Gaps in large language model awareness, usage, and perceptions in the United States: Evidence from a nationally representative longitudinal survey

Abstract

Large language models (LLMs) have the potential to benefit users in both their work and personal lives, but which groups are quickest to adopt them? To investigate awareness, usage, and perceptions of LLMs among US adults across socio-demographic groups—and to track changes over time—we administered a two-wave survey using a nationally representative, probability-based online panel of 12,224 US residents. Across two survey waves spanning 1 year, we observed marked gaps in usage: groups more likely to use LLMs included men, younger adults, those with college education and higher incomes, individuals in more analytical occupations (e.g. STEM), Democratic-leaning respondents, and those with above-median cognitive ability, internet literacy, and openness to experience. These usage gaps do not appear to be declining and, in many cases, seem to be widening over time. Our analyses indicate that these disparities are associated with differences in both access-related factors (e.g. income, occupation, digital skills) and individual traits and preferences (e.g. openness to experience, political orientation). Overall, our data provide a dynamic picture of the rapidly evolving exposure to, adoption of, and attitudes toward LLMs in the US population.

Significance Statement.

As large language models transform how people work and learn, understanding who adopts these technologies—and who does not—has critical implications for equity and economic opportunity. Using a nationally representative longitudinal survey of 12,224 adults in the United States, we reveal substantial and, in many cases, widening digital divides. Men, younger adults, those with higher education and incomes, and individuals in analytical occupations are significantly more likely to use large language models, with these gaps persisting or expanding over time rather than narrowing. We also find gaps in people’s reasons for using AI and perceptions of it. Given that persistent gaps in how people use new technologies can exacerbate existing inequalities, the present findings suggest that ongoing research in this area is needed.

Introduction

Advancements in generative artificial intelligence (Gen-AI) are transforming society. In particular, large language models (LLMs)—a type of Gen-AI that processes and generates human-like text and code—promise significant gains in productivity, creativity, and knowledge to users who become proficient with them.^a At the same time, recent evidence points to possible downsides associated with LLM utilization (5–7). Historically, both awareness and usage of new technologies—and their subsequent potential benefits and risks—have varied across subpopulations due to differences in pre-existing attitudes, resources, and access to information about new tools (8, 9). As companies and policymakers seek to harness LLMs’ potential, it is critical to assess who is engaging with these technologies and who is not (10, 11). To that end, this article examines patterns of awareness and use of LLMs across diverse sub-groups of the US population over time, identifying emerging digital divides and their driving factors.

To date, the adoption of LLMs has been rapid but far from pervasive. ChatGPT—the most popular LLM—launched in November 2022 and reportedly reached over 200 million unique users across the world within three months (12, 13). In the United States, a survey by the Pew Research Center ($N\approx 10,000$) found that the fraction of adults who had ever used LLMs rose from 18% in July 2023 to 23% in February 2024 (14). A YouGov poll of online (nonprobability sampled) respondents (15) ($N=\mathrm{2,113}$) found a usage rate of 32% in April 2024, whereas a study based on the Qualtric’s Real-time Population Survey ($N\approx \mathrm{5,000}$) estimated that nearly 40% of the working-age population had used generative AI by November 2024 (16).^b

These studies also found some evidence of gaps in adoption across demographic groups. For example, both (14, 16) documented that the use of generative AI increased with income and education, decreased with age, and that women were less likely to use LLMs than men. An investigation of workers in Denmark showed similar patterns: LLM usage was more common among men, younger individuals, higher-achieving workers, and those with less work experience (19). The gender gap has been identified in international studies as well: a review of 14 datasets (a mix of surveys and web-traffic data) from multiple countries finds that women are less likely to use LLMs than men in nearly all regions (20).

However, many of the aforementioned studies are based on nonrepresentative, convenience samples. This raises concerns about selectivity, which can lead to biased estimates of adoption even after applying demographic weights (21, 22). Specifically, participants who opt into convenience panels may differ from the general population not only on observable characteristics but also on unobservable traits, including motivation, technological savvy, or attitudes toward new technologies, which are directly relevant to the outcomes being studied. Although the Pew Center uses a probability-based sample, its design is repeated cross-sectional, which prevents analyses of individual-level changes in awareness and adoption over time. Moreover, the underlying composition of gaps in usage and the extent to which they persist over time remain unexplored.

To the best of our knowledge, this article presents the first probability-based, longitudinal survey tracking American adults’ awareness and use of LLMs. We completed two rounds of data collection (one each in 2023 and 2024) within the Understanding America Study (23), a nationally representative, probability-based online panel of ∼15,000 US residents. Our study has several advantages over existing research. First, its probability-based sampling enables broader generalization compared to studies relying on convenience samples, allowing us to make more accurate inferences about the entire US population. Second, the availability of rich background data about participants permits a deeper analysis of how a range of individual traits that have previously been linked to technology adoption—including cognitive ability and personality traits (24–26)—determine LLM awareness, usage, and adoption. By simultaneously accounting for these individual traits alongside demographics, socioeconomic status, occupation, political preferences, and urbanicity, we can identify the independent predictive power of each of these factors. Third, exploiting the panel structure of our data, we can gauge how such gaps evolve over time. Finally, beyond examining gaps in awareness, usage, and adoption of LLMs, we also document differences in reported reasons for using LLMs and beliefs about how useful or harmful LLMs are.

We document substantial sociodemographic disparities in LLM awareness and usage. Both are more likely among younger individuals, those with higher education and socioeconomic status, people in cognitively demanding jobs, and respondents with greater cognitive ability, stronger internet skills, and a greater openness to experience. In line with previous research, we also find substantial gender and racial disparities in LLM usage: women are less likely to use LLMs than men, Black respondents are less likely than White respondents, and Asian respondents are more likely than White respondents. We benchmark these gaps against those from different occupations and personal characteristics (like political affiliation, personality traits, cognitive abilities, and Internet skills). Moving beyond prior work, we quantify the extent to which observable individual characteristics contribute to the observed demographic disparities. We estimate that differences in socioeconomic factors and personal attributes explain up to half of the racial gaps, but only about one-third of the gender gap.

Importantly, we find that many of the same factors that predict usage in the population also predict new adoption among prior nonusers. In fact, none of the gaps we observe narrowed between waves. Although the frequency of LLM usage among users remained stable, the reasons for use shifted. High-frequency users increasingly reported using LLMs for work-related tasks, signaling that this group is beginning to integrate LLMs into their professional routines. In contrast, low-frequency users continued to cite curiosity or entertainment as the main reasons for use. This emerging divergence may signal a growing gap in who stands to benefit most from LLM adoption, as well as who may be more exposed to its potential downsides.

Data and methods

Sample

We developed a longitudinal survey to measure the use of AI (generative or otherwise) in the US population, and administered it through the Understanding America Study (UAS) panel in Fall 2023 (wave 1, September–October 2023) and Spring 2024 (wave 2, April–July 2024). The UAS is a probability-based online panel comprising ∼14,000 US residents recruited through address-based sampling.^c This mitigates most selection biases associated with opt-in or convenience samples and ensures relatively comprehensive population coverage. Once enrolled in the panel, UAS members are invited to participate in surveys two to three times per month. A set of 22 core surveys, collecting information on employment, health, cognitive ability, digital literacy, personality, and psychological traits, is administered to the entire panel every 2 years. This approach provides us with rich background data for all participants, which we integrate with our AI surveys.

A total of 12,224 UAS panel members completed at least one wave of our AI survey. The number of respondents was 9,936 in wave 1 and 11,162 in wave 2. For both waves, the conditional response rate—that is, the share of active UAS members who completed the survey—was 76%, while the unconditional response rate, obtained by accounting for the panel’s recruitment (empanelment) rate, was ∼11%. After merging data on occupation, political preferences, cognitive ability, Internet literacy, and personality traits from other UAS surveys, and excluding respondents with missing information on key variables, our final analytic sample comprises 8,754 individuals in wave 1, 8,964 in wave 2, and 8,121 who completed both waves. The demographic composition of our analytic sample closely resembles that of the full UAS sample, with modest under-representation of Black, Hispanic, and younger respondents (aged 18–29). To ensure population representativeness and generate population-level estimates, we apply sample weights throughout our analysis. These weights account for differential inclusion probabilities of population groups implied by the UAS sampling procedure and adjust the distributions of gender, race, age, education, and geographic location (as defined by Census regions) to match those of the broader US population.^d

Key survey questions and variables

Below, we describe some of the key questions in our AI survey and variables from UAS data that we integrate into our analyses. Further details are provided in the Appendix.

Survey questions

Q1. Awareness and usage of LLMs.

Participants were asked, “Have you heard about or used AI applications that create human-like text or code, such as ChatGPT, Bard/Gemini, or Bing Chat?”^e They could select one of four responses: “I have never heard about them,” “I have heard about them but never used one,” “I have used them,” or “I don’t know.”

Q2. Frequency of usage.

Participants who reported having used these tools (henceforth, “users”) were asked to indicate how frequently they used them. They could select one of five responses: “Rarely,” “Less than once a month,” “Once or a few times per month,” “Once or a few times per week,” or “Every day.”

Q3. Reasons for using LLMs.

Users were asked: “What do you use AI applications such as ChatGPT, Bard/Gemini, or Bing Chat for? Please select all that apply.” They could choose one or more of fourteen options, including “out of curiosity,” or “for work-related tasks.”^f

Q4. Perceived usefulness or harm.

Users were asked “How useful do you find AI applications such as ChatGPT, Google Bard/Gemini, or Bing Chat? Please indicate on a 5-point Likert scale ranging from “not at all useful” to “extremely useful.” Next, they were asked an identical question about the perceived harmfulness of these tools (“How harmful do you find…”).

Study outcomes

The responses to questions 1 to 4 serve as our main outcomes of interest. Based on responses to question 1, we construct a 4-item categorical variable that encodes individual engagement with LLMs as never heard, heard but not used, used, and do not know.

Frequency of usage (question 2) is grouped into three categories, with respondents classified as “high-frequency users” if they use LLMs every day or once or a few times per week, “medium-frequency users” if they use them once or a few times per month, and “low-frequency users” if they use them rarely or less than once a month. For the regression analysis, we also created an indicator equal to 1 if the individual is a high-frequency user, and 0 otherwise.

Based on responses to question 3, we constructed nonmutually exclusive indicators equal to 1 if the user reported each specific reason for usage, and 0 otherwise. Finally, responses to question 4 were used to create two indicators capturing whether the respondents perceived LLMs as “very” or “extremely useful” and “very” or “extremely harmful,” respectively.

Covariates

Most of the covariates are a straightforward encoding of participants’ survey responses. Those whose construction is less obvious are described below:

Occupation.

Respondents were asked to select their occupation from the top tier of Standard Occupation Codes (SOCs; see https://www.bls.gov/soc/), and then presented with more detailed tiers based on their choice. As missing data increase with each level, we rely on first-tier SOCs in the analysis to maximize sample size. We also include separate indicators for individuals who have never worked, did not complete the SOC survey, or were unable to match their occupation to a first-tier SOC.^g

Big-5 personality traits.

Respondents rated themselves on 44 items, which were used to create measures of the Big-5 personality traits (29), namely: openness to experience, conscientiousness, extroversion, agreeableness, and neuroticism. For our analyses, individuals are classified as scoring at or above the sample median on each trait vs. below it.

Cognitive tests.

Respondents answered 15 quiz questions (with dichotomous correct/incorrect choices) within each of these three modules: Number Series, Picture Vocabulary, and Verbal Analogies. The questions were derived from the Woodcock-Johnson IV battery of cognitive tests (30), which aims to assess quantitative reasoning and verbal abilities. For each module, we classified individuals as scoring at or above the sample median vs. below.^h

Advanced Internet skills.

Respondents rated their proficiency in seven online tasks—such as writing or commenting online, or downloading apps and software—using a Likert scale (see the Appendix for the full list). A composite index was created by summing the respondents’ self-ratings across these items, with higher values indicating greater proficiency in advanced internet tasks. For our analyses, we classify individuals as scoring at or above the sample median vs. below.

Political affiliation.

Respondents were asked with which political party they most closely identified. Possible responses were Democratic, Republican, Independent, Libertarian, Green, some other party, or not aligned with any party. Those who identified as Independent or not aligned were then asked if they leaned toward the Democratic party, the Republican party, or neither. We use these responses to construct a categorical variable with four groups: (i) Democrat or leaning Democrat, (ii) Republican or leaning Republican, (iii) Independent or unaligned, and (iv) Other (Green party, Libertarian, etc.).

Analytic strategy

We conducted four sets of analyses.

Analysis 1 presents descriptives on the distribution of individuals across the four categories of LLM awareness and usage (“Don’t know,” “Never heard,” “Heard but not used,” or “Used”) by demographic, occupational, and personal characteristics as well as across waves.

Analysis 2 examines the drivers of different forms of engagement with LLMs, including awareness and usage (analysis 2a), usage frequency (analysis 2b), and adoption between waves (analysis 2c). We estimate Multinomial Logit models for multivariate outcomes (analyses 2a and 2b) and standard Logit models for binary outcomes (analysis 2c).

Analysis 3 investigates the two largest demographic disparities in usage, those arising along the gender (analysis 3a) and race (analysis 3b) dimensions. We apply nonlinear decomposition techniques suitable for binary outcomes (31) to quantify the contribution of observable factors to these gaps.

Analysis 4 explores two additional domains relevant to individuals’ interaction with LLMs: their reasons for using LLMs (analysis 4a) and their perceptions of LLM usefulness and harmfulness (analysis 4b). For both outcomes, we present unconditional and conditional analyses.

We use survey weights throughout the analysis to produce population-representative estimates. Standard errors are clustered at the respondent level unless otherwise noted.

Analysis 1: patterns of LLM awareness and usage

Figure 1 presents an overview of LLM awareness and usage across waves 1 and 2. Usage increased over time: In Fall 2023, 17.8% of respondents reported having used LLMs, and 68.5% had either heard of or used them. By Spring 2024, these figures rose to 23.6% and 72.3%, respectively.

Fig. 1.

Awareness and usage of LLMs by wave. $N=\mathrm{8,754}$ in wave 1 and 8,964 in wave 2. Sample weights are used to ensure population representativeness.

Figure 2 further segments differences in LLM use by participants’ sociodemographic characteristics (panel 1), occupations (panel 2), and personal traits (panel 3).ⁱ As shown in the first panel of Fig. 2, LLM usage varies significantly by sociodemographic characteristics. Focusing on wave 2, we observe a substantial gender gap, with 27.2% of men reporting having used LLMs compared to 20.0% of women. Racial differences are also pronounced: 46.3% of Asians have used LLMs, compared to 24.5% of Hispanics, 22.6% of Whites, and 15.2% of Black respondents. Usage declines by approximately 7 percentage points (pp) with each additional decade of age and rises sharply with both education and income: 10.6% of respondents with a high school diploma or less report having used LLMs, compared to 19.2% among those with some college, and 39.5% among those with a bachelor’s degree or higher. Similarly, 38.6% of respondents with household incomes above 100,000 USD report LLM usage, nearly twice the rate of those with household incomes between 60,000 and 99,999 USD. These usage patterns across sociodemographic factors are already present in wave 1.

Fig. 2.

Percentage of US residents who had used LLMs at least once at the time of waves 1 and 2; means by various subgroups. Sample weights are used to ensure population representativeness. $N=8,754$ in wave 1 (Sep–Oct 2023) and 8,964 in wave 2 (Apr–Jul 2024).

As shown in the second panel of Fig. 2, LLM usage also varies significantly by respondents’ occupation. Adoption rates are generally higher—and increasing more rapidly—among individuals employed in fields involving cognitive, analytical, or creative tasks, with 60.4% of those working in Computer and Mathematics reporting having used LLMs by wave 2, compared to 49.0% of those in Legal occupations, and 44.0% of those in Arts, Entertainment, and Sports (44.0%). In contrast, usage is much lower among respondents employed in manual labor-intensive occupations, such as Installation and Maintenance (13.1%), Production (11.7%), and Building and Grounds Cleaning (10.1%).

Finally, the third panel of Fig. 2 examines LLM usage by political affiliation and individual traits. Respondents affiliated with smaller political parties (Green, Libertarian, or others) report the highest usage rates (37.7% in wave 2), followed by Democrats or those leaning Democrat (27.9%), Independents (23.7%), and Republicans or those leaning Republican (16.2%). Usage also increases with cognitive ability, and varies across the Big-5 personality traits: it is higher among individuals with greater openness to experience, but lower among those with higher conscientiousness and agreeableness. Internet skills emerge as the strongest predictor of LLM usage, with a difference of over 25 pp between respondents whose advanced Internet skill score is above the sample median (37.0%) and those with a score below the sample median (11.6%) in wave 2, up from a 21 pp difference in wave 1.

Analysis 2: determinants of LLM engagement

To assess how various factors independently predict awareness, usage, usage frequency, and adoption of LLMs, we pool data across the two waves and estimate multivariate regressions, controlling for the joint influence of potentially correlated variables.

2a: predictors of LLM awareness and usage

The first three columns of Table 1 show estimated marginal effects from a Multinomial Logit model, where the dependent variable captures respondents’ self-reported exposure to LLMs, classified into three categories: “Never heard of LLMs,” “Heard, not used LLMs,” and “Used LLMs.”

Table 1.

Predictors of LLM awareness and usage, usage frequency, and adoption.

	Awareness and usage			Frequency	Adoption
	(1)	(2)	(3)	(4)	(5)
	Never heard of LLMs	Heard, not used LLMs	Used LLMs	High-frequency user	W1 nonuser to W2 user
Wave = 2	−3.6a (0.7)	−2.1b (0.8)	5.6a (0.5)	3.7c (1.5)
Female	7.3a (1.1)	−1.1 (1.3)	−5.3a (1.0)	−7.6a (2.2)	−1.7 (1.1)
Race (omit: White)
Black	2.9 (1.7)	−1.8 (2.1)	−1.8 (1.6)	7.2 (4.8)	−2.2 (1.6)
Asian	−7.6a (2.2)	−3.0 (2.6)	11.1a (2.0)	10.5b (3.7)	0.5 (2.4)
Other	−1.1 (2.2)	−0.7 (2.8)	1.8 (2.2)	−6.1 (3.5)	−1.4 (2.4)
Hispanic	−0.4 (1.5)	−0.3 (1.8)	0.3 (1.3)	1.1 (3.2)	0.4 (1.5)
Age (omit: 18–29)
30–39	5.8a (1.6)	2.1 (2.1)	$-{9.3}^{\mathrm{a}}$ (1.8)	2.3 (3.0)	−7.4b (2.5)
40–49	8.1a (1.7)	3.7 (2.2)	$-{13.2}^{\mathrm{a}}$ (1.9)	−3.3 (3.4)	$-{9.6}^{\mathrm{a}}$ (2.6)
50–59	12.5a (1.8)	4.7c (2.3)	$-{18.2}^{\mathrm{a}}$ (1.9)	−1.0 (3.8)	$-{10.8}^{\mathrm{a}}$ (2.6)
60$+$	16.6a (1.8)	8.1a (2.3)	−24.0a (1.9)	−8.0c (3.4)	$-{14.7}^{\mathrm{a}}$ (2.6)
Education (omit: High school or less)
Some college	$-{4.8}^{\mathrm{a}}$ (1.2)	2.5 (1.5)	3.5b (1.2)	−2.4 (3.2)	2.7c (1.3)
Bachelors or more	$-{11.3}^{\mathrm{a}}$ (1.4)	5.0b (1.8)	8.8a (1.4)	3.5 (3.4)	4.2b (1.4)
Household income (omit: $30,000 or less)
30,000–59,000 USD	1.2 (1.4)	1.6 (1.7)	−0.7 (1.4)	−2.0 (3.9)	−3.4c (1.7)
60,000–99,000 USD	0.1 (1.5)	3.2 (1.9)	−0.9 (1.4)	−7.7c (3.9)	−1.7 (1.8)
100,000 USD or more	−4.3b (1.7)	2.5 (2.0)	4.6b (1.5)	−3.9 (3.8)	−0.3 (1.9)
Location (omit: Metropolitan)
Micropolitan	1.2 (1.3)	−0.0 (1.7)	−1.1 (1.3)	1.4 (3.2)	−0.7 (1.4)
Small town/Rural	3.3c (1.6)	−1.1 (2.0)	−1.2 (1.7)	7.1 (5.5)	−0.8 (1.9)
Unknown	−3.8 (2.3)	0.1 (2.9)	2.6 (2.3)	−0.7 (4.1)	−2.5 (2.3)
Working	−0.2 (1.2)	−2.6 (1.4)	2.9b (1.1)	−1.5 (2.8)	2.6c (1.3)
Politics (omit: Democrat, reg. or lean)
Republican, reg. or lean	4.7a (1.3)	−0.3 (1.5)	$-{4.9}^{\mathrm{a}}$ (1.2)	−7.0c (2.8)	−2.9c (1.3)
Independent or unaligned	2.5c (1.2)	−2.3 (1.4)	−1.3 (1.1)	−3.1 (2.5)	−1.2 (1.3)
Green party, Libertarian or other	−9.0b (2.8)	4.0 (3.1)	3.1 (2.1)	−0.4 (4.0)	−2.3 (2.3)
Median or above internet skills	$-{9.0}^{\mathrm{a}}$ (1.0)	0.3 (1.3)	10.5a (1.0)	4.4 (2.3)	4.4a (1.1)
Median or above cognition score in:
Numeric module	−3.6b (1.1)	2.2 (1.4)	4.2a (1.0)	−5.0c (2.5)	0.7 (1.2)
Picture & Vocab module	$-{4.3}^{\mathrm{a}}$ (1.2)	4.5b (1.4)	1.8 (1.1)	−1.8 (2.5)	−0.6 (1.2)
Verbal analogy module	−2.7c (1.1)	3.8b (1.4)	1.3 (1.0)	$-{8.6}^{\mathrm{a}}$ (2.5)	1.6 (1.2)
Median or above on personality trait:
Agreeableness	3.4b (1.1)	−1.4 (1.3)	−1.4 (0.9)	−0.6 (2.2)	−0.1 (1.1)
Conscientiousness	0.8 (1.1)	1.8 (1.3)	−1.2 (1.0)	−1.0 (2.2)	−0.4 (1.1)
Extroversion	2.3c (1.0)	−3.0c (1.2)	0.6 (0.9)	1.4 (2.1)	0.3 (1.1)
Neuroticism	1.4 (1.1)	−1.7 (1.3)	0.3 (1.0)	−1.0 (2.3)	1.2 (1.1)
Openness	$-{6.1}^{\mathrm{a}}$ (1.0)	3.0c (1.2)	4.2a (0.9)	3.0 (2.1)	1.8 (1.1)
Heard, not used, LLM in Wave 1					6.7a (1.4)
N	17,332	17,332	17,332	3,826	6,173
Standard Occupation Code (omit: Sales)
Management	−3.9 (2.9)	−3.1 (3.5)	6.7c (2.7)	11.2 (6.7)	4.4 (3.0)
Business & Financial operations	−2.6 (3.0)	−0.5 (3.6)	1.3 (2.7)	−7.0 (5.9)	2.4 (2.9)
Computer & Math	−14.4a (3.3)	−1.8 (4.3)	14.2a (3.2)	5.4 (6.1)	5.0 (3.7)
Architecture & Engineering	−4.1 (4.9)	9.9 (5.4)	−4.3 (3.1)	−5.9 (7.2)	−1.4 (3.1)
Life, Physical & Social sciences	−6.3 (5.8)	1.2 (6.7)	3.5 (3.9)	−0.0 (9.4)	5.8 (5.2)
Community & Social service	3.4 (4.0)	−3.2 (4.8)	−2.7 (4.0)	3.9 (10.9)	0.9 (3.9)
Legal	−6.1 (4.5)	−9.3 (5.2)	16.9a (4.4)	0.8 (7.3)	9.7 (6.1)
Educational instruction & Library	0.6 (3.1)	−5.1 (3.4)	4.0 (2.5)	−2.5 (5.9)	11.5 a (3.1)
Arts, Entertainment, Sports & Media	−3.0 (4.6)	−0.5 (5.0)	4.9 (3.4)	0.5 (7.5)	6.2 (3.7)
Healthcare practitioners & Technical	−0.5 (3.2)	2.1 (3.7)	−2.7 (2.7)	−11.0 (6.3)	−1.1 (2.5)
Healthcare support	2.8 (3.0)	−3.4 (3.7)	−2.2 (2.9)	−4.6 (7.3)	−4.4c (2.2)
Protective service	9.2 (5.3)	1.4 (5.7)	−11.3 a (3.1)	−17.2c (6.8)	−5.6c (2.3)
Food preparation & serving related	−2.2 (3.2)	3.4 (3.9)	−3.7 (3.1)	−9.2 (7.3)	−1.0 (3.0)
Building, Cleaning & Maintenance	−0.6 (4.2)	5.4 (5.5)	−5.7 (4.5)	10.4 (16.7)	−2.1 (3.7)
Personal care & service	−5.0 (4.6)	5.6 (5.8)	1.2 (5.3)	−7.4 (11.8)	3.7 (5.9)
Office & Administrative support	−5.7c (2.8)	7.4c (3.5)	−1.8 (2.7)	−6.7 (6.6)	2.6 (2.7)
Farming, Fishing & Forestry	−2.6 (5.3)	4.6 (5.9)	−3.9 (5.4)	−18.3c (7.9)	9.5 (7.6)
Construction & Extraction	5.1 (4.7)	−6.0 (4.9)	−1.6 (4.0)	−11.2 (7.5)	4.8 (5.1)
Installation, Maintenance & Repair	−2.7 (3.9)	5.7 (4.9)	−4.3 (3.6)	1.5 (11.6)	0.7 (4.0)
Production	5.4 (3.4)	0.1 (4.8)	−6.6 (3.9)	5.1 (12.6)	−0.5 (3.9)
Transportation & Material moving	−2.4 (3.5)	3.8 (4.5)	−0.0 (3.6)	−10.6 (7.9)	3.6 (4.0)
Military specific	−8.5 (10.4)	5.7 (10.4)	1.2 (5.3)	25.7c (12.8)	10.5 (7.8)
Never worked	2.2 (2.6)	−5.0 (3.3)	0.2 (2.7)	3.6 (7.3)	3.1 (2.9)
Occ. not stated	−7.6b (2.4)	3.0 (3.0)	3.7 (2.3)	−3.2 (5.4)	5.8c (2.5)
Occ. not on the list	0.5 (2.4)	−0.6 (2.8)	−1.5 (2.1)	−0.4 (5.5)	2.1 (2.2)
N	17,332	17,332	17,332	3,826	6,173

Standard errors in parentheses, clustered at the respondent level. ${ }^{\mathrm{a}}P<0.001$, ${ }^{\mathrm{b}}P<0.01$, ${ }^{\mathrm{c}}P<0.05$.

Both awareness and usage of LLMs increased between waves 1 and 2, with the share of respondents reporting LLM use rising by 5.6 pp ($P<0.001$), and the share who had never heard of them declining by 3.6 pp ($P<0.001$).

Women were less likely than men to have used LLMs and more likely to be unaware of them (5.3 pp, $P<0.01$, and 7.3 pp, $P<0.001$, respectively). After controlling for observable factors, the racial and ethnic gaps observed in Fig. 2 are no longer statistically significant, except for Asians, who remain 11.1 pp ($P<0.001$) more likely than the reference group (Whites) to have used LLMs and 7.6 pp ($P<0.001$) more likely to be aware of them.

The monotonic decline in LLM usage with age documented in Fig. 2 is also apparent in the multivariate analysis. Compared to 18- to 29-year-olds, 30- to 39-year-olds are 9.3 pp ($P<0.001$) less likely to use LLMs, with the gap widening to 24.0 pp ($P<0.001$) among those aged 60 and older. Similarly, awareness declines with age, with a 16.6 pp ($P<0.001$) gap between those aged 60+ and their youngest counterparts (18–29 year olds).

Awareness and usage increase with each education level and are higher among respondents with household incomes above $100,000.

Political affiliation and personal traits are independent predictors of LLM usage. Republican-leaning respondents are 4.9 pp ($P<0.001$) less likely than Democrat-leaning ones to use LLMs and 4.7 pp ($P<0.001$) more likely to be unaware of them. Internet skills remain a strong predictor of LLM usage (10.5 pp, $P<0.001$) and also of awareness, with higher-skilled respondents 9.0 pp ($P<0.001$) less likely to be unaware of LLMs. Higher scores across all cognitive modules are associated with greater awareness of LLMs, but only performance on the number series task is significantly associated with usage. Among the Big-5 personality traits, higher openness to experience is associated with a 4.2 pp ($P<0.001$) increase in usage and a 6.1 pp ($P<0.001$) increase in awareness. Higher agreeableness and extroversion are associated with lower awareness but show no statistically significant association with usage.

Finally, the regression results suggest that part of the occupational gradient in LLM usage observed in Fig. 2 reflects differences in other variables correlated with occupation, as only some occupational differences remain statistically significant after adjusting for additional covariates. Compared to the reference category (Sales), LLM usage is 16.9 pp ($P<0.001$) higher among users in Legal occupations, 14.2 pp ($P<0.001$) higher in Computer & Math, and 6.7 pp ($P<0.05$) higher in Management. In contrast, usage is significantly lower among those in Protective Services. Respondents employed in Computer & Math are also more likely to be aware of LLMs; employment in Office & Administrative Support is associated with increased awareness but not usage.

2b: predictors of usage frequency

As shown in Fig. 3, most users engage with LLMs relatively infrequently. Approximately half report being low-frequency users (using LLMs rarely or less than once a month), one-quarter are medium-frequency users (a few times a month), and one-quarter are high-frequency users (a few times per week or daily). Usage frequency appears to be stable for the majority of users, with 62% of respondents remaining in the same usage frequency category (low/med/high) in wave 2 as in wave 1.

Fig. 3.

Within-individual changes in usage frequency between wave 1 and wave 2. “High frequency users” use LLMs daily or within the last week. “Medium frequency users” use LLMs a few times a month, and “Low frequency users” use LLMs once a month or less. Block heights are scaled to sample size (N).

Column 4 of Table 1 reports marginal effects of sociodemographic and individual characteristics on the likelihood of frequent LLM use, estimated from a Logit model using pooled data on LLM users from both waves.

Several of the factors previously found to be associated with LLM usage also predict usage frequency. For example, women are 7.6 pp ($P<0.001$) less likely than men, and Asian respondents are 10.5 pp more likely than the reference racial group (Whites) ($P<0.01$), to be frequent users. High frequency usage is also less likely among individuals in the oldest age category, those leaning Republican, and those scoring lower on the numeric and verbal analogy modules of the cognition test.

2c: predictors of LLM adoption among wave 1 nonusers

The longitudinal design of our survey enables us to examine the factors that predict LLM adoption by Spring 2024 among individuals who did not use LLMs in Fall 2023. This analysis addresses the critical question of whether previously identified usage gaps across groups are narrowing or widening over time. We examine this question in column 5 of Table 1, which reports marginal effects from a Logit model predicting adoption in wave 2 among respondents who were nonusers in wave 1.

We find no evidence of narrowing gender or racial gaps, as the gender and race indicators are not statistically significant. Meanwhile, age and socioeconomic gaps appear to be widening. The probability of LLM adoption in wave 2 among nonusers in wave 1 decreases progressively with each older age category; individuals with higher education levels and those currently employed are more likely to adopt LLMs between waves.

Republican-leaning respondents, who were less likely than their Democrat-leaning counterparts to use LLMs in wave 1, are also less likely to adopt them over time (−2.9 pp, $P<0.05$). Internet skills continue to play a role, with individuals scoring above the sample median more likely to become new adopters (4.4 pp, $P<0.001$).

Adoption is also positively associated with prior awareness of LLMs. Among wave 1 nonusers, those who had heard of LLMs are 6.7 pp ($P<0.001$) more likely to use them in wave 2 than those who had not.

Finally, there is limited evidence that adoption rates vary across occupations. Relative to those employed in Sales (the reference category), adoption is more likely among those in Educational Instruction & Library, and less likely among those in Healthcare Support and Protective Services.

Analysis 3: decomposition of gender and racial gaps in LLM usage

As previously illustrated in Fig. 2, there are sizable unconditional differences in LLM usage across gender and race between Blacks and Whites and Asians and Whites. After accounting for observable individual characteristics (as shown in Table 1), those gaps narrow, and the one between Black and White respondents is no longer statistically significant. In this section, we use Fairlie decomposition (31) to assess the extent to which observable differences between gender and racial groups explain disparities in LLM usage.

3a: decomposition of the gender gap

In the pooled sample, women are 7.6 pp ($P<0.001$) less likely than men to report LLM use. The first panel of Fig. 4 explores how differences in observable characteristics between men and women affect this gap. Negative (positive) contributions of a given variable indicate that equalizing that variable across genders would reduce (increase) the usage gap.

Fig. 4.

Decomposing gender and racial gaps in LLM usage. The contributions of individual categories (e.g. Black, Hispanic, Asian, Other) are estimated separately but shown combined in the figure.

The largest contributor to the gender gap is the difference in scores on the numerical module of the cognitive test. Since these scores are positively associated with LLM use and lower on average for women, equalizing them would reduce the gap by 1.68 pp ($P<0.001$). Lower education and household income among women account for an additional 0.88 ($P<0.001$) and 0.67 pp ($P<0.01$), respectively, with smaller contributions from differences in racial composition, Internet skills, and openness.

These effects are partially offset by age differences. Women in the sample are, on average, younger than men, and LLM use declines with age. Therefore, equalizing age across genders in the sample would increase the usage gap by 2.74 pp ($P<0.001$).

The sum of all contributions is 1.87 pp ($P<0.05$), explaining roughly one-third of the gender gap in usage. The remaining two-thirds reflect factors not captured by the observables in our analysis.

3b: decomposition of racial gaps

We now turn to assessing LLM usage gaps for Black and Asian respondents relative to Whites (no usage gaps were observed for Hispanics or individuals in the “Other” race category in the raw data). Unconditionally, Blacks are less likely to use LLMs than Whites (the usage gap is −5.91 pp, $P<0.001$), whereas Asians are substantially more likely to use them (27.42 pp, $P<0.001$). The right panel of Fig. 4 assesses the contributions of observable characteristics to these gaps.

For Black respondents, the usage gap relative to Whites is partly explained by lower education levels (−1.55 pp, −1.04 pp, −0.76 pp, $P<0.05$). In addition, smaller negative contributions come from being less likely to be female, working in occupations with lower LLM exposure, having lower Internet skills, and lower employment rates.

In contrast, Asian respondents display higher scores on these same dimensions relative to Whites. Higher educational attainment increases the gap by 2.69 pp ($P<0.001$), and greater representation in occupations with high LLM exposure, particularly technical fields such as Computer and Mathematics, contributes 2.72 pp ($P<0.001$). Additional positive contributions come from higher household income, higher scores on the numerical module of the cognitive test, lower likelihood of being female, higher Internet skills, and higher employment rates.

Political affiliation and age also contribute to these racial gaps. Both Black and Asian respondents are more likely than Whites to lean Democratic and less likely to identify as Republican. Since Democratic affiliation is positively associated with LLM use and Republican affiliation is negatively associated with it, this reduces the Black-White gap and increases the Asian-White gap. Moreover, both groups are younger in our sample, on average, than White respondents, which further reduces the Black-White gap while increasing the Asian-White gap.

Overall, differences in observable characteristics explain approximately half (3.03 pp, $P<0.01$) of the Black-White gap and half of the Asian-White gap (13.04, $P<0.001$) in LLM usage. The remaining portions of both gaps remain unexplained.

Analysis 4: reasons for use and perceptions of LLMs

4a: reasons for using LLMs

The most commonly reported reasons for using LLMs, shown in descending order of prevalence in the first panel of Fig. 5 (“All Users”), are curiosity, work-related tasks, entertainment, improving communication, learning about the world, personal tasks, creative writing, and looking for health-related information. The frequency with which these reasons were reported are largely stable over time, except for curiosity, which saw a significant decline of 9 pp ($P<0.05$), and health-related info, which increased by 4 pp ($P<0.05$).

Fig. 5.

The percentage of LLM users who report using LLMs for each reason (respondents could choose more than one). Values for the first panel (“All”) are cross-sectional: that is, X $\to$ Y implies X% of wave 1 users and Y% of wave 2 users cited that reason. Values for the remaining three panels are constructed after restricting data to users who report low-, medium-, or high-frequency of usage in both waves. Respondents whose usage frequency changes between waves are excluded from those three panels.

The next three panels of Fig. 5 indicate that the reasons for LLM usage are closely tied to usage frequency. High-frequency users are most likely to use LLMs for work-related tasks or to improve communication. The share of high-frequency users citing work reasons for LLM usage increases from 68% to 76% between waves, whereas the share citing curiosity declines from 53% to 41%. Conversely, low-frequency users primarily use LLMs out of curiosity, with over 80% reporting this reason in both waves.

Table 2 shows Logit marginal effects of demographics and individual characteristics on the likelihood of reporting each of the four most common reasons for LLM usage. Women are less likely than men to use LLMs out of curiosity (−10.3 pp, −8.7 pp, $P<0.05)$, but more likely to use them to improve communication (7.0 pp, $P<0.001$). Black, Asian, and Hispanic respondents are more likely than Whites to use LLMs to improve communication, and Asians are 9.2 pp ($P<0.01$) more likely to use them for work-related tasks.

Table 2.

Predictors of the top four reasons for using LLMs.

	(1)	(2)	(3)	(4)
	Out of curiosity	Entertainment	Work tasks	Improve communication
Wave = 2	−10.3a (2.4)	−5.4b (2.5)	4.1 (2.4)	3.7 (2.4)
Female	−10.3a (2.7)	−8.7c (2.7)	1.6 (2.6)	7.0c (2.6)
Working	−8.0b (3.1)	−2.4 (3.6)	27.3a (3.2)	2.4 (3.2)
Race (omit: White)
Black	−7.0 (5.3)	−4.5 (4.9)	10.2 (5.3)	23.5a (5.5)
Asian	6.1 (3.8)	−4.4 (4.3)	9.2b (4.2)	13.3c (4.1)
Other	12.2b (5.1)	3.4 (6.0)	−2.5 (5.2)	1.9 (4.9)
Hispanic	−2.0 (3.7)	−0.4 (3.9)	8.0b (3.7)	8.1b (3.8)
Age (omit: 18–29)
30–39	2.1 (4.2)	−2.4 (4.4)	−4.7 (4.3)	−7.0 (4.2)
40–49	4.2 (4.4)	4.5 (4.8)	−13.6c (4.5)	−10.0b (4.4)
50–59	−5.3 (4.8)	−12.1b (4.9)	−6.0 (4.9)	−7.7 (4.7)
60+	−3.7 (5.0)	−13.6c (5.1)	−15.6c (5.2)	−7.8 (5.0)
Education (omit: High school or less)
Some college	−3.1 (4.1)	0.4 (4.5)	−1.0 (4.6)	−0.5 (4.3)
Bachelors or more	−2.9 (4.0)	−5.4 (4.4)	13.4c (4.5)	4.1 (4.2)
Household income (omit: $30,000 or less)
30,000–59,000 USD	12.4c (4.7)	3.5 (4.7)	−2.0 (4.8)	1.2 (4.2)
60,000–99,000 USD	12.0b (4.7)	2.1 (4.7)	2.4 (4.8)	3.9 (4.4)
100,000 USD or more	11.9c (4.5)	5.1 (4.6)	8.1 (4.6)	9.0b (4.0)
Location (omit: Metropolitan)
Micropolitan	−4.0 (3.7)	−0.0 (3.7)	7.5b (3.8)	−2.7 (3.5)
Small town/Rural	1.2 (5.6)	−10.4 (5.4)	8.5 (5.6)	−0.5 (5.8)
Unknown	−9.2 (6.9)	9.9 (6.5)	4.0 (5.7)	−1.3 (5.8)
Politics (omit: Democrat, reg. or lean)
Republican, reg. or lean	−1.8 (3.4)	−3.7 (3.4)	−0.3 (3.3)	3.5 (3.4)
Independent or unaligned	−0.8 (2.9)	−0.6 (3.0)	−3.3 (3.0)	0.5 (2.9)
Green party, Libertarian, or other	3.6 (5.0)	2.8 (5.0)	0.4 (4.6)	−5.0 (4.5)
Median or above internet skills	1.3 (2.9)	6.4b (2.9)	2.8 (2.9)	0.0 (2.9)
Median or above cognition score in:
Numeric module	10.6a (2.9)	−2.9 (3.0)	−3.1a (2.8)	2.9 (2.7)
Picture & Vocab module	9.0c (2.9)	−2.4 (3.0)	3.8 (3.0)	1.9 (2.9)
Verbal analogy module	4.9 (2.8)	6.3b (2.8)	−3.8 (2.8)	5.0 (2.7)
Median or above on personality trait:
Agreeableness	5.3b (2.6)	−1.5 (2.7)	−1.3 (2.6)	2.9 (2.7)
Conscientiousness	2.7 (2.7)	−8.3c (2.7)	−6.6b (2.6)	−4.3 (2.6)
Extroversion	−0.5 (2.5)	3.0 (2.6)	6.0b (2.5)	3.0 (2.4)
Neuroticism	3.7 (2.6)	−1.2 (2.8)	−2.5 (2.7)	−1.0 (2.7)
Openness	0.6 (2.6)	6.2b (2.7)	4.5 (2.5)	4.2 (2.5)
Occ. exposure to AI (Omit: medium exposure)
Top 5	−0.5 (3.1)	−6.8b (3.1)	4.8 (3.0)	−2.5 (3.0)
Bottom 5	1.8 (6.3)	−2.5 (6.6)	−13.2 (6.9)	−3.9 (6.8)
Never worked	−5.7 (6.9)	14.0 (7.3)	−3.9 (7.9)	−8.6 (6.6)
Occ not on list	8.8b (3.8)	5.6 (4.2)	−6.1 (3.8)	−6.3 (3.9)
Occ Missing	3.5 (3.8)	−3.4 (4.0)	−5.9 (4.0)	−2.4 (3.8)
N	2,614	2,614	2,614	2,614

Standard errors in parentheses, clustered at the respondent level. ${ }^{\mathrm{a}}P<0.001$, ${ }^{\mathrm{b}}P<0.05$, ${ }^{\mathrm{c}}P<0.01$.

Using LLMs for curiosity increases with household income but has no significant relationship with education. It is also more likely among individuals with higher scores in cognitive tests, and among those with greater agreeableness. Usage for entertainment is less common among individuals aged 50 and above, but more common among those with higher Internet skills, greater openness to experience, and lower conscientiousness.

LLM use for work tasks is more likely among younger, more educated individuals, as well as among those with higher scores on numeric cognitive tests, lower conscientiousness, and higher extroversion. Noticeably, it shows no association with the degree of AI exposure within the individual’s occupation.

Finally, LLM use for improving communication is more frequent among women, racial minorities (Black, Asian, Hispanic), and those with household incomes above $100,000.

Table S3 in the Appendix focuses on the predictors of the next four common reasons for using LLMs, including learning about the world, getting help with personal tasks, for creating writing, and seeking health information. Change in reasons for using LLMs section in the Appendix provides an analysis of the longitudinal changes in the number of reasons for using LLMs.

4b: perceived usefulness and harmfulness of LLMs

Overall, users report positive experiences with LLMs. In both waves, 43–45% rate their experience as very or extremely useful, while an additional 44% find it somewhat useful (see Appendix Fig. S4).

Perceptions of usefulness change little between wave 1 and wave 2. Column 1 of Appendix Table S4 examines which individual characteristics are associated with the likelihood of rating LLMs as “very” or “extremely” useful. Notably, all non-White respondents are significantly more likely than Whites to report LLMs as very or extremely useful. Most other demographic and personal traits show no significant association with the likelihood of rating LLMs as highly useful, except for respondents aged 60 or older (who see LLMs as less useful, −18.3 pp, $P<0.001$), and those with Internet skills at or above the sample median (who see LLMs as more useful, 7.2 pp, $P<0.05$).

Less than 10% of users in our sample view LLMs as “very” or “extremely” harmful. For completeness, column 2 of Appendix Table S4 reports the analysis of individual characteristics associated with this perception, though no statistically significant associations are found.

General discussion

This article presents findings from a nationally representative, longitudinal survey of more than 12,000 Americans, offering the most comprehensive assessment to date of how awareness, usage, and perceptions of LLMs are evolving in the US population. As LLMs hold the potential to enhance productivity and transform labor markets but also entail important risks, it is essential to understand who is currently using these tools, who is not, and who is likely to adopt them in the future. Our findings reveal persistent, and in some cases widening, gaps in LLM usage, while also shedding light on key factors that contribute to these disparities.

Overall, awareness of LLMs continues to grow, reaching 72% in spring 2024 (up 3.8 pp since fall 2023). Usage also increased by 5.8 pp during the same period, but remains far from ubiquitous, with 23.6% of respondents reporting having used LLMs in spring 2024. Consistent with some prior analyses (e.g. (14, 16, 19, 20)), but extending them to examine a broader set of features, we find marked sociodemographic disparities in LLM usage. Men, younger adults, individuals with higher levels of education and household income, and those employed in more analytical occupations are significantly more likely to report using LLMs. We also observe differences in usage by race, whereby Asian respondents are significantly more likely to use LLMs than Whites, and Black respondents significantly less likely to do so. There was no gap in usage between Hispanics and Whites. These findings align with evidence from the digital divide literature, which documents persistent socioeconomic and demographic disparities among early adopters of digital technologies such as the Internet and smartphones. (32–36)

Our analysis also reveals disparities across additional user characteristics beyond standard sociodemographic controls. Specifically, LLM use is more common among more politically liberal respondents and those with above-median cognitive ability, Internet skills, and openness to experience. These results are consistent with prior evidence that technology adoption in general, and digital technologies in particular, tend to favor novelty-seeking, technologically sophisticated, and higher-cognitive-ability individuals (24, 25, 37, 38).

To better understand the inequalities in LLM usage, we examine how our observable measures help to explain the usage gaps by gender and race. Differences in observable characteristics account for about half of the racial gaps, with the Black-White gap becoming statistically nonsignificant in multivariate analyses. In contrast, only one-third of the gender gap is explained by observables, suggesting that other factors, such as differing levels of trust in generative AI or varying perceived utility, may contribute to the persistent differences in usage between men and women.

Although some of the gaps in LLM use appear to be stable over time (e.g. the gender gap), our analyses indicate that others are increasing. Age, education, income, political preferences, and Internet skills significantly predict which individuals among those who were nonusers in wave 1 went on to adopt LLMs by wave 2, contributing to the widening of gaps in usage across these dimensions. The growing gaps may be concerning because they suggest that already advantaged group members, such as those from higher socioeconomic classes or with more education, are quickly positioning themselves to benefit most from generative AI tools in the future. In this possible modern-day illustration of the Matthew effect (39), those with existing advantages and opportunities might be able to leverage AI to gain even greater advantage.

The reasons for using LLMs are also changing over time. We observed a decline in use motivated by curiosity and entertainment and a rise in use for more productive activities, particularly work-related tasks. These increases in productive use were concentrated among high-frequency users. In contrast, low-frequency users continued to cite curiosity and entertainment as their main motives. This emerging divide between high- and low-frequency users mirrors earlier findings in the literature on digital inequality (including access to computers and the Internet), which emphasized that disparities in technology usage are compounded by differences in the benefits derived from that usage (32).

An important contribution of our study is showing that core predictors of technology uptake emphasized in the Technology Adoption Model (TAM) are also relevant for LLM adoption in the general population. These include factors associated with perceived ease of use (such as cognitive ability and technological savvy) and perceived usefulness (such as occupational exposure)—the central elements of the original TAM (40, 41)—as well as others considered in more recent extensions, such as personality traits (42). At the same time, our results point to additional influences, such as political orientation, not traditionally encompassed by TAM. While our data allow us to assess several determinants of technology adoption emphasized in TAM as well as some novel ones, they do not permit a direct assessment of others, such as social influence or technology-related anxiety or enjoyment (43, 44). Understanding their relevance for LLM adoption remains an open question for future research.

There are several limitations to our study. First, our data are US-specific and may not generalize to other contexts. Second, self-reported measures of awareness and usage could be imprecise. Third, although including two survey waves captures initial usage trends, we may still miss longer-term shifts or new use cases. Future longitudinal research could investigate the sustained impacts of LLM adoption on productivity, career trajectories, and societal trust in these technologies. Finally, while we analyzed individual and demographic predictors, future work should delve deeper into contextual factors, such as organizational policies, peer influences, and systemic barriers, to better understand the mechanisms that drive or hinder adoption.

Despite these caveats, our findings offer a robust, data-driven foundation to inform and motivate further research on the societal impacts of LLMs. Generative AI technologies, and LLMs in particular, hold the potential to improve individual outcomes in multiple domains, for instance by enhancing worker productivity, streamlining research processes, and supporting learning (4, 45–48). Yet, alongside these promises, there are risks stemming from privacy concerns, hallucinations, information verification challenges, and the possibility that LLM users acquire lower-quality knowledge due to excessive “cognitive offloading.” (5–7, 47, 49) By identifying the groups most likely to adopt and use LLMs, this article provides an essential guide for future research aiming to understand who stands to benefit or risks harm from this technology, and can help guide strategies intended to mitigate biases in the impacts of LLM technologies.

Supplementary Material

Supplementary material is available at PNAS Nexus online.

Funding

This work is supported in part by funds from the John S. and James L. Knight Foundation (GR-2022-65028) and the National Institute on Aging (U01AG077280). The project described in this article relies on data from survey(s) administered by the Understanding America Study, which is maintained by the Center for Economic and Social Research (CESR) at the University of Southern California. The content of this article is solely the responsibility of the authors and does not necessarily represent the official views of USC or UAS.

Author Contributions

M.A.: conceptualization; formal analysis; funding acquisition; writing-original draft; writing-review and editing. M.C.: conceptualization; formal analysis; writing-original draft; writing-review and editing. N.F.: conceptualization; funding acquisition; project administration; writing-review and editing. J.N.: conceptualization; software; formal analysis; investigation; visualization; writing-review and editing. J.S.: conceptualization; project administration; writing-review and editing

Data Availability

The data underlying this article are available in Open Science Framework at https://doi.org/10.17605/OSF.IO/HAZDR. The data are publicly available on the Understanding America Survey website: https://uasdata.usc.edu/index.php (datasets UAS 574 and UAS 607) and can be downloaded and used by any registered user. Data collection and dissemination of UAS surveys is overseen by the Biomedical Research Alliance of New York (BRANY), IRB 22-030-1044, which approved this study. Participants give their informed consent to take all UAS surveys.