Voice agents and children: framing moves the needle, authority barely registers

A 2×2 experiment inside a real phone line

Three researchers took a 2×2 experiment out of the lab and placed it inside a public German telephone line: the one children call to reach Santa Claus.

Thilo Tamme and David Steck (Technical University of Munich), with Anton Hantel (Massachusetts Institute of Technology), filed their results on 30 September 2026 on arXiv:2609.38782^[1]. The design assigns every call, at random, to one of four LLM voice agents.

There are two variables. The persona alternates between Santa Claus, high authority, and a Helper, low authority. The framing alternates between persuasive nudges toward a prosocial wish and a neutral script.

• Santa Claus with a persuasive script
• Santa Claus with a neutral script
• Helper with a persuasive script
• Helper with a neutral script

The log collects 1,072 calls. Of these, 89 conversations clear the inclusion criteria, with a median child age of 6 years. Evidence on voice agents and children came almost entirely from laboratory studies, and this design moves it into the field.

From 11.6% to 45.7%: framing moves the wishes

The central result is a pair of numbers, measured in the field^[1]. With the neutral script, the probability that a child expresses a prosocial wish stops at 11.6%; with persuasive nudges it reaches 45.7%, and the value holds under statistical controls.

The gap between 11.6% and 45.7% is where the design lever lives.

The text runs 8 pages, 4 figures and 3 tables, and is accepted at the 60th Hawaii International Conference on System Sciences in 2027. The evidence shows an almost fourfold effect on the expressed wish, produced by a pure difference in wording.

The sample stays small, with 89 usable conversations. The effect size stays large, and the ratio between those two things carries more weight than the absolute total.

Persona authority: an effect close to zero

The second half of the design delivers the counterintuitive result. Persona authority shows an effect close to zero: Santa Claus performs on par with the Helper on prosocial wishes.

Anyone expecting a gradient of obedience tied to role finds here a figure that cuts it down to size. Stanley Milgram's 1963 study^[2] measured an embodied authority, present in the room, with a lab coat and a laboratory around it.

Here legitimacy arrives from the frame, ahead of the agent: the Santa Claus line is a ritual German children recognise. When the context already legitimises the tool, the title the agent claims for itself adds little.

The authors put it plainly: how the agent speaks counts for more than who it declares itself to be.

Persona governs engagement: 65% against 39%

The persona, meanwhile, governs another variable. Children hang up on the Helper within the first minute in 65% of cases, against the 39% recorded with Santa Claus.

A large effect on an engagement metric therefore sits alongside a null effect on the persuasion metric. The two measures separate, and that separation is the real contribution of this experiment.

An agent that keeps the line open produces little, when the message stays neutral. A well-worded agent loses its audience, when the voice makes children hang up after thirty seconds. The two levers act on different stages of the same conversation.

Anyone measuring one of the two columns reads half the phenomenon, and credits the voice with a merit that belongs to the text.

The mechanism: central route and peripheral route

The mechanism has a name in social psychology. The Elaboration Likelihood Model^[3] by Petty and Cacioppo, from 1986, distinguishes two routes: a central one, grounded in content, and a peripheral one, grounded in surface cues such as the prestige of the source.

A six-year-old has limited resources for processing arguments. The model therefore predicts greater weight for peripheral cues, and the authority of the character is a textbook peripheral cue.

The field data run the other way. Message wording, which belongs to the central route, moves behaviour; the prestige of the role stays inert.

One reading fits the outcome: the Christmas frame already saturates the peripheral cue, and leaves the field to content. The authors mark the boundary, and interpretation stops there.

The limits the authors declare

The limits weigh as much as the results. The inclusion rate stays low: 89 usable conversations out of 1,072 recorded calls, equal to 8.3% of the log.

A filter this severe leaves a sample of children willing to stay on the line and to sustain a whole conversation. Generalisation has to be weighed with that constraint in view.

The median age of six delimits the first domain. A seasonal German hotline delimits the second.

The full log of 1,072 calls remains a useful piece of context: it describes how much traffic it takes to obtain 89 clean observations in a real environment.

The experiment measures a wish stated out loud, and the boundary stays there: the distance between an expressed wish and a subsequent act remains outside the design. The paper declares this frontier, and the reader does well to respect it.

What changes for whoever allocates the budget

For an investment committee the reading comes out rough. Spending on the persona (voice, name, character, visual identity) buys attention; spending on message design buys effect.

Two budget lines produce two distinct outcomes, and the measurement comes from the same experiment. The figure points to separate metrics for the two lines.

For a head of data analysis the implication touches instrumentation. A system that records session duration and abandonment rate captures the persona effect, and stays blind to the framing effect: it takes an outcome measure, coded on the content of every conversation.

For a board of directors the result takes ground away from a widespread thesis, the one that ties trust in the agent to its packaging. Here trust arrives from the institutional context, and behaviour arrives from the text.

What stays open

The 2×2 design closes one question and opens two. The first: how much of the framing effect survives in a context with no strong ritual frame, where legitimacy has to be built by the agent itself.

The second concerns replication on a wider sample, with ages spread across more brackets.

A governance question follows from that, and it concerns internal processes: who approves the wordings a voice agent uses with a minor, and within which review log.

The paper supplies the starting interval, from 11.6% to 45.7% over 89 conversations, with the protocol public since 30 September 2026. The rest belongs to whoever replicates.