What Californians think about AI

Executive Summary

On a warm Tuesday evening in Fresno, a retired union organizer and a twenty-something software engineer found themselves in rare agreement: AI was going to change everything, and not for the better. Their politics could not be more different, one a lifelong Democrat, the other a libertarian who has never voted blue, but both distrusted the tech companies building the tools.

Their unlikely consensus reflects a deeper pattern. This report presents findings from a major statewide research project, combining a survey of 1,400 Californians, focus groups, and in-depth interviews, revealing how the AI debate has broken out from tech conferences into dinner tables, workplace break rooms, and political town halls. Three-quarters of Californians have heard about AI in the past month; two-thirds have tried it themselves. And behind that engagement lies a cautious optimism, a hope that innovation might lower grocery or energy bills, improve access to healthcare, or ease the morning commute, but it is a hope shadowed by a deeper unease: whether the hands shaping these tools are building a society for all, or entrenching one built by and for the few.

A bipartisan belief that the benefits will be captured by the few. Fifty-nine percent of Californians believe the real winners of any AI revolution will be corporations and ultra-elites, specifically, a small number of individuals we can name, and not ordinary working and middle-class people. This is not a partisan finding: 55% of Republicans share this concern. Barely 20% believe working and middle-class families will benefit at all. And this is more than an economic equity concern. Persistently, people told us it was something bigger, a constitutional-style concern about the stability of society and democracy itself. As one participant put it: "This is going to negatively impact the traditional upper classes, too. If you don't own the corporation, you're screwed."

A clear mandate for state-led regulation. An overwhelming 70% of Californians want strong, enforceable laws to govern AI. They reject voluntary industry standards and self-policing outright, seeing them as inherently and fatally compromised by profit motives. The California state government is significantly more trusted than the federal government to implement effective measures, consistent with a state that has historically set the pace on consumer protection, environmental standards, and data privacy.

But trust comes with conditions. Many are deeply skeptical of legislators' independence, fearing they are already compromised or captured by the very industry they are tasked with regulating. The public's message is blunt: "Speak with experts, not lobbyists." "Are their campaigns funded by these companies? You can't serve two masters." Voters' fears crystallize into a concrete agenda: robust privacy protections; anti-bias rules with real enforcement in housing, hiring, and finance; worker protections that require automating companies to fund retraining; and a ban on lawmakers holding stock in AI-affected companies.

Five distinct mindsets shape the politics. Californians' attitudes to AI don't divide neatly along demographic lines. Instead, this research identifies five mindset segments, Market Optimists (14%), Hopeful Regulators (23%), Pragmatic Skeptics (14%), Alarmed Populists (19%), and the Cautiously Disengaged (30%), each defined by their values and emotional response to AI rather than their age or zip code. Understanding these tribes is strategically essential: they reveal which communities are already on board, which can be won over, and which will actively resist. The real strategic pivot lies with Pragmatic Skeptics and Alarmed Populists, they are concerned and persuadable, making them the best bet for building political momentum.

The stakes extend far beyond California's borders. Just as environmental standards, consumer protections, and data privacy laws born here have rewritten the rules elsewhere, so too could the state's decisions on AI. The companies building these systems are headquartered here. Legislation passed in Sacramento isn't just shaping the terms of access to a lucrative market, it's setting the rules at the very source of where the technology is developed. History shows how far that influence can reach: in 1966, California set vehicle emissions standards tougher than the federal government's, and within a decade automakers across the US and Asia had redesigned engines to comply. For AI, that influence is even more direct.

If we fix the fears, we free the gains. That trust is the key to converting caution into permission, giving banks, hospitals, schools, and government the license to harness AI and release the productivity dividend we've been promised. For lawmakers who say they want to capture AI's benefits, this research charts a clear path: demonstrate independence, legislate for fairness, transparency, and accountability, and the permission to innovate will follow. Without that independence, the mandate, and the benefits, collapse.

Ultimately, the question isn't whether AI will change California, but whether Californians will have any say in how.

Methodology

This report draws on a statewide survey of 1,400 Californians (1,000 core sample plus oversamples of Black, AAPI, and tech-savvy adults) conducted by Diffusion in partnership with TechEquity Collaborative, Lake Research Partners, and Voss Strategy, fielded 29 April–8 May 2025 (±3.1% margin of error). The study used multivariate techniques, including regression, factor analysis, and cluster analysis, to uncover the underlying belief structures shaping public views on AI, complemented by two online focus groups of 10 participants each. A detailed methodology is provided in Section 8.

This report is accompanied by two published opinion pieces: "Mapping California's AI Tribes" on TechPolicy Press, and an analysis of the research's implications for California's regulatory leadership.

1. Overall
Sentiment

1.1 Cautious optimism, deep unease

California is both the birthplace of these technologies and the place where public opinion about them is most intensely felt. Three-quarters of Californians have heard about AI in the past month; two-thirds have tried it themselves. The level of engagement is striking, this is not a distant policy debate for most people, but a lived reality they encounter through their phones, their workplaces, and their children's schools.

At first, there's an optimism about "innovation." People's stories often begin positively. AI tools are saving teachers hours on lesson plans and helping small business owners manage their books. When asked about AI's impact on broad categories like innovation (+51 net), healthcare (+26 net), or economic growth (+18 net), Californians lean positive.

But that optimism unravels quickly when you ask a simple question: "Who do you think will benefit from these technologies?" It's a question that interrupts the usual mental shortcuts, nudging people to pause, step back, reflect on their assumptions, and weigh the issue on their own terms. That moment of re-evaluation reveals a profound anxiety lying just beneath the surface. Ultimately, 55% of Californians say they are more concerned than excited about the future of AI, while just 33% are more excited.

The sense of pace adds urgency. Nearly half (48%) of Californians believe AI is developing far too quickly, driven by corporate interests rushing products to market without adequate oversight. Just 4% think it's moving too slowly. This sentiment resonates powerfully in community focus groups, where the anxiety is visceral: "It's already happening way too fast, and it's only going to get worse." "Companies are ramming things through too fast, and it feels sketchy. Honestly, it's freaking me out." "I just keep thinking… who is going to step in? Somebody has to do something before it goes too far."

1.2 The lynchpin question: who benefits?

The distributional question is the one that concentrates the mind. Fifty-nine percent of Californians believe the real winners of any AI revolution will be the wealthiest households and corporations, specifically, a small number of individuals we can name, and not ordinary working and middle-class people. The gap is stark: 59% say the wealthiest, just 20% say working people, and 20% are not sure. This is a bipartisan view: 55% of Republicans share this concern (compared to 64% of Democrats). Barely one in five believe working and middle-class families will benefit at all.

But this is more than just an economic equity concern. Persistently, people told us that this is bigger, it's a deeper, constitutional-style concern about the stability of society and democracy itself. Many believe that ultra-elite figures like Elon Musk, Mark Zuckerberg, and Sam Altman aren't merely profiting from the technology; they're skirting democratic processes and using it to quietly rewrite the rules of society in their own favor.

The language people use is striking for its intensity and specificity. "I'm more concerned that big corporations with all their money and all their lobbyists will control whatever commission we set up, and we'll lose control of it." "If the company makes more than what the fine is going to be, they're just going to keep breaking the law… they think they're above the law and nobody stops them." "They're only building this stuff to cut us out and keep more for themselves. Even if they mean well, their investors will force it."

Jobs and workers' rights top the list of immediate anxieties. When asked about AI's impact on specific domains affecting ordinary people, the picture flips sharply negative: jobs (-28 net), workers (-24 net), workers' rights (-23 net), and income inequality (-20 net). The concern is not merely theoretical. Californians perceive AI as actively reshaping the workplace, exacerbating income inequality, and eroding personal freedoms.

1.3 Risks and priorities

When asked to rank the specific risks they associate with AI, Californians place economic and privacy concerns at the top. The creation of fake photographs and videos, the reduction of personal privacy, and the spread of disinformation are seen as both extremely likely and extremely concerning. Job displacement, surveillance, and data misuse feel imminent and personal. Existential risks, autonomous weapons, nuclear control, generate concern but feel more distant.

A note on privacy: in focus groups, "privacy" functions as a shorthand for a much broader set of material harms. When Californians say they are worried about privacy, they are not primarily talking about data collection in the abstract, they are describing the fear that their personal information will be used against them: to discriminate in hiring, to inflate prices, or to deny services. Privacy, in this context, is the plain-spoken way of talking about algorithmic discrimination and data-driven exploitation.

These AI-specific fears align with Californians' broader issue priorities. When asked what issues they want their elected leaders to prioritize most, inflation and rising prices tops the list (31%), followed by jobs and the economy (18%), threats to democracy (13%), and housing affordability (9%). The connection between these everyday economic anxieties and AI's perceived threats is direct: people see AI as something that could make these existing problems worse, not better.

2. Policy &
Regulation

2.1 A mandate for state-led regulation

Despite deep cynicism, there remains a clear mandate for state-led regulation. An overwhelming 70% of Californians say we need strong laws to force companies to make AI safe and secure, that voluntary guidelines aren't enough. Just 18% believe we should trust companies to innovate responsibly without government interference.

The California state government, in voters' eyes, is uniquely positioned to lead this regulatory charge. Californians are 9 points more likely to trust the state government to control AI than the federal government, a significant gap that reflects both the state's regulatory tradition and, in the current moment, deep skepticism of federal leadership on technology issues.

2.2 Trust is fragile and conditional

But that trust comes with conditions. Many are deeply skeptical of legislators' independence, fearing they are already compromised or captured by the very industry they are tasked with regulating. The yearning for independence is a dominant theme in the qualitative research:

"Speak with experts, not lobbyists." "I want you all to ensure that AI companies cannot lobby politicians to make decisions that benefit them." "Are their campaigns funded by these companies? You can't serve two masters." "They're completely captured by industry, I wouldn't trust anything legislators say." "I worry they'll just get bought off." "If you ensure that no people are making these decisions that have financial ties or stocks… all of our politicians basically do insider trading."

The favorability data reinforces this picture. Labor unions are the most favorably viewed actors in the AI debate (+41 net), while Elon Musk (-21), Republicans in the state legislature (-22), and Silicon Valley executives (-10) are viewed unfavorably. Even tech industry executives as a category are essentially neutral (+2). Democrats in the state legislature hold a modest positive rating (+13), but voters' underlying message is clear: legitimacy in the AI debate comes from independence, not from party affiliation or industry credentials.

For lawmakers who say they want to capture AI's benefits, this research charts a clear path: demonstrate independence from industry, legislate for fairness, transparency, and accountability, and the permission to innovate will follow. That trust, in turn, is the key to converting caution into permission, giving banks, hospitals, schools, and government the license to harness AI and release the productivity dividend we've been promised. Without that independence, the mandate, and the benefits, collapse.

2.3 The voters' blueprint

These fears crystallize into a consistent, four-pillar policy agenda, a voters' blueprint for AI rules:

• Privacy first. Voters see privacy as the most urgent risk. Many said AI could "destroy their lives" if left unchecked. They backed the right to pre-emptively opt out of AI systems, and mandatory labeling of AI-generated content and tools. They want protection before harm occurs.
• Rules with teeth. Voters backed clear bans on discrimination in areas like housing, hiring, and finance, with enforcement mechanisms carrying meaningful penalties, not vague values statements. They want rules with teeth, not vague principles.
• Job protection. Voters said companies that automate should be required to create new jobs or pay into a fund for those displaced, and face financial penalties if they automate irresponsibly. They see fairness at work as a litmus test for AI policy.
• Earned trust. Voters don't believe California is acting independently of tech interests. They called for stronger oversight laws, a ban on policymakers holding AI-affected stocks, and mandatory tracking and reporting on job losses and other labor and environmental impacts. Support hinges on independence from tech.

Critically, when respondents are given additional information about each policy area, support either holds steady or increases across all segments. The voters' blueprint is not a product of ignorance, it is reinforced by knowledge. Privacy protection reaches 86% favor after messaging (up from 81%), civil rights protections reach 80% (up from 73%), and anti-discrimination rules reach 78% (up from 73%).

3. The AI
Compass

3.1 Segmentation methodology

When it comes to artificial intelligence, public opinion does not neatly divide along familiar lines of age, income, party, or geography. A young engineer in San Jose may see the world through the same lens as a retired venture capitalist in Santa Barbara, while a high school teacher in Fresno may share more with a factory worker in Fremont than with her tech-sector neighbors.

This is why this research takes a different approach, one that maps not who people are demographically, but how they think about AI, what they value, and what they fear. The method identifies and tracks distinct "mindset tribes" within the electorate, defined by their attitudes, values, behaviors, worldviews, and emotional responses rather than their age or zip code.

This isn't an academic exercise. Behavioural research has long shown that worldviews predict political behaviour better than demographics, especially for technologies as abstract, hyped, and fast-moving as AI. Inspired by the Yale Program on Climate Change Communication's "Six Americas" segmentation, which mapped distinct public mindsets on climate change and has since shaped communication strategies worldwide, this study applies the same attitudinal clustering approach to AI policy for the first time. Where Yale's work revealed that Americans' climate attitudes are driven by values and worldview rather than demographics, we find the same is true for AI.

3.2 Meet the Five Mindsets

Each of the five segments is a coherent cultural tribe, a community shaped by shared stories and sources of trust. At one pole, Market Optimists reflect the language and risk appetite of Silicon Valley venture circles. At the other, Alarmed Populists carry deep anxieties about lost jobs, remote decision-makers, and powerful machines they can neither see nor control. In between lie groups whose attitudes hinge on concrete issues, economic fairness, workplace surveillance, and whether policymakers can demonstrate real independence from industry.

Understanding these tribes is strategically essential. Policymakers cannot persuade, mobilize, or build a durable consensus without knowing which communities are already with them, which can be won over, and which will actively resist. This approach reveals where common ground can be built across seemingly opposed constituencies, and where the fault lines may be too deep to bridge.

3.3 Segment attitudes & cross-tabs

The five mindsets diverge sharply across key attitudinal questions. The charts below show how each segment responds to questions about favorability toward AI actors, the role of government, the pace of development, the case for regulation, trust in federal institutions, and policy preferences.

The policy heatmaps below show initial and final policy preferences broken down by segment, revealing which groups shift most when given additional information.

3.4 Geographic distribution

Los Angeles and the Bay Area are the center of gravity for all five segments, but their compositions differ in ways that matter for outreach and mobilization.

LA County is the single largest concentration point, accounting for 26% of the total sample. It concentrates Market Optimists (31%) and Hopeful Regulators (35%), making it the epicenter of both pro-innovation and pro-regulation energy. One in four Alarmed Populists (25%) and one in five Pragmatic Skeptics (20%) also reside here, meaning LA is a microcosm of the full attitudinal spectrum.

The Bay Area accounts for 20% of the total sample but punches above its weight in Pragmatic Skeptics (25%) and Hopeful Regulators (21%). It also holds 20% of the state's Alarmed Populists. This mix of tech-exposed skepticism and regulatory enthusiasm makes the Bay a natural proving ground for policy messages that balance innovation with accountability.

Orange County and San Diego lean pro-innovation. Orange County over-indexes on Market Optimists (13% of that segment from just 8% of the total population), while San Diego's profile more closely mirrors the state average. Both regions show lower concentrations of Alarmed Populists.

Inland Empire, Fresno, and Sacramento over-index on Alarmed Populists and the Cautiously Disengaged. Sacramento in particular concentrates Pragmatic Skeptics (16% of that segment from 12% of the total). That makes it a useful location for regulation-focused messaging. Fresno and the Inland Empire show elevated Alarmed Populist shares (10% and 13% respectively, against their total population shares of 7% and 11%), reflecting the anxiety-driven attitudes concentrated in communities further from the tech industry's center of gravity.

The Central Coast and North regions have smaller populations and roughly proportional segment shares, though the North shows a slight lean toward Hopeful Regulators relative to its size.

3.5 Strategic targeting: shiftable and sympathetic

When we consider who is most open to shifting, who is already paying attention, and who holds social and political sway, two groups stand out: Pragmatic Skeptics and Alarmed Populists.

They are concerned, engaged, and persuadable, our best bet for building momentum. The others are either locked in or largely disengaged.

Market Optimists are locked in: their pro-innovation, anti-regulation stance is ideologically anchored, and their low electoral participation limits their political sway. Hopeful Regulators are already allies, they need to be maintained, not converted. The Cautiously Disengaged, while large (30%), are unlikely to mobilize around AI unless personally impacted, and their shallow engagement makes them susceptible to counter-messaging.

That leaves the two shiftable segments. Pragmatic Skeptics (14%) are young, female-skewing, and strongly Democratic but under-registered. Their skepticism is grounded in practical concerns, jobs, surveillance, economic fairness, and they respond to concrete policy proposals rather than abstract arguments. Mobilizing them could expand the pro-regulation electorate. Alarmed Populists (19%) are older, highly engaged civically, and deeply anxious. They are already paying attention and already showing up, at rallies, at public meetings, at the ballot box. What they need isn't activation but direction: visible action that channels their anxiety into support for specific policy.

Together, the two segments are 33% of California's adult population. They share a common distrust of corporate self-regulation, a belief that government should act, and a responsiveness to messages grounded in tangible protections. The strategic opportunity is to unite them around a shared policy agenda, even though their emotional starting points (functional skepticism vs. existential alarm) and their narrative frames (Machine vs. Mythical/Science Fiction) differ sharply.

4. Voter
Demographics

The AI Compass segments in the previous section are the most powerful lens for understanding Californian opinion on AI. With a technology this abstract, fast-moving, and unevenly experienced, traditional demographic categories, gender, age, race, education, are weaker predictors of attitudes than the underlying worldviews, values, and emotional responses captured by attitudinal clustering. The pattern, worldview beating demographics, is well documented in research on emerging technologies.

Nonetheless, demographic breakdowns are the variables that voters, strategists, and legislators most commonly use to think about the dynamics of the electorate. This section identifies the most notable demographic patterns in the data. They should be read as electoral context for the deeper structural analysis, not as the primary organizing framework.

Note: Unless otherwise stated, the findings in this section are drawn from the 1,248 registered voters within our 1,400-person sample, reflecting the electoral lens of this analysis.

4.1 Gender and age

When reviewing the data for correlations that move together, the most significant demographic indicator is gender. Among registered voters under 50, men are net 44 points more positive than women about AI's future. Stratifying further by education and age widens the gap dramatically: college-educated men under 50 are +24 net excited about AI's future, while non-college-educated women aged 50 and over are −61 net concerned, an 85-point spread.

These two groups inhabit fundamentally different realities. Younger, college-educated men remain wary of elite influence but, on balance, see AI as an exciting career opportunity. For older, non-college-educated women, it reads as an uncontrolled threat to themselves, their families, and broader society. These groups also respond to different frames: younger adults engage with economic opportunity messaging, while older voters, especially women, respond more to reassurance and protection.

The electoral implication is significant. Older women are the most concerned cohort in our data, and that combination of high concern and high turnout makes them a powerful constituency shaping the political dynamics on this issue. Privacy concerns are particularly acute among women: among registered voters, 46% of women are extremely concerned about AI-enabled deepfakes, compared to 34% of men. Among women over 50, that figure rises to 49%. The result is two parallel conversations within the electorate, one about opportunity, one about threat, running simultaneously and often talking past each other.

4.2 Race and ethnicity

Although gender, age, and education are by far the most significant demographic factors, there are also distinctive racial patterns in AI attitudes. In simple terms, those who have experienced structural discrimination are more sensitive to how these tools get built and used.

Black voters are cautious, but more trusting of government than other groups. Among registered voters, 29% of Black Californians believe AI will increase discrimination (vs. 26% overall), and they show some of the strongest support for civil rights protections in AI: 82% favor new regulation (vs. 75% overall). Notably, they are also the most confident in government's ability to control AI: 42% trust the state government on AI, compared with 36% overall. For Black voters, AI regulation is a matter of principle and process rather than a personal crisis.

Latino voters express a greater sense of urgency and economic risk. They are the most concerned about AI's impact on jobs: 58% are concerned AI will replace lower-paying jobs (vs. 52% overall) and 59% are concerned it will reduce wages (vs. 55% overall). On workers' rights, 47% see AI's net impact as negative, vs. 43% overall. They are also the most convinced that AI will exacerbate economic inequality, with 63% believing that only elites will benefit (vs. 60% overall). This economic anxiety makes them strongly supportive of job protection policies. The research also reveals a messaging pattern specific to this group: promises to "retrain workers for the AI future" tend to backfire among Latino voters, as they hear them as confirmation that job losses are already on the way. Messages emphasizing corporate accountability,"companies should pay a price if they automate without contributing to the community", perform significantly better.

Asian American and Pacific Islander voters are the most AI-familiar demographic: 70% have used large language models, and 17% work in the tech sector (compared to a median of 7% for other non-white groups). They are favorable toward regulation,43% favor new regulations "a great deal," 75% favor new regulations overall, but not as acutely as other communities of color (for comparison, 49% of Latino voters favor new regulations "a great deal," while Black communities are 82% favorable for new regulations as a whole). Their proximity to the industry gives their support for regulation a distinctive, informed quality.

The variation here suggests that race is ultimately a weaker predictor of AI attitudes than economic position or industry proximity. The distinctive patterns within each group are better explained by their relationship to the technology and the labor market than by race itself.

4.3 Regional patterns

The geographic distribution of AI attitudes, introduced in Section 3.4, also reveals important regional dynamics when viewed through an electoral lens.

In the Bay Area, proximity to the industry breeds a particular kind of skepticism. AI literacy is highest here (77% familiar with AI, 67% have used large language models), but so is home-turf scrutiny: Silicon Valley executives draw a thin 41/35 favorability split among Bay Area registered voters, and 75% support strong laws. This is not abstract concern, it is lived experience and local grievance from people who have watched the tech industry's promises collide with its reality. Messaging that acknowledges the value of innovation while demanding accountability,"guardrails so innovation can thrive without the public paying the costs again", resonates here. Leaders who can credibly criticize tech excess while championing innovation may find an unusual coalition.

In Los Angeles, the Inland Empire, and the Central Valley, between 53% and 59% of voters are concerned AI will replace low-paying jobs, with the Inland Empire the most anxious. AI is seen less as an opportunity and more as a direct threat to livelihoods, from Hollywood and the creative industries through to service work and logistics. Protection framing lands: "clear rules, real penalties, worker-first safeguards." These audiences want tangible action and enforcement with teeth, principles without consequences are unlikely to persuade.

Sacramento is the outlier: at 63% support for strong regulation (vs. 72% or higher elsewhere), it shows notably lower enthusiasm for the legislative response. This may reflect a particular cynicism about legislative politics from people who live closest to the Capitol and harbor doubts that new laws change outcomes. Competence framing should land here: "clear standards, real oversight, and enforcement that actually sticks."

Across all regions, however, support for strong regulation remains a clear majority position, the variation is in intensity, not direction.

5. Narrative
Frames

5.1 The core frame

When asked to describe AI to a friend who doesn't know much about it, the most common narrative frame across Californians is Machine, AI as a "complex calculator," a "powerful tool… [that can] help solve problems." This framing positions AI as a designed system, which is politically useful: it can be changed, rebuilt, and regulated. It makes policy intervention conceptually viable.

But beneath that shared structure, each segment tells a subtly different story.

5.2 Frame dynamics by segment

Market Optimists favor the Machine frame (24.9%) but are uniquely drawn to Body/Embodiment framing (23.0%), describing AI as "an artificial person helping you work out your problems" or "a brain just like ours inside a computer." This group sees AI as something human-like, not a tool, designed to assist and extend our cognition. The Body metaphor introduces a sense of autonomy, something with its own agency or inner logic. A light-touch regulation approach is preferred; control feels unnatural to them. This risks downplaying systemic risks.

Hopeful Regulators lean most heavily on the Machine frame (43.0%), with notable engagement in Mythical/Science Fiction (21.4%) and Body/Embodiment framing (11.9%). Their language blends functionality with openness to systemic transformation, imagining what AI might become, not just what it is. They are captivated by potential, and their language orients toward what AI could become, not what it currently does. They are receptive to governance, especially when framed as visionary or future-proofing.

Pragmatic Skeptics are the most functionally minded: over half (51.8%) use the Machine frame. Their language is generally dry, technical, and stripped of emotion, AI as merely an appliance or infrastructure, "just a different computer." The low metaphor density reflects emotional distance and wariness of hype. They reject sentimental appeals. Credibility comes from restraint, clarity, and hard limits.

Alarmed Populists stand apart. They are the only group not to center the Machine frame. Instead, they rely heavily on Mythical/Science Fiction (30.1%) and Power/Hierarchy framing, seeing AI as a tool of elite control or a runaway dystopia. Their metaphors are more vivid, often hyperbolic, and driven by suspicion: "Knower of all," "AI is capable of ending our civilization as we know it," "intellectual robot of the future, like on space ships," and explicit analogies to Terminator, Star Wars, and other science-fiction futures. AI in their telling is distant, fantastical, and dangerous, not made by humans, but unleashed upon them. Trust must be earned through visible constraint of elite actors. Reassurance alone backfires.

The Cautiously Disengaged mostly use Machine (37.1%) and Mythical/Science Fiction (19.6%) language. This suggests cultural familiarity without deep conviction, a blend of absorbed tropes and vague caution. AI is vaguely familiar, faintly unsettling, but not urgent. Their echoes of both function and fiction, without commitment to either, suggest that effective messages will be grounded, calm, and concrete, signaling stewardship, not spectacle.

5.3 The China question

The data uncovers a vulnerability in the China competitiveness narrative, but not in the way you might expect. When respondents were asked how important it is for America to advance AI (without mentioning China), net importance was +20. When the question was reframed as advancing AI "before China does," net importance jumped to +35, a 15-point lift overall.

But the segment-level picture reveals sharp divergence that makes the China frame strategically dangerous.

Market Optimists, the most pro-innovation segment, actually recoil at the mention of China. Their net importance drops 33 points (from +77 to +44) when China is introduced. For a group motivated by personal gain and market logic, the geopolitical frame feels like government overreach, not innovation support.

Alarmed Populists surge in the opposite direction: their net importance jumps 41 points (from −39 to +2) when China is mentioned. They feel powerless and betrayed, unsure who to blame, and China offers a convenient villain. In focus groups, they said: "I think it's in another country… I think it's China or… They're way ahead of us" and "They don't need to have fly-over balloons any more, they already know everything about us over here." But what they really want is a champion who will defend their values and jobs.

Pragmatic Skeptics shift modestly (+8 points), but their framing is revealing. They feel exploited and dismissed, and place the blame on domestic elites and big tech. They see "China" talk as a distraction, and want proof that AI rules will put the public, not corporations, in control. As one participant said: "Does this competition just continue until AI replaces every single job, [making shareholders] as wealthy as they possibly can be?" Another asked: "Aren't the US corporations buddying up with the government?… they have contracts with the military, so it's all in tandem."

The implication is clear: opponents can exploit the China frame to undermine regulation by casting it as a competitive handicap. Policymakers should avoid simple competitive framing and anchor messages in relatable domestic impacts. The research's own messaging, tested through A/B experiments, shows that domestic framing consistently outperforms geopolitical framing across the segments that matter most.

6. Messaging
& Testing

6.1 A/B test results

The research tested a series of paired message frames to understand which approaches shift opinion most effectively. Each A/B test presents respondents with two alternative ways of framing the same policy issue, then measures which generates stronger support across the five mindset segments. Key tests included: "people" vs. "a person" having oversight over AI decisions, trust in AI-made decisions (initial vs. post-messaging), advancing AI for its own sake vs. to beat China, and real-world scenarios involving supermarket pricing and rent-setting by AI systems.

The rent-hike test is especially revealing. Respondents were told that after the Los Angeles fires, software systems detected a reduced supply of housing and automatically started to increase rents, and asked whether this was a good or bad thing. When the scenario was framed as an example of "artificial intelligence," 70% called it a bad thing (net −53). When it was framed as "unrestricted technology," that figure rose to 75% (net −61). The gap suggests that the abstract label "AI" may actually soften public reaction compared with language that foregrounds the absence of rules. For communicators, the implication is clear: naming the lack of constraint is more powerful than naming the technology itself.

6.2 Dial trace results

In addition to A/B testing, the research employed dial testing, tracking real-time audience reactions as policy messages are read aloud. Each line represents one of the five mindset segments, showing how their approval rises and falls at specific phrases and turning points within a message.

6.3 Strategic synthesis

The chart below summarizes the overall effectiveness of each message tested, showing the pre- and post-exposure shift in support across the full sample.

The messaging data reveals a consistent pattern. After exposure to the research's policy messages, support for privacy protections rose from 74 to 82 net favor (+8), upholding civil rights rose from 65 to 74 (+9), and anti-discrimination regulations rose from 64 to 68 (+4). These gains were broadly distributed across segments, but the largest shifts came from the groups that started furthest from the pro-regulation position: Market Optimists gained +8 on privacy and +9 on civil rights protections, while Alarmed Populists gained +14 on civil rights, the single largest segment-level shift on any policy question.

Across the five dial-tested messages, one finding stands out: the "Fair Costs, Shared Gains" message, which leads with rising prices, names corporations using AI and personal data to push costs higher, and closes with a call for fair pricing guaranteed by law, tops both instant and considered scores in every segment except Market Optimists, who prefer the industry's "Trust Innovators" pitch. The pattern holds for all three of the research's own messages: tangible, pocketbook-first framing consistently outperforms more abstract or existential appeals.

On human oversight, 70% of Californians rated it important after messaging (up from 66%), with the strongest gains among Market Optimists (+7) and Cautiously Disengaged (+9). On trust in AI decisions, distrust deepened slightly after messaging (from 49% to 52%), with the sharpest increases among Pragmatic Skeptics (+9), suggesting that more information makes this segment more wary of unaccountable AI, not less.

Language choices matter at a granular level. Split testing among registered voters (n=1,248) shows that "Silicon Valley executives" elicits more negative sentiment than "tech industry executives." Inside California, "Silicon Valley" reads as a specific power center with a locally charged reputation; outside the state, "tech industry" reads more neutrally as a sector. For California-specific communications, the more geographically grounded framing carries sharper rhetorical force.

There is also a messaging asymmetry around workforce transition. Promises to "retrain workers for the AI future" tend to backfire, voters hear them as confirmation that job losses are on the way. Messages that emphasize corporate accountability ("companies should pay a price if they automate without contributing to the community") perform significantly better. This pattern is especially pronounced among Latino voters, 58% of whom are concerned AI will replace low-paying jobs and 59% of whom believe it will reduce people's wages. The implication for policymakers is clear: lead with accountability and enforcement, not with retraining pledges that inadvertently validate the threat.

Taken together, the messaging data points to a six-part recipe for effective AI communication.

• Start with tangible impact. Show how AI touches everyday life so people feel it immediately.
• Emphasize shared values. Use widely held principles as motivating anchors.
• Define actors, actions, and consequences. Link corporate choices to real outcomes so responsibility is clear.
• Connect personal harms to systemic causes. Show that individual experiences are symptoms of systems that need structural fixes.
• Stick to a single narrative frame. Whether "building" or "machine," consistency strengthens coherence.
• Define concrete action. Offer visible policies rather than abstract principles.

For California legislators, these findings offer actionable guidance. Prioritize tangible, immediate issues, like job security, data privacy, and economic fairness, to mobilize Pragmatic Skeptics and Alarmed Populists effectively. Emphasize legislative independence from industry, clearly delineating between expert consultation and corporate lobbying. Craft messages that resonate emotionally and practically, turning abstract concerns into specific and measurable policies. The data shows that good messaging shifts attitudes and builds durable coalitions across segments that appear, at first glance, to have little in common.

7. Conclusion

7.1 If we fix the fears, we free the gains

The central insight of this research is not that Californians are afraid of AI. It is that their fear is conditional, and that resolving it unlocks something far more valuable than compliance. Clear rules convert caution into permission, giving banks, hospitals, schools, and government the license to harness AI and release the productivity dividend we've been promised.

The data bears this out. After exposure to concrete policy messaging, support for privacy protections rose to 86%, civil rights protections to 80%, and anti-discrimination regulations to 78%. Human oversight gained importance across every segment. Even Market Optimists, the most resistant to regulation, shifted meaningfully toward supporting enforceable rules. The permission to innovate doesn't come despite regulation; it comes because of it.

7.2 Key takeaways

Mandate. Voters want California to set enforceable AI rules now. Seventy percent favor strong laws over voluntary guidelines, and this is a bipartisan position. The mandate is not ambiguous.

Anxiety. They fear ultra-elites are rewriting the rules to entrench disadvantage. Fifty-nine percent believe only corporations and the wealthiest individuals will benefit. This isn't abstract anxiety. It's a specific concern about power and democratic control.

Conditional trust. They are wary of government motives, but want to have faith. Californians trust the state government nine points more than the federal government to regulate AI, but that trust is conditional on demonstrated independence from the industry being regulated. Voters are blunt: "Speak with experts, not lobbyists."

Rules you can see. Trust hinges on visible protection of fairness, rights, and social norms. The four-pillar voters' blueprint, privacy, rules with teeth, job protection, and earned trust, emerged consistently across surveys, focus groups, and message testing. Voters want rules they can point to, not principles they have to take on faith.

Industry skepticism. They do not trust industry to self-police, influence, or co-write the guardrails. The rejection of voluntary standards is near-universal outside Market Optimists. Even Hopeful Regulators, who are excited about AI's potential, insist on mandatory regulation with real enforcement.

Unlock trust. Enforceable rules convert anxiety into a platform for innovation and investment. If legislators can commit visibly to independence, legislate for fairness, transparency, and accountability, the permission to innovate will follow. Without that independence, the mandate, and the benefits, collapse.

7.3 California's regulatory moment

California isn't just the birthplace of these technologies, it's the world's regulatory testbed. From emissions standards to privacy protections, rules written in Sacramento have a habit of shaping markets from New York to New Delhi.

History shows how far that influence can reach. In 1966, California set vehicle emissions standards tougher than those of the federal government. With an economy that would rank fourth in the world if it were a country, and laws easy for other states to adopt, California's rules soon became the default. Within a decade, automakers across the US, and later in Asia, had redesigned engines to comply. The economic logic was simple: California's market power made compliance the price of entry.

For AI, that influence is even more direct. The companies building these systems are headquartered here. Legislation passed in Sacramento isn't just shaping the terms of access to a lucrative market, it's setting the rules at the very source of where the technology is developed. Like the automakers then, AI firms will adapt their products to meet the toughest standards in their home jurisdiction and their biggest markets.

If the state now sets clear, enforceable AI rules, it can do more than regulate; it can set the global benchmark for fairness and accountability. The question isn't whether AI will change California, but whether Californians will have any say in how. The mandate is clear and urgent: act decisively, in public, and in the public interest.

8. Methodology

This research was conducted by Diffusion in partnership with TechEquity Collaborative, with polling and analysis by Lake Research Partners and Voss Strategy.

Survey design

The quantitative survey was fielded online among 1,400 Californian adults from 29 April to 8 May 2025. The core sample of 1,000 respondents was designed to be representative of the state's adult population by age, gender, ethnicity, geography, and political affiliation, with a margin of error of ±3.1% (with larger margins for subgroups). The study also included targeted oversamples of Black Californians (n=100), Asian American and Pacific Islander Californians (n=100), and technology-savvy adults (n=200) to ensure robust sub-group analysis across communities that are disproportionately affected by AI deployment. Post-stratification weighting was applied by gender, age, race, region, and education to ensure representativeness of the adult California population.

Technology-savvy respondents were defined as registered voters who met all of the following criteria: they consumed at least two news sources multiple times per week; engaged in at least two political activation behaviors (e.g., signing petitions, attending meetings); used three or more online platforms regularly; and posted political or social opinions on social media at least a couple of times per week. Tech-savvy adults made up 12% of the final weighted sample.

Qualitative research

The study also included a series of focus groups and in-depth interviews across multiple California communities, providing the qualitative depth that grounds the quantitative findings in lived experience. Participants were recruited from diverse geographic, ethnic, and socioeconomic backgrounds to ensure the voices reflected in this report span the full range of Californian perspectives.

Segmentation

The AI Compass segmentation model was developed through a two-stage quantitative classification procedure. First, exploratory factor analysis was applied to a subset of attitudinal batteries covering values, worldviews, risk perceptions, and emotional responses to AI. This identified underlying latent dimensions within the data by examining which items loaded most strongly on each factor. Initial results suggested one dominant factor per battery, though substructure within one or more batteries was also possible.

K-means cluster analysis was then conducted using the top-loading items from each factor. Models with four, five, and six clusters were tested and evaluated based on the clarity of interpretive distinctions between segments, aiming to optimize both analytic usefulness and policy relevance. The process looked for consistent profiles, such as "base support," "strong opposition," and "mixed/ambivalent" segments, and further disaggregated large clusters (N≈600) in higher-cluster solutions to identify meaningful subtypes. The five-cluster solution was selected as the most robust and interpretively clear.

CHAID analysis

To identify demographic and attitudinal subgroups driving differences in AI sentiment, a CHAID (Chi-squared Automatic Interaction Detection) analysis was conducted using responses to attitudinal questions about AI's future direction. Responses were recoded to a five-point ordinal scale ranging from 1 ("much more concerned") to 5 ("much more excited"). CHAID then split the sample into mutually exclusive nodes based on statistically significant interactions with key demographic and behavioral predictors (e.g., age, gender, education, tech use). Each terminal node represents a distinct subgroup with internally similar and externally distinct attitudes. Mean scores were computed for each node to interpret group differences, with higher values indicating greater excitement about AI.

AI literacy index

Each respondent's AI Literacy Score was calculated as the average of their z-standardized responses to eight five-point Likert items measuring familiarity with AI's outcomes, risks, core concepts, and enabling technologies. These values were then linearly rescaled to a 0–100 scale, where higher scores reflect greater conceptual understanding. For interpretability, respondents were assigned to four ordered literacy segments using pre-defined cut-offs derived from the distribution observed in the Californian sample: the high AI literacy segment have scores between 75–100, the advanced segment is between 63–75, the moderate segment is 50–62, and the lowest segment have scores of less than 50.

This approach draws on decades of behavioral research showing that attitudes and values are more predictive of political behavior than demographic categories when it comes to emerging technologies. The resulting segments are defined by how people think about AI, not by who they are demographically, though each segment shows distinctive demographic and geographic patterns as a consequence of those shared worldviews.

Narrative frame analysis

To map the narrative logic embedded in public discourse, we applied the Narrative Frames typology, originally developed by the lead author at the University of Cambridge, to a corpus of approximately 1,300 open-text responses to the questions: "Imagine you were trying to explain artificial intelligence (AI) to a friend who doesn't really know much about it. In your own words, how would you describe AI, what it does, what it might mean for us all?" and a second open-text prompt later in the survey.

Each response was processed through a two-stage qualitative coding pipeline. First, metaphor identification: responses were screened using the Pragglejaz Group's Metaphor Identification Procedure (MIP) to detect metaphorical expressions, with a second pass applying Charteris-Black's Critical Metaphor Analysis (CMA) framework to identify figurative constructions contributing to narrative coherence and persuasive force. Second, frame assignment: identified metaphorical fragments were coded to one of 22 primary narrative frames (Machine, Body/Embodiment, Mythical/Science Fiction, Power/Hierarchy, Journey, and others) and, where applicable, one of 28 sub-frames.

Each respondent was thereby assigned a full "frame profile," indicating which narratives were invoked and which were absent. This allowed us to quantify frame prevalence, blending, and clustering across the sample, and to compare public responses to previously analyzed elite discourse, surfacing key divergences in how AI is conceptualized by policymakers versus the public. Frame frequencies were cross-tabulated by segment to reveal how each group's underlying worldview shapes the language they use to understand AI.

Message testing

The study employed two complementary testing methodologies. A/B split testing presented respondents with paired message frames on the same policy issue (e.g., "people" vs. "a person" having oversight; advancing AI for its own sake vs. to beat China) and measured which generates stronger support across the five segments.

Dial testing incorporated a real-time test in which every respondent rated each message on a 0–100 slider that opened at 50 and had to be moved at least once every two seconds, generating a continuous approval trace. Messages were shown in random order and weighted in line with the main survey. Traces were smoothed, and mean scores, as well as lifts by the five voter segments, were calculated with the same 95% confidence standards used throughout the study. This method was chosen over other message testing methodologies because the dial captures the immediate emotional pulse: where approval rises, stalls, or plunges, revealing which words persuade, which phrases trigger concern, and ultimately which messages are most likely to shift support for sensible AI safeguards with statistical confidence. Messages were tested across five thematic areas: fair costs, accountability, jobs, safety, and trust.