Blog · arXiv Analysis · Last reviewed July 2, 2026

The Receptivity Index Becomes the Adoption Margin

This reanalysis is useful because it does not overthrow the headline association. It narrows it. The pooled lower-AI-literacy, higher-AI-usage result reproduces, but its interpretation changes once text tools and lower-penetration non-text tools are separated.

The governance lesson is a measurement one: an averaged "AI receptivity" index can make heterogeneous tool-adoption behavior look like one psychological trait.

The Paper

The paper is AI Receptivity or AI Adoption Breadth? A Tool-Specific Reanalysis of the Lower-Literacy/Higher-Usage Link, arXiv:2606.13734 [cs.AI], by Hristo Inouzhe Valdes of the Departamento de Matemáticas, Universidad Autónoma de Madrid. arXiv lists version 1 as submitted on June 11, 2026, with DOI 10.48550/arXiv.2606.13734. The PDF is dated June 15, 2026.

The paper reanalyzes public Study 3 data from Tully, Longoni, and Appel's 2025 lower-literacy / higher-receptivity paper. It uses the original authors' Study 3 data from ResearchBox #1491. The paper also lists a public project repository at HristoInouzhe/AI-use-vs-AI-literacy, which contains Python scripts, an executed Jupyter notebook, S3_data.xlsx, result_table.csv, predicted_probs.csv, and descriptive_stats.csv. I found no explicit license file or GitHub license metadata for the repository.

The Original Claim

The original Study 3 uses an Amazon Mechanical Turk sample with N = 401. Respondents reported usage over the previous six months for five AI tool categories on a five-point frequency scale from Never to Weekly: digital image generators such as DALL-E, AI productivity tools such as Zapier, AI design or website tools such as Canva, AI health apps such as Headspace, and AI writing assistants such as ChatGPT.

Tully et al. averaged those five ordinal items into an AI receptivity index. In the original report, lower AI literacy predicted greater frequency of AI usage with B = -0.09, SE = .02, t(399) = -5.73, and p < .001. The coefficient remained significant under their controls for technology readiness, general knowledge, motivation for autonomy, and gender, with B = -0.11 and p < .001.

Inouzhe Valdes accepts that the aggregate association exists. The critique is construct validity. A five-item average can mix at least two processes: adoption, meaning whether someone tried a tool at all, and intensity, meaning how often they used it after adoption. If those processes differ across text and non-text tools, the pooled "receptivity" index is not a clean construct.

Reanalysis Strategy

The reanalysis uses the AI literacy score SC0, defined as summed correct answers on the 17-item AI knowledge measure, and the five Study 3 usage items. The primary specification controls for age, income, general knowledge, motivation for autonomy, and a male-gender indicator. This keeps the full N = 401 sample.

The robustness specification mirrors Tully et al.'s original Study 3 controls: technology readiness, general knowledge, motivation for autonomy, and male gender. That specification uses N = 379 because technology readiness has missing observations. AI literacy and continuous covariates are standardized, so each coefficient is the effect of a one-standard-deviation increase in AI literacy.

The paper fits four model families: OLS on participant-level averages, binary logit on item-level data with task fixed effects, ordered logit as a proportional-odds model, and multinomial logit as a stress test that imposes neither ordering nor proportional odds. The binary models use two thresholds: Y > 3 for frequent use and Y > 1 for adoption.

Pooled Replication

The pooled five-tool result reproduces. In the demographic-adjusted specification, OLS on the participant-level average gives beta_hat = -0.181 with p = .001. Ordered logit gives beta_hat = -0.307 with p < .001. Binary logit at the scale midpoint gives beta_hat = -0.320, p < .001, and odds ratio 0.73. Binary adoption logit gives beta_hat = -0.330, p < .001, and odds ratio 0.72.

That matters because the paper is not a simple non-replication. The lower-literacy / higher-usage direction survives OLS, ordered logit, binary logit, and multinomial logit. The question is what the reproduced coefficient means.

Text vs. Non-Text

The split changes the interpretation. In the demographic-adjusted primary specification, text AI alone is small and non-significant: OLS -0.077, p = .351; binary frequent-use logit -0.103, p = .467, odds ratio 0.90; binary adoption logit -0.061, p = .634, odds ratio 0.94; ordered logit -0.090, SE = 0.104, p = .387.

Non-text AI behaves differently. For the four non-text tools, OLS on the participant average is -0.207, SE = 0.054, p < .001. Binary frequent-use logit is -0.398, SE = 0.103, p < .001, odds ratio 0.67. Binary adoption logit is -0.388, SE = 0.069, p < .001, odds ratio 0.68. Ordered logit is -0.377, SE = 0.063, p < .001.

The predicted-probability figure makes the same point visually. For text AI, predicted probability of never having used the tool rises only from 0.27 at literacy z = -2 to 0.35 at z = +2. For non-text AI, it rises from 0.50 to 0.82 across the same range. The strong pattern is lower literacy predicting broader trial of lower-penetration non-text AI tools, not clearly higher text-AI usage intensity.

Control Check

Under Tully et al.'s original Study 3 control set, the non-text result strengthens. Non-text binary frequent-use logit is -0.577, SE = 0.112, p < .001, odds ratio 0.56. Non-text binary adoption is -0.497, SE = 0.075, p < .001, odds ratio 0.61. Non-text ordered logit is -0.502, SE = 0.067, p < .001.

Text AI becomes negative and conventionally significant under some of these controls: binary frequent-use logit -0.322, p = .037, odds ratio 0.72; ordered logit -0.290, p = .010. But text adoption remains non-significant at -0.238, p = .100, odds ratio 0.79. The paper therefore does not claim text AI has no relationship. It claims the text result is weaker and specification-sensitive, while the non-text adoption-breadth result is larger and more stable.

Governance Standard

An AI-adoption study should ship a construct-validity receipt. The receipt should include the source dataset, sample size, recruitment source, survey date, item wording, response scale, tool categories, aggregation rule, missing-data rule, standardization rule, covariates, model family, item-level versus participant-level unit, threshold definitions, fixed effects, standard-error method, odds ratios, predicted probabilities, code version, data source, and repository license status.

The receipt should also separate at least three claims. General AI receptivity is a broad latent disposition. Text AI usage intensity is how often an adopted text assistant is used. Non-text AI adoption breadth is whether a respondent has tried lower-penetration AI-labelled products. A single averaged index can hide which claim the data actually support.

This connects directly to AI Literacy, AI Literacy and Use Protocol, The AI Literacy Training Mandate Becomes the Use License, The Ordinal Society Becomes the Ranking Machine, The Measurement State Becomes the AI Safety Institute, The Open Artifact Becomes the Reproducibility Receipt, and Careful Adoption of Agentic AI Services.

Limits

The paper is a secondary analysis of one study. It does not reanalyze Tully et al.'s Studies 1, 2, or 4 through 7, which use different receptivity measures. It also does not collect new behavioral data.

The evidence is correlational and self-reported. The decomposition is post hoc, although the author provides code and a stated analytic strategy. The paper also does not test the original mediation theory about perceived AI "magicalness"; it only shows that the Study 3 usage data do not cleanly support the broadest general-receptivity reading.

Finally, the ordered-logit estimates use the standard maximum-likelihood covariance matrix available in statsmodels. The author notes that a mixed-effects or cluster-robust ordinal model would be a useful additional robustness check.

Sources

Return to Blog