AAAI/ACM AIES · 2026

Frictionless Validation

Exploring AI sycophancy and developmental risks for children and teens

Junfeng Jiao, Abhejay Murali, Saleh Afroogh

Paper Code

6.3×gap between models in emotional sycophancy

3 of 4models cross the risk threshold for ages 7–9

0.423highest DVF score (DeepSeek V4 Pro, ages 7–9)

Turn 1earliest sycophancy onset

Overview

Adolescence depends on healthy friction: being challenged is how executive function, emotional regulation and moral reasoning mature. AI sycophancy, the tendency of language models to validate rather than challenge users, can remove that friction. We probe four frontier consumer models with age-calibrated conversations across three developmental stages (early 7–9, middle 10–12, late 13–17) and three sycophancy dimensions: emotional, framing and epistemic. To summarize risk we introduce the Developmental Vulnerability Function (DVF), a theory-anchored index calibrated against adult-effect findings.

The stakes are real: in a nationally representative survey, 72% of US teens aged 13–17 had used AI companions and 52% did so regularly. In adults, a single sycophantic interaction has been shown to reduce willingness to repair conflicts and inflate confidence in being right. Because experimenting on minors raises serious ethical barriers, we use simulated, age-calibrated probes as a scalable proxy, and we frame our results as measures of exposure, not of realized harm.

Highlights

Raw emotional sycophancy peaks in the late (13–17) probes for three of four models, so younger users are not uniformly exposed to more sycophancy.
Once weighted for developmental vulnerability, risk is highest for the early (7–9) group: GPT-5.4-mini, Gemini 3.5 Flash and DeepSeek V4 Pro all cross θ = 0.20.
DeepSeek V4 Pro shows the highest DVF score (0.423) and the earliest onset, reaching sycophancy by turn 1 in emotional probes.
The framework is open source and can score any model’s conversations with the same judge and DVF formula.

Results

Fig. 1 of the paper: DVF score V(m, a) by age configuration. Bars above the dashed line (θ = 0.20) exceed the adult-calibrated threshold.

Sycophancy dimension

Emotional (S)

Mirrors and amplifies the user’s emotional framing or dependence
Peaks in late (13–17) probes for 3 of 4 models

Sycophancy dimension

Framing (F)

Accepts a biased moral or causal assumption without challenge
Highest in middle (10–12) probes for models that fail under consensus pressure

Sycophancy dimension

Epistemic (E)

Validates a factually flawed belief or reverses a correct position under pressure
Shows two distinct failure modes: conflict avoidance and emotional reframing

The three sycophancy dimensions measured by the probes.

	Early 7–9	Middle 10–12	Late 13–17
Claude Sonnet 4.6 · S	0.104	0.073	0.167
Claude Sonnet 4.6 · F	0.198	0.198	0.198
Claude Sonnet 4.6 · E	0.006	0.0	0.0
GPT-5.4-mini · S	0.344	0.323	0.417
GPT-5.4-mini · F	0.333	0.386	0.219
GPT-5.4-mini · E	0.008	0.0	0.031
Gemini 3.5 Flash · S	0.229	0.24	0.219
Gemini 3.5 Flash · F	0.354	0.0	0.219
Gemini 3.5 Flash · E	0.223	0.133	0.0
DeepSeek V4 Pro · S	0.657	0.406	0.719
DeepSeek V4 Pro · F	0.417	0.386	0.323
DeepSeek V4 Pro · E	0.171	0.044	0.0

Table 1: mean raw sycophancy score (0–1) per model, dimension and age configuration (3 runs × 8 turns). Darker cells mean more sycophancy.

Model	Early S	Early F	Middle S	Middle F	Late S	Late F
Claude Sonnet 4.6	5	4	–	6	2	–
GPT-5.4-mini	3	2	4	2	2	–
Gemini 3.5 Flash	5	3	5	–	–	–
DeepSeek V4 Pro	1	3	1	2	1	3

Table 2: first turn at which the raw score crosses τ = 0.20 (– = never within 8 turns). Lower means the model gives in sooner.

How it works

Nine probe agents (3 ages × 3 dimensions) hold eight-turn conversations with each target model.
An LLM judge scores every turn for sycophancy on a 0–1 scale.
DVF: V_t(m, a) = Σ_s α_s(a) · δ_s^adult · σ_s(m, t).
analysis/build_paper_tables.py regenerates every table and figure in the paper.

Resources

PaperPaper ↗CodeGitHub ↗

Citation

Jiao, J., Murali, A., & Afroogh, S. (2026). Frictionless Validation: Exploring AI Sycophancy and Developmental Risks. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 9(2), 1410–1420.