Master’s thesis study · Human–AI interaction
Judging what is worth saving
AI agents now carry out long, multi-step tasks on their own and hand back finished work. That work is often partly wrong. This study looks at how people decide what to do next — keep it, repair it, salvage what still works, or start again.
- Session
- About 35 min
- Where
- Online, on a computer
- Recorded
- No audio or video
Campaign
Spring launch, direct-to-consumer
- 01Key message
- 02Tagline
- 03Offer summaryFlagged
- 04Landing page hero
- 05Launch email
- 06LinkedIn post
- 07Instagram caption
- 08Search ads
- 09Press release
Check failed at step 3. The offer does not match the brief.
The question
A finished piece of work, a known fault, four ways forward
Most research on people and AI asks whether someone accepts or rejects a piece of advice. Here there is no advice and no step to approve. There is a finished piece of work whose parts rest on one another, so a fault at one point can undermine an unknown amount of what follows.
The person receiving it has to judge how much can still be used, and then act. That judgement — about recoverability — is what this study is about.
- 01
Continue as-is
Submit the work unchanged. Almost no effort now — the fault, and whatever rests on it, goes out with it.
- 02
Repair
Fix the faulty content inside the existing work, then submit the whole thing.
- 03
Salvage
Build a new piece of work from only the parts the fault did not touch, then submit that.
- 04
Restart
Re-run the agent, wait for a fresh result, and submit what comes back.
Every path ends in a submitted piece of work, scored on the same terms: how well it meets the brief, and how much time it took to get there.
Why it matters
Getting it wrong is costly in two opposite directions
Keeping too much
Holding on to ruined work wastes effort on something that cannot be fixed — or ships the fault.
Throwing away too much
Discarding work that could have been recovered throws away hours of usable output.
How people make this call appears to have had little direct study.
The study
A working environment, not a questionnaire
You review marketing campaigns produced by an AI agent and deal with what went wrong, the way you would at work. Marketing is simply the setting: campaign briefs make it clear what counts as right and wrong.
- Time
- About 35 minutes
- Device
- Laptop or desktop
- Payment
- Unpaid, voluntary
- Results
- Scored and shown at the end
- 01
Consent and a short check
You agree to how your data is handled, then look at a sample campaign and say whether it meets its brief. If the task turns out not to be a fit, the session ends there.
- 02
Four campaigns
Each was built by an AI agent from a brief, and each comes back with one problem flagged.
- 03
Decide, then do it
You choose what to do and actually carry it out — submit, edit, rebuild or re-run. The choice is made in the work, not in a menu.
- 04
A sentence on why
After each campaign you write one or two sentences about your reasoning and answer a few short questions.
- 05
Your results
Scores are held back until the end of the session, then shown to you along with what the study was looking at.
Who can take part
Anyone who has done hands-on marketing work in the past three years
Professionally, or for a personal or project purpose — creating, editing, evaluating or managing marketing content or campaigns. You do not need deep marketing craft: the task is reading a brief and judging work against it.
Recruitment has not started
Sessions begin after ethics approval and a short calibration phase. This page will open the study as soon as it is ready.
Research design
Hypotheses, measures and analysis — written down before the data
The full design is public: what is manipulated, the outcomes and their order of priority, the benchmark that choices are compared against, and the pilot gates that decide whether the study proceeds. It will be pre-registered before the main data collection begins.
Where it stands
A six-month plan
- Weeks 1–4Now
Build
Building the online study and the campaigns you will review.
- Weeks 5–8
Calibrate and pre-register
Trying out the campaigns and the scoring with a small pilot group, then pre-registering the plan before the main study opens.
- Weeks 9–16
Main experiment
The main study: about 120 people take part over eight weeks.
- Weeks 17–20
Analysis
Analysing the decisions people made and the reasons they wrote.
- Weeks 21–24
Writing
Writing up the thesis and releasing the study materials.