AI Attachment Report | 2026

Methodological
Annex

Study Sample

n=998 young adults (age 18–35) in the US. February 2026.

Missing data were minimal (<1–2%) for key variables.

Weekly AI Use: Responses about AI usage frequency were recoded into a binary variable indicating “weekly or more” usage. Participants reporting AI use several times a week, about once a week, or daily/almost daily were coded as weekly users.
Likert Scale Items: Items capturing attitudes, reliance, or emotional closeness to AI were originally collected on a 5-point scale. For clarity in analysis, scores of 4 or 5 were categorized as “high,” while scores 1–3 were considered “low.”
Relationship Status: The descriptive statistics reported three possible relationship statuses: “Single” (which included “Casual dating”), “In a relationship”, “Married” (which included “long-term relationship”). For regression analysis this variable was recoded as binary: participants who reported being in a relationship, marriage, or long-term partnership were coded as in a “stable relationship”; all other responses were coded as “not in a stable relationship”. Participants who selected the option “Prefer not to say” were reported as missing.
Gender: Gender was recoded into three categories: “Man,” “Woman,” and “missing” for participants who reported non-binary or preferred not to disclose.
Human Replacement Index (HRI): Count score 0–6 capturing substitution behaviors. Scores of at least 1 were used to calculate % of users with at least one substitution behavior.

Chi-square tests examined relationships between categorical variables, such as loneliness, gender, relationship status, and AI-related behaviors.
Logistic regression models were applied for binary outcomes (e.g., weekly AI use or high relational reliance), producing odds ratios and confidence intervals. Because odds ratios can overestimate the effect size when outcomes are common, we converted them into approximate risk ratios. This was done by taking the odds ratio and dividing it by one plus the baseline risk in the reference group times the difference between the odds ratio and one.
In simple terms: Risk Ratio ≈ Odds Ratio ÷ [1 − baseline probability + (baseline probability × Odds Ratio)].
This methodology underpins statements such as weekly emotional AI users being 4x more likely to identify AI as primary support.
Negative binomial regression was used for count outcomes (HRI scores), which were overdispersed. These models provided incidence rate ratios to quantify the association between predictors and relational preference scores. This confirmed that lonely individuals exhibited over 70% higher HRI scores, independent of age, gender, and relationship status.