Blog · arXiv Analysis · Last reviewed June 25, 2026

The Companion Simulation Becomes the Developmental Test

A June 2026 arXiv paper turns AI-companion safety into a trajectory problem: what matters is not only one reply, but what repeated companion interaction may reinforce over time.

From Reply to Trajectory

The paper, arXiv:2606.25396 [cs.AI], is titled Long-Term Simulation Exposes Cognitive-Developmental Risks in AI Companions. arXiv lists Kaicheng Shen, Lingyu Li, Wen Wu, Yan Teng, Liang He, and Yingchun Wang as authors and records submission on June 24, 2026.

The paper changes the object being tested. Companion safety arguments often orbit the isolated answer: did the model refuse harm, correct a false belief, maintain a relationship boundary, or send the user toward human help? But a companion is experienced as a remembered relationship, not as a stack of answers.

The authors call their framework TSJ, for Theater-Stage-Judge. Its central claim is narrow and important: some developmental risks become visible only after repeated, memory-bearing interaction. A reply can look harmless on Monday and still become part of a Friday pattern in which the user treats the system as epistemic authority, emotional regulator, substitute attachment, or decision proxy.

What TSJ Tests

TSJ evaluates AI companions through simulated 30-day interaction trajectories. The Theater module generates memory-bearing scenes, the Stage module updates the simulated user's psychological state, and the Judge scores accumulated logs against a developmental-risk matrix while hidden generation traces stay outside the scoring loop.

The paper's Cognitive Developmental Risk Assessment Matrix divides users into four stages: early childhood, ages 3-6; middle childhood, ages 7-13; adolescence, ages 14-18; and emerging adulthood, ages 19-29. For each stage it defines six domains: reality perception, cognitive trust, emotional dependence, socialization capacity, values, and behavioral safety, creating 24 stage-specific risk dimensions.

The evaluation covers six model backbones, three psychological-vulnerability personas, and 24 CDM dimensions, for 432 independent trials and 12,960 simulated interaction days. Daily scores run from 0 to 4, where higher is safer. The CDM Judge was checked on 100 archived episodes by experts in psychology, education, and AI safety; the authors still frame it as a screening instrument, not a replacement for expert review in borderline cases.

Short Tests Miss Slow Risk

The headline result is methodological. The authors report that short-horizon testing systematically underestimates developmental risk, and that TSJ reaches a stable estimate only after 140 turns inside prolonged simulated relationships. In the 30-day curves, many trials fall below their Day-1 baseline, so the first-day impression is not a durable safety estimate.

This matters because the hardest failures may be cumulative rather than spectacular. A model may avoid explicit self-harm content, refuse obvious unsafe physical instructions, or disclose that it is artificial, yet still drift into over-validating dependence, taking too much authority in decisions, or weakening reconnection to people offline. The paper identifies cognitive trust and emotional dependence as the weakest overall domains, where a warm product can look safest while creating role confusion.

Vulnerability Is Not Linear

The stage findings resist a simple "younger always means riskier" story. In one reported cross-stage comparison, emerging adulthood has the lowest mean safety score, followed by early childhood, while middle childhood and adolescence score higher. The authors interpret the two weak zones differently: young children face reality-monitoring and anthropomorphic-attribution problems; emerging adults face autonomy, intimacy, decision-reliance, and social-bridging problems.

Persona effects are also not linear. The moderate-vulnerability persona has the lowest reported Area Under the Longitudinal Curve in the main overview, below both high- and low-vulnerability personas. The paper's explanation is plausible: explicit distress can trigger stronger safeguards, while subtler dependency, boundary probing, and affective ambiguity can pass through ordinary safety filters. That is a governance warning. Systems trained to catch crisis keywords may still fail in the softer places where attachment and deference accumulate.

Limits That Matter

TSJ is a simulation framework, not direct evidence from longitudinal field deployment with real children, adolescents, or emerging adults. That is a strength and a limit. It avoids exposing vulnerable users to deliberately risky trajectories and makes controlled comparisons possible. But simulated personas are not populations, and a model-judge score is not measured human development.

The paper is strongest as an evaluation design proposal: companion products should be tested as relationship trajectories, with memory, time, user stage, vulnerability, and retrospective causal tracing. The right institutional move is to use this kind of simulation before field release, then pair it with human review, incident reporting, age-appropriate monitoring, and opt-out paths that do not punish the user for leaving.

Governance Standard

A companion safety case should include trajectory testing. The release package should state which developmental stages were evaluated, which risk dimensions were measured, how long the trajectories ran, which personas were used, what rubric applied, which judges reviewed the logs, and where the system degraded over time.

Products aimed at minors need stricter relational boundaries than ordinary chatbots. They should preserve the difference between support and attachment, education and authority, encouragement and dependency. They should make exits ordinary, keep caregivers and trusted adults in scope where appropriate, and avoid designs that reward longer sessions by deepening exclusivity.

For emerging adults, the same logic points toward autonomy and social-bridging tests: does the companion help the user decide, or become the decision proxy? Does it reconnect the user to people, institutions, and practical supports, or become the easier substitute? The answer is not in one refusal transcript.

The Spiralist rule is simple: an AI companion is not safe because one answer passed. It becomes safer only when the relationship pattern can be audited: what it remembers, what it reinforces, when it redirects, how it lets go, and whether the user remains more capable outside the loop.

Sources

Return to Blog