Source-linked AI summary
Persona and Persuasive Framing in AI Voice Agents: A $2\times2$ Field Experiment with Children
Thilo Tamme, David Steck, Anton Hantel
TL;DR
Evidence on how conversational-agent design shapes children’s susceptibility to persuasion has come largely from laboratory studies. This paper reports a randomized field experiment in a public German Santa hotline that varied agent persona and persuasive versus neutral framing. Persuasive framing roughly quadrupled prosocial wish expression, while persona authority showed no evidence of increasing persuasion among children who engaged, though persona shaped who stayed on the line.
Problem
Evidence on how conversational-agent design shapes children’s susceptibility to persuasion has come largely from laboratory studies.
Source: §1 · p. 1Method
A randomized 2×2 field experiment in a public Santa hotline varied whether an LLM voice agent presented as Santa or Santa’s Helper and used persuasive or neutral framing.
Source: §1 · p. 1, §3.1 · p. 3Results
Persuasive framing roughly quadrupled prosocial wish expression, while high-authority persona showed no evidence of greater persuasion among children who engaged and persona shaped who stayed.
Source: §6 · p. 6Takeaways & Limitations
In this hotline setting, the conversation script rather than the speaker’s claimed authority moved children’s wishes, while persona affected engagement.
Source: §6 · p. 6The analytical sample was small (n = 89), and the single-culture, single-ritual setting with masked AI identity limits generalization to disclosed-AI settings.
Source: §5.3 · p. 6
Abstract
from arXiv · showhide
Conversational agents increasingly interact with children, yet evidence on how their design shapes children's susceptibility to persuasion comes almost entirely from the lab. We report a $2\times2$ randomized field experiment embedded in a public German Santa Claus telephone hotline. Children's calls were randomly routed to one of four LLM voice agents varying persona (Santa, high authority, vs. Helper, low authority) and framing (persuasive nudges toward prosocial wishes vs. neutral). Of 1,072 logged calls, 89 conversations (median age 6) met inclusion criteria. Persuasive framing raised the probability of a prosocial wish from 11.6% to 45.7%, robust to controls. Persona authority showed a near-zero effect: Santa did not outperform the Helper. Persona instead shaped engagement; children hung up on the Helper far more often within the first minute (65% vs. 39%). Where context already lends an agent legitimacy, how it speaks shapes children's compliance more than who it claims to be.
1. Introduction
The study examines how conversational-agent design affects children’s susceptibility to persuasion, an important question as voice agents become common in children’s lives. It tests persona and conversational framing in a field experiment and reports that framing, more than persona authority, shaped prosocial wishes.
1. Introduction
- The 2×2 field experiment varied whether the agent introduced itself as Santa or Santa’s Helper and whether it used persuasive or neutral framing.The outcome was whether a child spontaneously expressed a prosocial wish.Source: §1 · p. 1
- The study provides rare in-the-wild causal evidence on child–agent persuasion: framing increased prosocial wish expression roughly four-fold, while persona authority showed no detectable effect on wishes.The authors also report that persona affected whether callers stayed in the conversation.Source: §1 · p. 1
- The field paradigm paired consent-gated access, a fictional authority figure, capped nudges, anonymized data, and a transcript-based manipulation check.The authors describe the check as one many field studies omit.Source: §1 · pp. 1–2
- Fewer than one in ten logged calls yielded an analyzable child conversation, a screening yield the authors report as a planning parameter for future field studies.They also report call duration and child speech share as planning parameters.Source: §1 · p. 2
2. Theoretical background
The theoretical background treats agents’ social cues and claimed authority as potential influences on children’s responses, while also proposing that conversational framing may shape persuasion. It motivates testing these ideas in a child field experiment, where such evidence is scarce.
2.1. Agents as social actors
- Voice agents can elicit social responses, and anthropomorphic design is associated with increased trust and compliance intentions in service contexts.Children attribute intelligence and feelings to voice assistants and may form parasocial relationships with their characters.Source: §2.1 · p. 2
2.2. Persona authority and source effects
- The authors expected Santa’s culturally legitimate authority to elicit more prosocial wishes than the Helper under persuasive framing.The background notes that authority can be double-edged when legitimacy is lacking.Source: §2.2 · p. 2
2.3. Framing, peripheral route, and age
- The authors expected persuasive framing to increase prosocial wishes and younger children to be more susceptible, proposing peripheral cues as relevant to children’s responses.The background describes persuasion knowledge as developing with age.Source: §2.3 · p. 2
- The authors state that prior child field experiments had not manipulated persona authority and persuasive framing while measuring a behavioral outcome.They situate the gap within concerns that HRI research relies heavily on laboratory studies.Source: §2.3 · p. 2
3. Methodology
The study embedded a randomized 2×2 experiment in a public German Santa hotline, varying agent persona and conversational framing. From 1,072 logged calls, 89 conversations entered the analytical sample, with exclusions differing by persona but not framing.
3.1. Setting and design
- Calls were randomly routed to one of four LLM voice agents in a 2×2 design crossing Santa versus Santa’s Helper with persuasive versus neutral framing.The agents shared guardrails and a technical stack; only the persona introduction and framing script differed.Source: §3.1 · p. 3
- The deployment used adult consent gating, a fictional rather than institutional authority figure, capped nudges, age-appropriate guardrails, caller-controlled hang-ups, and anonymized identifiers.These constraints were built into the deployment’s design.Source: §3.1 · p. 3
3.2. Sample
- The sample was formed by sequentially excluding calls outside the study window, short calls, repeat calls, adult callers, unintelligible calls, and a jailbreak attempt.Calls of one minute or less could not have reached the treatment stage.Source: §3.2 · p. 3, Table 1 · p. 3
- 89 of 1,072 logged calls entered the analytical sample; exclusions differed by persona but not framing.A usable stated age was available for 70 children, with a median age of 6.Source: §3.2 · p. 3
window (December 15–25).
The study defines prosocial wishes by their primary beneficiary and codes them manually, with a conservative automated analysis used as a robustness and fidelity check. It tests the hypotheses using logistic regressions and corroborates binary contrasts with Fisher’s exact tests.
window (December 15–25).
- The experimental cells, automated nudge detection, and prosocial-wish expression are reported for a sample of 89 conversations.The supplied table caption identifies these as the table’s contents.Source: Table 2 · p. 3
3.3. Measures and coding
- A prosocial wish is one whose primary beneficiary is someone other than the child, and manual coding was checked against a conservative automated classifier with 87% agreement (Cohen’s κ = .73).The automated analysis flags nudges in agent turns and other-oriented wishes in child turns.Source: §3.3 · p. 4
3.4. Analysis
- Logistic regressions tested framing, persona contrasts, and age; Fisher’s exact tests corroborated the binary contrasts.The framing analysis included full-sample and core-sample models, plus a robustness model with controls.Source: §3.4 · p. 4
4. Results
Persuasive framing increased prosocial wish expression, whereas persona did not differentiate wish outcomes; manipulation checks show the intended nudges were usually delivered.
4.1. Manipulation check
- 76% of persuasive-condition conversations contained at least one detected prosocial nudge, versus one borderline instance among 43 neutral conversations.Most persuasive calls without a detected nudge ended before the script reached the nudging stage; control-condition contamination was negligible.Source: §4.1 · p. 4
4.2. Persuasive framing (H1)
- 45.7% of children expressed prosocial wishes under persuasive framing, versus 11.6% under neutral framing (odds ratio 6.4, p < .001).The effect remained significant with controls for nudge count, age, duration, and days to Christmas.Source: §4.2 · p. 4, Figure 2 · p. 4
4.3. Agent persona (H2)
- 46.2% of Santa-persuasive and 45.0% of Helper-persuasive conversations contained prosocial wishes; the persona contrast was near zero (β = 0.05, p = .94).Both persuasive cells outperformed the Santa-neutral baseline; the two persuasive personas were statistically indistinguishable.Source: §4.3 · p. 5, Figure 3 · p. 5
experimental cell (95% Wilson CIs).
Age did not predict prosocial wish expression, while persona was associated with early caller disengagement; the gap emerged before nudges were delivered.
4.4. Age (H3)
- Age did not predict prosocial wish expression (β = 0.06, p = .47), and the age contrast was underpowered.A posthoc simulation estimated roughly 160 observations per group would be needed for 80% power to detect the observed persona difference.Source: §4.4 · p. 5
4.5. Persona and early disengagement
- 64.6% of Helper calls ended within the first minute, versus 39.2% of Santa calls (OR = 2.82, p < .001).The gap appeared under both framings and across alternative cutoffs from 30 to 120 seconds; the analysis remains exploratory.Source: §4.5 · p. 5, Figure 4 · p. 5
- 40.5% of Helper callers hung up without saying an intelligible word, versus 8.9% of Santa callers (OR = 6.99, p < .001).These calls were ended by the caller; nudges had not yet been delivered, so this early-disengagement analysis was unconfounded by framing, though exploratory.Source: §4.5 · p. 5, Figure 4 · p. 5
5. Discussion
Framing produced a large, robust effect alongside a near-zero persona estimate; the authors interpret the ritualized setting as already supplying legitimacy. Persona shaped who engaged, while framing shaped compliance among those who stayed.
5.1. Framing as the primary lever
- Framing produced a large, robust effect alongside a near-zero persona estimate; the authors interpret the ritualized setting as already supplying legitimacy.They found no evidence that the authority figure added persuasive force once children were inside the legitimate frame.Source: §5.1 · p. 5
- 65% of Helper calls versus 39% of Santa calls ended within the first minute, and 41% versus 9% ended without speech; expectation and voice remain rival explanations.The authors cannot separate deference to authority from callers’ expectations or the personas’ different voices.Source: §5.1 · p. 6
- Santa belief was high (M = 4.16/5) and AI awareness low (M = 2.21/5), so the Helper may have inherited Santa’s authority and weakened the intended contrast.The authors identify separating persona authority from institutional legitimacy as a next step.Source: §5.1 · p. 6
5.2. Implications for design and policy
- Conversation scripting is cheap to change but hard for parents and regulators to see; a lexical classifier recovered treatment scripts and achieved acceptable agreement with manual coding (κ = .73).The authors frame transcript-level auditing as feasible for detecting hidden persuasive design.Source: §5.2 · p. 6
- Prosocial wishes served as a benign stand-in for persuasion that could also serve commercial ends, motivating transcript-level scrutiny.The authors note that manipulative persuasion can be harder to catch when it is embedded in conversation.Source: §5.2 · p. 6
5.3. Limitations and future research
- Claims are bounded by the small analytical sample (n = 89), unverified ages, parental presence, one culture and ritual, masked AI identity, and single-call outcomes.Durable attitude or behavior change is unknown, and the setting limits generalization to disclosed-AI contexts.Source: §5.3 · p. 6
6. Conclusion
In a randomized field experiment inside a real Santa hotline, persuasive conversational framing roughly quadrupled the share of children expressing a prosocial wish. The conclusion is that persona shaped who listened, while framing shaped what they wished for.
6. Conclusion
- Persuasive conversational framing roughly quadrupled the share of children who expressed a prosocial wish in a randomized field experiment inside a real Santa hotline.Among children who stayed on the line, the authors found no evidence that a high-authority figure was more persuasive than a low-authority assistant.Source: §6 · p. 6
- The persuasive power of a child-directed agent lay in conversation design, a layer described as cheap to change and difficult for parents or regulators to see.A simple classifier could reliably identify which script a child had heard; the study examined a benign wish.Source: §6 · p. 6
2.0. UNICEF Office of Global Insight and Policy.
The supplied section consists of bibliography entries spanning conversational agents, children's perceptions, AI persuasion knowledge, identity disclosure, and digital-human customer service.
2.0. UNICEF Office of Global Insight and Policy.
- The cited bibliography entry identifies a 2024 work by J. Watson, F. Valsesia, and S. Segal, and its visible title begins “Assessing AI”; no further title text is shown.Source: §2.0 · p. 8
- The listed works include research on children's perceptions of conversational agents and identity disclosure in voice-chatbot design.The entries identify CHI 2020 proceedings and an advance online publication in Management Science.Source: §2.0 · p. 8
- The bibliography includes a work on celebrity digital-human customer-service agents.The entry names the Proceedings of the 56th Hawaii International Conference on System Sciences.Source: §2.0 · p. 8