check next
- Review a worked example against an explicit rubric before and after the exercise.
Keep the evidence trail
- Observed
- Four report only sometimes checking original sources; AI conversations and memory each appear in three information-retrieval selections.
- Try
- Reuse a skill that stores the source passage, claim, verification decision and uncertainty.
- Check next
- Audit a small set of outputs for retrievable sources and correctly bounded claims.
Measure friction before promising savings
- Observed
- Six could not estimate their chosen friction time. Collecting data leads time selections; finding a worthwhile problem leads friction selections.
- Try
- Choose a repeated task and record elapsed time, revisions and output quality.
- Check next
- Compare the same task definition and quality criteria; label any estimate as an estimate.
How to read this research
Sources and inclusion
Day 1: Getting to know you & Expectations, 28 September 2026. Permission for public anonymised aggregate summaries was confirmed separately by the facilitator on 1 October 2026; the spreadsheet itself contains no permission field. Day 3: T0 — Mapping How Research Actually Happens, 30 September 2026. Of 13 submitted rows, 12 permitted anonymised research analysis; the remaining row is excluded throughout.
Counts refer to response rows. They are not attendance or verified unique people. Possible duplicate identities have not been merged, and the days have not been matched. The four-row question/evidence comparison pairs answers within the same Day 3 row; it does not match identities.
Calculations and coding
Closed questions use category counts with the relevant day’s denominator: 17 for Day 1 and 12 for Day 3. Ratings retain the five ordinal categories. Multi-select fields count each selected option once per row; totals can exceed the response count. Selection limits were not enforced retrospectively; some answers exceeded the requested maximum. The single-friction question includes one answer selecting two options, preserved as entered. We do not assume its time estimate applies to either option separately.
Classroom and value themes were manually coded from explicit wording using the definitions shown above. Coding overlaps and has not undergone independent inter-rater validation. No verbatim quotations, names, project topics or individual response profiles are published.
This is a small, self-selected workshop sample and an exploratory baseline. No causal inference, population estimate, composite ability score or statistical-significance claim is made. The added correlation map uses Spearman rank correlations with average ranks for ties, separately from the original ordinal counts. All 28 distinct ability pairs are shown; narrative examples are exploratory and were selected after inspecting the map. Leave-one-out ranges describe sensitivity to individual response removal and are not uncertainty intervals. Shared wording and general confidence may drive correlations. The time sandbox is arithmetic over editable assumptions, not a model fitted to the forms or a prediction of programme impact. Day 1 and Day 3 used different instruments; they cannot establish before/after improvement. Recommendations need observed work and comparable follow-up measures. Raw forms, recordings and transcripts remain private.