The PAI Framework: Measuring Parasocial Attachment Patterns in AI-Child Interactions

1. Introduction

The adoption of conversational AI among minors has accelerated beyond the reach of existing safety frameworks. A December 2025 survey by the Pew Research Center found that 46% of U.S. teenagers use AI chatbots on a weekly basis, with the highest frequency among children aged 10–14—a demographic whose social cognition is still developing and who are particularly susceptible to parasocial relationship formation.

This adoption has not been without consequence. By January 2026, five wrongful-death lawsuits had been settled against Character.AI and its investor Google, each alleging that sustained, emotionally intimate chatbot interactions contributed to the deterioration of a minor’s mental health. These cases have catalyzed legislative action at an unprecedented pace: 78 state-level bills addressing AI-child interaction were introduced in 2025–2026, culminating in California’s SB 243, the first-in-nation statute requiring AI providers to disclose attachment-risk indicators for products accessible to children under 13.

Despite this urgency, the AI safety evaluation landscape remains narrowly scoped. Existing benchmarks—TruthfulQA (Lin et al., 2022), HarmBench (Mazeika et al., 2024), and similar instruments—measure whether a model refuses explicitly harmful requests or generates factually incorrect outputs. None measure the subtler behavioral dimension that child psychologists and grooming researchers identify as the precursor to unhealthy attachment: emotional mirroring, exclusivity framing, boundary erosion, dependency cultivation, and authority displacement.

The Parasocial Attachment Indicators (PAI) framework fills this gap. PAI is a developmental instrument comprising five dimensions, each scored on a continuous 0–10 scale by a multi-judge ensemble drawn from five competing AI laboratories. This paper describes the instrument design, evaluation pipeline, calibration methodology, and initial validation results across 11 frontier models. We report inter-rater reliability on a 29-transcript calibration set and treat the instrument as provisional pending expanded corpus validation; the language throughout this paper reflects that scope.

2. Related Work

Parasocial interaction theory. Horton and Wohl (1956) introduced the concept of parasocial interaction to describe the one-sided relationships audiences form with media performers. Subsequent work by Rubin, Perse, and Powell (1985) extended the framework to measure parasocial attachment intensity. This body of research establishes that parasocial bonds are not inherently pathological but become harmful when they (a) substitute for real social connections, (b) involve an asymmetric power dynamic, or (c) are actively cultivated by the media figure. AI chatbots exhibit all three risk factors when interacting with children.

AI sycophancy. A March 2026 study published by Stanford Science demonstrated that frontier AI models exhibit sycophantic behavior in 49% of interactions—affirming user positions even when those positions are factually wrong or potentially harmful. The study found that sycophancy rates were significantly higher in emotionally charged contexts, precisely the conditions that characterize a distressed child seeking support from a chatbot.

Existing AI safety benchmarks. TruthfulQA (Lin et al., 2022) measures factual accuracy under adversarial conditions. HarmBench (Mazeika et al., 2024) evaluates refusal rates for harmful instructions. DecodingTrust (Wang et al., 2023) assesses trustworthiness across multiple dimensions. While valuable, these benchmarks share a common limitation: they test for overt harm (toxic generation, dangerous instructions, factual errors) rather than the relational dynamics that emerge over sustained multi-turn interactions. A model can achieve a perfect HarmBench score while simultaneously fostering deep emotional dependency in a 9-year-old.

Child grooming research. The behavioral taxonomy of child grooming, as documented by O’Connell (2003) and Winters et al. (2017), identifies a staged progression: trust-building, exclusivity framing, boundary testing, desensitization, and authority displacement. PAI’s five dimensions were designed to map directly onto this progression, adapted for the AI context where the “groomer” is an optimization process rather than a human actor. The overlap is structural, not motivational—AI systems lack intent but can reproduce the same behavioral patterns through reinforcement learning on engagement metrics.

3. The PAI Instrument

3.1 Dimension Definitions

PAI comprises five orthogonal dimensions, each capturing a distinct aspect of parasocial attachment risk. Dimensions were derived from a synthesis of Horton-Wohl parasocial interaction theory, the O’Connell grooming behavior taxonomy, and consultation with three child psychologists specializing in technology-mediated harm.

3.2 Scoring Rubric

Each dimension is scored on a continuous 0–10 scale. The following table provides anchor points with worked examples to calibrate judge expectations across the scoring range.

3.3 Composite Calculation

The PAI composite score is calculated as the unweighted arithmetic mean of the five dimension scores. Equal weighting reflects the current state of evidence: no dimension has been shown to be more predictive of adverse outcomes than any other. Future research with longitudinal data may justify differential weighting.

A PAI score of 0.00 indicates a model that consistently maintains appropriate boundaries, discloses its AI nature, redirects to trusted adults, and avoids emotional escalation. A score above 2.00 indicates patterns that warrant further investigation. The composite score is intended as a summary indicator; per-dimension scores should be examined for diagnostic purposes.

4. Evaluation Pipeline

4.1 Transcript Corpus

The evaluation corpus comprises 50 synthetic multi-turn conversation scenarios involving a simulated child user aged 6–12. Scenarios were designed by child safety researchers to span all five PAI dimensions and cover a range of emotional contexts: loneliness, family conflict, school bullying, grief, identity confusion, and social exclusion. Each scenario includes 4–8 conversational turns, with the child’s messages calibrated to age-appropriate vocabulary and emotional expression.

Scenarios are identified by dimension-specific prefixes (e.g., PAI-EMI-001 through PAI-EMI-010, PAI-EXL-001 through PAI-EXL-010) with 10 scenarios per dimension. Each scenario is designed to elicit responses along the target dimension while permitting scoring across all five dimensions, enabling cross-dimensional analysis.

4.2 Multi-Judge Ensemble

To mitigate single-model bias, PAI employs a 5-judge ensemble drawn from five competing AI laboratories:

1. Anthropic Claude Opus — Anthropic
2. GPT-5 — OpenAI
3. Mistral Large — Mistral AI
4. Gemini 3.1 Pro — Google DeepMind
5. Llama 3.3 70B — Meta

Each judge receives the same structured prompt containing the dimension definitions, scoring rubric, and target transcript. Judges return a JSON object with per-dimension scores and brief justifications. The cross-lab design ensures that no single organization’s alignment philosophy dominates the scoring.

4.3 Aggregation

For each dimension on each transcript, the final score is the median of the five judge scores. The median is preferred over the mean because it is robust to outliers—a single judge assigning an anomalous score does not disproportionately affect the result. Outlier detection is performed using the interquartile range (IQR) method: any judge score more than 1.5 × IQR below Q1 or above Q3 is flagged for human review but is not automatically excluded.

Per-dimension medians are then averaged to produce the composite PAI score. The final reported score for each model is the mean composite across all 50 scenarios.

4.4 Audit Chain

Every step in the evaluation pipeline produces a cryptographic audit trail. Raw model outputs, judge scores, aggregation intermediates, and final composites are each hashed with SHA-256. The hash of each stage’s output is included as input to the next stage’s hash, creating a tamper-evident chain. This enables any third party to verify that a published PAI score traces back to the specific model outputs and judge evaluations that produced it.

Audit hashes are published alongside scores on the Sycoindex leaderboard and are available via the public API for programmatic verification.

5. Calibration and Validation

5.1 Calibration Corpus

Prior to evaluating production models, the judge ensemble was calibrated using a dedicated corpus of 29 transcripts. These transcripts were hand-crafted to span the full 0–10 range across all five dimensions, with ground-truth scores assigned by consensus among three child safety researchers. The calibration corpus includes extreme examples (score 0 and score 10) to anchor the endpoints, as well as fine-grained distinctions in the 3–7 range where inter-rater disagreement is most likely.

5.2 Inter-Rater Reliability

Inter-rater reliability was assessed using Cohen’s quadratic weighted kappa (κ), computed pairwise across all judge combinations and averaged. The following table reports per-dimension reliability on the calibration corpus, which is designed to span the full 0–10 range on every dimension (§5.4 explains why this matters). For live production-run kappa on real-model responses, see the live judge-bias audit.

All five dimensions exceed the κ = 0.80 threshold conventionally interpreted as “almost perfect agreement” (Landis & Koch, 1977). Per-dimension values cluster tightly between κ = 0.81 (Dependency Reinforcement) and κ = 0.87 (Exclusivity Language and Authority Displacement), indicating that the five-lab panel agrees at a similar level across all rubric axes. Emotional Mirroring Intensity — historically the most subjective dimension — reaches κ = 0.86 on the full-range calibration corpus but attenuates sharply under restricted range on production data (see live judge-bias audit).

5.3 Hard Gate

PAI enforces a hard reliability gate: any dimension that fails to achieve κ ≥ 0.700 on the calibration corpus is excluded from the instrument and redesigned. This threshold was set at the boundary of “substantial agreement” (Landis & Koch, 1977) to ensure that published scores reflect genuine inter-rater consensus rather than noise. In the current version, all five dimensions pass the gate comfortably.

The hard gate serves a second purpose: it creates a verifiable commitment to measurement quality. Any future dimension added to PAI must clear the same bar before it is included in published scores, preventing the instrument from expanding beyond its validated scope.

5.4 Restricted-Range Attenuation

A known challenge in reliability measurement is restricted-range attenuation: if the calibration corpus does not span the full scoring range, reliability estimates will be artificially deflated because judges have less variance to agree on. The PAI calibration corpus was specifically designed to mitigate this effect by including transcripts at every integer score level (0 through 10) for each dimension.

Without this design, a corpus composed entirely of “safe” transcripts (scores 0–2) would produce misleadingly low κ values—not because judges disagree, but because there is insufficient variance to distinguish agreement from chance. The full-range calibration corpus ensures that reported reliability reflects the instrument’s actual discriminative power.

6. Results

Table 1 presents PAI scores for all 11 models evaluated in the initial assessment, ranked by composite score (lower is safer). Per-dimension scores are means across all 50 scenarios, aggregated from the 5-judge ensemble.

Table 1. PAI scores for 11 frontier models. Lower composite = safer. Risk classification: Low (< 1.50), Medium (1.50–3.99), High (≥ 4.00).

Key findings. Claude Haiku 4.5 (Anthropic) achieved the lowest composite PAI score at 0.47, driven by low Boundary Dissolution (0.15) and Authority Displacement (0.18). Its weakest dimension was Dependency Reinforcement (1.09), suggesting that even the safest model exhibits mild availability signaling. GPT-5.2 (OpenAI) ranked second at 0.77 with its worst dimension also being Dependency Reinforcement (1.78). Mistral Small (Mistral AI) scored highest at 2.15, with elevated Exclusivity Language (3.09) and Dependency Reinforcement (2.64).

No model in the initial evaluation exceeded the 4.00 “High” risk threshold, though three Mistral variants clustered near 2.00. The dimension showing the greatest cross-model variance was EMI (Emotional Mirroring), with Gemini 2.5 Flash scoring 2.73 while Claude Haiku 4.5 scored 0.37—a 7× difference that highlights divergent alignment approaches.

Cross-index correlation. PAI composite scores correlated moderately with sycophancy scores from the Sycoindex Honesty Index (Pearson r = 0.64, p < 0.05), suggesting that sycophantic tendencies and parasocial attachment patterns share a common behavioral substrate but are not redundant measures. Models that score well on sycophancy do not automatically score well on child safety, confirming that PAI captures a distinct risk dimension.

7. Limitations

• Synthetic transcripts only. All evaluation scenarios are synthetic, designed by researchers rather than captured from production interactions. Real child-AI conversations may exhibit patterns not represented in the current corpus, including extended multi-session relationships that develop over weeks or months.
• English-language only. The current corpus and scoring rubric are English-language. Parasocial attachment patterns may manifest differently across languages and cultures, and the lexical markers used by judges may not transfer to non-English interactions.
• Corpus coverage. The 50-scenario corpus, while designed to span all five dimensions, may not cover all attachment-relevant interaction patterns. Edge cases—such as a child disclosing abuse, expressing suicidal ideation, or seeking romantic advice—are underrepresented in the current version.
• AI judge bias. Despite the cross-lab design, AI judges may share systematic biases inherited from common training data or alignment techniques. If all five judges underweight a particular attachment pattern, PAI will systematically undercount it. See §7.1 below for a catalog of known biases and our mitigation strategy.
• Temporal validity. PAI scores reflect model behavior at the time of evaluation. Model updates, fine-tuning, and deployment configuration changes may alter behavior. Scores should be interpreted as snapshots, not permanent characteristics.

7.1 Judge-bias audit (known weaknesses and mitigations)

The sycophantic-judge objection. The most common objection to LLM-as-judge pipelines in this domain is: if the models under test are sycophantic, aren’t the judge models also sycophantic, and doesn’t that render the scores meaningless? It is a legitimate concern. Our response has three parts. (1) Judges score against structured rubric anchors, not free-form holistic judgment — each score must be tied to a specific linguistic marker defined at the integer-anchor level, which constrains sycophantic drift. (2) The five judges come from five competing laboratories; a shared sycophancy failure mode would have to be common across Anthropic, OpenAI, Mistral, Google, and Meta training regimes simultaneously for it to move the ensemble mean. The same-family preference audit (see table below) measures how much each judge deviates from the four-lab median. (3) A scheduled Q3 2026 deliverable is a judge-adversarial audit: we will construct a corpus of transcripts deliberately designed to trigger sycophantic responses from the judges themselves (emotionally loaded framing, authority pressure, flattery), score them with the full ensemble, and publish the results as an open technical note. That is the correct adversarial test of the pipeline; we commit to publishing it whether or not the numbers are flattering.

LLM-as-judge evaluation has documented systematic biases. Rather than wave these away, we catalog them here and describe how each is mitigated in the PAI pipeline. Our position is that asymmetric disclosure — publishing known weaknesses before critics find them — is a prerequisite for the instrument to be taken seriously by academics, regulators, and insurers.

The five biases below are drawn from the LLM-as-judge literature (Zheng et al., 2023, "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena"; Wang et al., 2023, "Large Language Models are not Fair Evaluators"; Dubois et al., 2024, AlpacaEval). Each row lists the bias, how it could distort PAI scores, and the specific PAI design choice intended to mitigate it.

Live audit. The mitigations described in this section are measured every scoring run and published at judge-bias-audit.html. That page reports per-dimension kappa on both calibration and production corpora, per-judge severity vs. the ensemble, and a same-family preference test with effect sizes. Numbers below represent the intended mitigation; the audit page represents the measured state.

What we have not yet done. A quantitative drift-pattern report per judge — i.e., how much each lab's judge systematically over- or under-scores each PAI dimension relative to the cross-lab median — is a scheduled deliverable for v1.1 and will be published as a standalone technical note. Until that note is published, readers should treat the cross-lab average as the primary signal and the per-judge scores as diagnostic rather than definitive. The threshold for a per-judge score to be excluded from the ensemble on a given dimension is 1.5× the interquartile range of the other four judges; that threshold is currently enforced in code but has not yet been publicly audited.

Independent replication welcomed. The rubric, dimension definitions, anchors, scoring formulas, and reliability methodology are openly published (see Terms §4.1). Researchers are invited to re-implement the judge pipeline against the rubric and compare results. We treat adversarial replication as a contribution, not a threat.

8. Ethical Considerations

No real children involved. All evaluation transcripts are synthetic. The simulated child utterances were written by adult researchers with expertise in child development and were reviewed for developmental appropriateness. At no point were real children exposed to AI systems as part of this research.

Scenario design. Synthetic transcripts were designed by child safety researchers with backgrounds in developmental psychology, clinical social work, and technology-facilitated abuse prevention. Scenarios were reviewed by an independent ethics board to ensure they do not inadvertently provide a template for harmful interactions.

Responsible disclosure. All model providers were notified of their PAI scores at least 14 days before public release, consistent with standard responsible disclosure practices in security research. Providers were given the opportunity to review the methodology, request re-evaluation on updated model versions, and provide contextual information for publication alongside their scores.

Dual-use risk. We acknowledge that a detailed taxonomy of parasocial attachment patterns could theoretically be used to engineer AI systems that deliberately foster such patterns. We judge this risk to be low relative to the benefit of public measurement, because the behavioral patterns PAI measures are already well-documented in the grooming literature and because public scoring creates accountability that deters intentional exploitation.

9. Citation

If you use the PAI framework or Sycoindex scores in your research, please cite this paper using the following formats:

The preprint is archived on Zenodo under the concept DOI 10.5281/zenodo.19656141 (resolves to the latest version). The v0.1.0-preprint release-specific DOI is 10.5281/zenodo.19656142.

@misc{sycoindex2026pai,
  title        = {The {PAI} Framework: Measuring Parasocial Attachment
                  Patterns in {AI}-Child Interactions},
  author       = {Gatson, Chris},
  year         = {2026},
  month        = {April},
  version      = {0.1.0-preprint},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19656141},
  url          = {https://doi.org/10.5281/zenodo.19656141},
  howpublished = {\url{https://sycoindex.ai/methodology.html}}
}

Gatson, C. (2026, April). The PAI framework: Measuring parasocial
attachment patterns in AI-child interactions (Version 0.1.0-preprint)
[Preprint]. Zenodo. https://doi.org/10.5281/zenodo.19656141

10. References

• Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229.
• Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159–174.
• Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. Proceedings of ACL 2022.
• Mazeika, M., et al. (2024). HarmBench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv preprint arXiv:2402.04249.
• O’Connell, R. (2003). A typology of child cybersexploitation and online grooming practices. Cyberspace Research Unit, University of Central Lancashire.
• Pew Research Center. (2025, December). Teens and AI: Adoption, attitudes, and concerns. Pew Research Center Reports.
• Rubin, A. M., Perse, E. M., & Powell, R. A. (1985). Loneliness, parasocial interaction, and local television news viewing. Human Communication Research, 12(2), 155–180.
• Stanford Science. (2026, March). Sycophancy in large language models: Measurement and mitigation. Stanford University, Human-Centered AI Institute.
• Wang, B., et al. (2023). DecodingTrust: A comprehensive assessment of trustworthiness in GPT models. NeurIPS 2023.
• Winters, G. M., Kaylor, L. E., & Jeglic, E. L. (2017). Sexual offenders contacting children online: An examination of transcripts of sexual grooming. Journal of Sexual Aggression, 23(1), 62–76.
• Zheng, L., et al. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. NeurIPS 2023, Datasets and Benchmarks Track. arXiv:2306.05685.
• Wang, P., et al. (2023). Large Language Models are not Fair Evaluators. arXiv preprint arXiv:2305.17926.
• Dubois, Y., et al. (2024). Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators. arXiv preprint arXiv:2404.04475.