06 · UX research
How people respond to a co-viewer presented as human or AI.
An exploratory study with 48 participants. I built the experiment, analysed the response logs in Excel, and connected behaviour with how participants interpreted their partner.
01 Research question
When we watch something with other people, we notice when they laugh, how strongly, and whether it feels right. AI is starting to join these moments as companions, co-viewers and voice agents.
I asked whether people respond differently when their co-viewer is introduced as human or AI. Both conditions used the same three laughter sounds; participants chose when to trigger them. I examined response frequency, sound choices and the experience of watching together.
The outcome is an exploratory account of behaviour and interpretation, followed by design questions for AI co-viewers.
02 Study design
48 participants took part in 24 paired sessions: 12 pairs in Human and 12 in Told-AI. Each pair watched the same 11-minute video in separate rooms. They could not see or talk to each other; A, S and D triggered laughter sounds heard by their partner.
The lightest response, a small laugh through the nose.
A restrained laugh, like trying to hold it in.
The strongest response, a clear open laugh.
Human condition. Told the co-viewer was another participant.
Told-AI condition. Told the co-viewer was an AI under development that reacts when it judges the moment fits.
The partner was another participant in both conditions. The three sound samples and the video were held constant, while the identity instruction changed. The timing and sequence of responses were generated by each pair, so the received audio was not identical across sessions.
The video contained 37 clips, coded as Normal, Amusing, Funny, Awkward, Offensive or Dark Humour. Three pilot participants independently checked my coding. Each key press was timestamped and linked to the clip being shown. These records capture a chosen sound response, rather than a recording of spontaneous laughter.
An 8-person pilot changed the setup. The on-screen soundboard drew attention away from the video, and the processed “AI” voice seemed to influence reactions itself. I moved the controls to the keyboard and used the same sound samples in both conditions to reduce those competing explanations.
03 Working with the data
I used Excel to align the logs with the video, check event validity and summarise responses. Keeping the cleaning decisions visible mattered: repeated key events could otherwise inflate the totals.
Converted clip boundaries to milliseconds and aligned event times with the video’s synchronisation point. Each event could then be assigned a clip and humour category.
Separated valid responses from records outside the analysis window. I then checked short-gap runs of repeated presses rather than treating every logged event as a separate response.
Used FILTER, COUNTIFS and PivotTables to inspect responses by condition, sound and content. The reported results use the final cleaned event table.
Kept counts separate from response shares. For humour categories with different exposure times, I compared responses per participant-minute instead of raw totals.
Workbook excerpt · final event table
Anonymised rows from order_B. Relative time is measured from the synchronisation point.
| Participant | Time (s) | Response | Clip | Category |
|---|---|---|---|---|
| p002 | 7.187 | SuppressLaugh | 1 | Normal |
| p002 | 35.812 | BigLaugh | 2 | Funny |
| p002 | 37.640 | BigLaugh | 2 | Funny |
| p001 | 38.703 | Nasal | 2 | Funny |
| p002 | 38.984 | BigLaugh | 2 | Funny |
Source: order_B · five Human-condition events · all marked Valid.
Timeline!D3 converts the video timestamp in B3 into milliseconds.
=IF(B3="","",VALUE(LEFT(B3,2))*60000+VALUE(MID(B3,4,2))*1000+VALUE(RIGHT(B3,1))*100)The final review flagged runs with at least three events from the same participant and response, with every adjacent gap ≤60 ms. Keeping the first event in each run removed 419 repeated records. The original tables were preserved alongside the audit.
valid events before the final review
repeat records removed
valid responses retained
Source: Cleaning_B_Audit · 26 flagged runs. The rule is a data-cleaning choice, not a claim that all rapid presses are accidental.
Reaction rate = valid responses ÷ (participants × clip duration in minutes). Both conditions had 24 participants and the same 72.6 seconds of Dark Humour. This makes the rate comparison explicit; it does not remove the fixed-order limitation.
04 Findings
Told-AI produced fewer valid responses overall. Within those responses, BigLaugh accounted for a larger share. Human still had more BigLaugh presses in absolute terms: 925 versus 819. These are different measures of participation.
valid laughter responses from 48 participants.
responses in the Human condition, against 1,896 in Told-AI.
of Told-AI responses were BigLaugh, against 37.5% in Human.
Told-AI participants ended up believing the co-viewer was an AI.
Self-reported attention was similar (5.46 in Human; 5.38 in Told-AI). That offers little support for an attention-only explanation, although these ratings do not establish why the response totals differed.
Offensive clips showed no consistent direction across the session: Told-AI had a higher reaction rate in the first third, while Human was higher in the second and third. A single overall total would hide that variation.
For Dark Humour, Told-AI had a higher reaction rate: 13.26 versus 10.06 responses per participant-minute. All Dark Humour clips appeared near the end, so content and session timing remain entangled. This is an observed pattern, not evidence that AI framing generally encourages dark humour.
Only 15 of 24 Told-AI participants finally believed they had watched with an AI; 7 thought it was a person and 2 were unsure. In the Human condition, 4 of 24 ended up unsure.
I reviewed all 24 pair timelines alongside participants’ accounts. Their explanations referred to response timing, silence, sound choice and humour taste. One Human-condition participant became unsure because the partner pressed “too much”; another interpreted an extreme response as human because they expected AI to stay within limits. These accounts help interpret the logs, but do not show when beliefs changed or who influenced whom.
05 Design implications
The study separates participation, sound choice and interpretation. I used those distinctions to propose the following directions. They have not yet been evaluated in a working AI system.
Observed
Told-AI had fewer responses but a higher BigLaugh share.
Design direction
Let the system decide whether to respond separately from which sound to use.
Next test
Compare those decisions separately and ask users about presence and appropriateness.
Observed
Offensive and Dark Humour did not show the same pattern.
Design direction
Consider the content before triggering laughter; a single response policy may miss important differences.
Next test
Vary content order and evaluate responses across several sensitive contexts.
Observed
The same behaviour could support different identity judgements; final AI belief was incomplete.
Design direction
Consider how users interpret response timing and silence alongside the identity introduction.
Next test
Collect repeated identity judgements rather than relying only on an end-of-session question.
I came away with a more specific question than whether people laugh more with AI: how does an identity introduction interact with what happens during shared viewing? The logs describe participation, while the questionnaire and pair accounts show how differently the same limited signals can be interpreted.
My part
Limits of this study: the analysis is descriptive and exploratory, without significance testing. Fixed clip order confounds content with session timing. Key presses record chosen sound responses, not spontaneous laughter. Identity belief was measured only after viewing. A few active participants have substantial weight in pooled totals, and participants responded in pairs, so individual responses are not independent observations. Close timing does not establish influence. Further work should vary clip order and examine participant- and pair-level variation.