🤖 AI

AI Receptivity or AI Adoption Breadth? A Tool-Specific Reanalysis of the Lower-Literacy/Higher-Usage Link

Reanalyzing public data from Tully et al. (2025), this study demonstrates that the reported link between lower AI literacy and greater AI receptivity is not a general phenomenon but a specific pattern where lower literacy predicts broader adoption of non-text AI tools, while showing no significant association with text-based AI usage.

Original authors: Hristo Inouzhe

Published 2026-06-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hristo Inouzhe

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a study that claimed, "People who know less about how a car engine works are actually more likely to drive cars." The original researchers found this to be true when they looked at all driving habits combined.

This new paper is like a mechanic pulling up to that same study and saying, "Wait a minute. Let's look closer at what kind of driving they are doing."

The author, Hristo Inouzhe, takes the original data and breaks it down into two very different categories of "driving":

The "Super Highway" (Text AI): Tools like ChatGPT that write for you. These are very popular; almost everyone has tried them, and people use them often.
The "Off-Road Trails" (Non-Text AI): Tools that generate images, build websites, or manage health apps. These are much newer and less familiar; most people have never even tried them.

Here is what the re-analysis found, using simple analogies:

1. The "Average" Was Hiding the Truth

The original study took the answers for all these different tools and averaged them into one big score, like mixing red, blue, and yellow paint to get a muddy brown color. They saw a clear pattern: Lower knowledge = More usage.

But the author argues that mixing these paints together hides the fact that the colors are actually very different. When you separate the "Text AI" paint from the "Non-Text AI" paint, the picture changes completely.

2. The "Text AI" Result: A Flat Road

When looking only at writing assistants (like ChatGPT), the link between low knowledge and high usage almost disappears.

The Analogy: Imagine a flat road. Whether you are a driving expert or a total beginner, you are just as likely to be driving a car on this specific highway. Knowing how the engine works doesn't really change whether you get in the car or not for this specific task.

3. The "Non-Text AI" Result: A Steep Hill

When looking at the other tools (image generators, health apps, etc.), the original pattern is very strong, but with a twist.

The Analogy: Imagine a steep hill. People with low knowledge are much more likely to take a step onto the path and try these tools. People with high knowledge are much more likely to stay at the bottom and say, "I've never tried that."
The Catch: The study shows that low-knowledge people are more likely to try these tools (adoption), but it doesn't prove they use them more often or more intensely once they start. It's like the difference between buying a ticket to a theme park (trying it) versus riding every single rollercoaster five times a day (intensive use). The low-knowledge group is better at buying the ticket, not necessarily at riding the rides more.

The Big Takeaway

The original study claimed: "People who know less about AI are generally more open to using it."

This paper says: "That's only half the story. It's more accurate to say: People who know less about AI are more likely to try out the newer, stranger, less familiar AI tools."

When it comes to the tools everyone already knows and uses (like writing assistants), knowing more or less about AI doesn't really change whether people use them. The "magic" of AI only seems to attract the less-knowledgeable crowd when the tool is something new and unfamiliar, like a picture generator or a health bot.

In short: The original study saw a crowd of people walking toward a building. This paper points out that the crowd is actually two different groups: one group is walking in because they love the building (Text AI), and the other group is walking in because they are curious about a strange, new door they've never seen before (Non-Text AI). The "curiosity" group is the one driven by low knowledge, not the "love" group.

Technical Summary: AI Receptivity or AI Adoption Breadth?

Problem Statement
This paper reanalyzes data from Study 3 of Tully et al. (2025), which reported a counterintuitive negative association between AI literacy and AI receptivity (i.e., lower literacy predicts higher usage). The original study aggregated self-reported usage frequencies across five heterogeneous AI tool categories (image generators, productivity tools, design services, health apps, and writing assistants) into a single continuous index. The authors of this reanalysis argue that this aggregation masks substantial heterogeneity. Specifically, the original methodology conflates two distinct behavioral processes: (1) intensity of use for familiar tools (text AI) and (2) adoption breadth for less familiar tools (non-text AI). Furthermore, the original approach treats ordinal Likert-scale data (1–5 frequency) as metric, a practice methodological literature (Liddell & Kruschke, 2018; Agresti, 2010) suggests can distort effect estimates.

Methodology
The authors re-examined the public data from Tully et al.'s Study 3 ( $N=401$ ) using a more granular, tool-specific approach.

Data: Five usage frequency items (1=Never to 5=Weekly) and an AI literacy score (sum of 17 correct answers).
Model Families: To address the ordinal nature of the data and the aggregation issue, the authors fitted four model families:
1. OLS: On participant-level averages (replicating the original specification).
2. Binary Logit: On item-level data with task fixed effects, testing two thresholds: "frequent use" ( $Y>3$ ) and "adoption" ( $Y>1$ ).
3. Ordered Logit (Proportional-Odds): The primary specification for ordinal Likert responses.
4. Multinomial Logit: A stress test treating categories as nominal without ordering assumptions.
Controls: The primary specification adjusted for age, income, general knowledge, motivation for autonomy, and gender. A robustness check utilized the original study's control set (technology readiness, general knowledge, motivation for autonomy, gender).
Decomposition: The analysis separated the data into "Text AI" (writing assistants) and "Non-Text AI" (the other four categories), recognizing that text AI had a well-populated distribution across all frequency levels, whereas non-text tools were heavily concentrated at the "Never" response.

Key Results

Replication of Aggregate Effect: The authors successfully reproduced the original negative association between AI literacy and aggregate AI usage across all model specifications (OLS, binary, ordered, and multinomial logit).
Heterogeneity by Tool Type: When decomposed, the relationship diverged significantly:
- Text AI: AI literacy was not a significant predictor of text AI usage in the demographic-adjusted specification (Ordered Logit $\beta = -0.090, p = .387$ ). The relationship was weak and specification-sensitive, becoming significant only under the original study's specific control set.
- Non-Text AI: AI literacy was a strong, robust negative predictor of non-text AI usage (Ordered Logit $\beta = -0.377, p < .001$ in the primary specification; $\beta = -0.502, p < .001$ in the original control specification).
Adoption vs. Intensity: The non-text effect is primarily driven by the adoption margin (whether a user has ever tried the tool) rather than the intensity margin (frequency of use). The odds ratio for ever having used a non-text AI tool was 0.68 per standard deviation increase in literacy.
Predicted Probabilities: Visual analysis showed that for non-text tools, higher literacy strongly predicted a higher probability of never having used the tool (rising from 50% to 82% across the literacy range). For text AI, the probability of non-use remained relatively flat.

Key Contributions

Methodological: The paper demonstrates that averaging heterogeneous ordinal items into a single index can obscure distinct behavioral mechanisms. It advocates for modeling ordinal adoption data using ordered-logit or binary-logit specifications rather than OLS on averaged indices.
Substantive: It refines the interpretation of the "lower-literacy, higher-receptivity" claim. The evidence does not support a general tendency for low-literacy consumers to be more receptive to AI across the board. Instead, it points to a specific pattern: lower-literacy consumers are more likely to have tried (adopted) low-penetration, non-text AI tools (e.g., image generators, health apps), but this does not extend to intensive use of high-penetration text tools.

Significance and Claims
The paper claims that Study 3, which measures self-reported past usage rather than stated preferences, does not support the broadest interpretation of Tully et al.'s theoretical claim (that lower literacy increases general AI receptivity due to "magical" perceptions).

Interpretation: The aggregate coefficient is a composite of a weak text-AI intensity effect and a strong non-text-AI adoption effect.
Managerial Implication: The finding narrows the strategic prescription for firms. Targeting low-literacy consumers may be effective for acquiring users for new or low-penetration non-text AI categories, but it does not necessarily predict higher intensity of use or receptivity to established text-based assistants.
Construct Validity: The paper suggests that "AI receptivity" may not be a unitary latent trait but rather a family of tool-specific decisions with different drivers.

The authors conclude that while the aggregate statistical association is real, its substantive meaning changes when decomposed, shifting the focus from a general psychological trait to a specific pattern of adoption breadth across heterogeneous tool categories.

1. The "Average" Was Hiding the Truth

2. The "Text AI" Result: A Flat Road

3. The "Non-Text AI" Result: A Steep Hill

The Big Takeaway

Technical Summary: AI Receptivity or AI Adoption Breadth?

More like this