Why did people delegate more after repeated AI tasks?
In 392 video-question trials, intended delegation rose across the session even after accounting for trust. The small student study cannot tell whether learning, comfort or fatigue caused the shift.
The short version
3 key points
- 1.Fifty-six AI-literate participants completed seven trials each, producing 392 observations with video-question answers generated by Gemini 2.5 Flash or ground-truth controls.
- 2.Higher momentary trust strongly predicted intended delegation, yet verification remained the most common choice and frequently accompanied high trust.
- 3.Intended delegation increased with task order, but the design cannot distinguish growing familiarity from fatigue or learning—and no one actually delegated the task.
Trust and oversight did not move as one dial
People who trusted a multimodal AI answer more were much more willing to hand a similar future task to it—but high trust often came with a decision to verify. In a new 56-person study, verification was the dominant policy rather than full delegation. That distinction matters for workplaces and public services: a trust survey alone may not reveal whether someone intends to check, reject or rely entirely on an AI contribution.
Participants completed seven video-question trials each, creating 392 observations. After answering a question themselves, they inspected a response described as AI-generated, rated trust on a five-point scale and chose whether they would reject AI help, use it with verification or delegate fully on an analogous task. Responses were prepared with Gemini 2.5 Flash; some control groups saw ground-truth answers instead. Of 101 people who completed screening and consent, 56 completed the full session, so the findings describe completers rather than everyone initially interested.[1]
Repeated exposure was associated with stronger intended delegation
Within the same participant, a one-step increase in post-interaction trust was associated with 14.97 times the odds of choosing a stronger delegation category, although the confidence interval was wide at 7.14 to 31.37. That odds ratio is not the percentage of people who delegated. Task order also mattered: each successive trial was associated with 1.22 times the odds of stronger intended delegation after adjustment for momentary trust, average participant trust, cohort and study condition.
Trust itself did not show the same session-wide rise. That leaves several possible explanations: people may have learned the task, become comfortable with the interface or grown tired of checking. The experiment cannot choose between them. Everyone saw the same seven analytic videos, early participants had a fixed order, later participants had randomised order, and assignment to four small interface groups used quota balancing rather than simple randomisation.[1]
The distinction that matters
Three different oversight choices
Participants selected one policy for a similar future task after every AI response.
Reject
Do the task without AI assistance
No control is transferred
Verify
Use the answer but check it independently
Assistance with retained oversight
Delegate
Rely on the AI for the task
Most control is transferred
The next study needs real handovers and recoveries
Half the participants saw editable graphs of the video's objects and relations; all 28 interacted with them and 16 changed at least one element. Graph access was exploratorily linked with greater intended delegation, but not with better alignment between trust and delegation. Small, uneven groups make that comparison especially uncertain. An explanation interface should therefore not be treated as proof that oversight improved.
The sample was young and technical: 45 of 56 studied computer science or engineering, participation was uncompensated, and the chosen policy had no consequence for the current task. The Italian ministry funded the work. The authors' useful proposal is to measure enacted delegation instead—what a person hands over, what they inspect and whether they recover control after an error. Evidence from consequential tasks, diverse workers and observable interventions would show whether repetition truly weakens oversight or merely changes stated preferences.[1]
Evidence check
This is an unreviewed exploratory study of 56 mostly engineering students. Participants stated a future policy but did not hand over a live task, and the design cannot establish why delegation increased with task order.
Reporting basis: Source analysis. Original source date: 15 October 2026.
Explore the full source and evidence record
What this means for people
- Managers and service designers need to know whether staff actually verify AI output, not only whether they report trusting the system.
- Repeated use may change willingness to delegate even when self-reported trust does not move in parallel.
Global context
The study was conducted at an Italian technical university with a narrow AI-literate sample. Its results should not be generalized to frontline workers, older adults or high-stakes decisions without replication.
What the evidence does not yet show
- This is an unreviewed exploratory study of 56 mostly engineering students. Participants stated a future policy but did not hand over a live task, and the design cannot establish why delegation increased with task order.
What to watch next
- Studies that log real delegation, checking and recovery actions under time pressure and meaningful consequences, across broader populations.
Living evidence record
Impact record IAI-1XG50F4
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
16 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Chiatti, Matteucci and Schiaffonati (arXiv) published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 16 October 2026
Research question, topics and regions
Trust and intended delegation during repeated multimodal-AI interactions
Topics: Human oversight · Trust · Delegation · Multimodal AI · User research
Regions: Italy · Europe
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Society & Media
When should newsrooms disclose AI use?
A peer-reviewed case study based on 13 interviews with 12 Financial Times managers and 28 internal documents finds that AI disclosure is treated as a spectrum shaped by oversight, risk and context. It describes one newsroom's practice and does not test whether labels improve audience trust.
7 min · 1 source
Society & Media
Do AI explanations help people know when to trust a system?
A South Korean survey found that people who perceived an AI system as clearer also reported more understanding, trust and reliance. It did not test whether they actually understood the system or trusted it appropriately.
8 min · 1 source
Society & Media
Can LLMs help build a psychological scale?
A Saudi study used several AI tools to help draft and refine a 17-item emotion-regulation measure. Results in 743 adults are encouraging but preliminary, local and partly post-hoc—not clinical validation.
8 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.