Rein In the Titans, Cushion the Workers: Where Americans Stand on AI Policy

Cover Memo

This report draws on the AI@Work survey (CHIP50 Wave 38.6), a survey of 23,462 U.S. adults (weighted n = 23,485) fielded June 21–July 13, 2026. All percentages are weighted to U.S. Census population targets and are population-representative; unweighted respondent counts are reported as a reliability guide only.

These are interim estimates. Wave 38.6 launched June 21, 2026 and data collection is ongoing; figures reflect responses received through July 13, 2026 and will be revised in a later release.

This is a basic (full-demographic) report. It presents five AI-policy attitudes and their relationship with the standard demographic set — party, education, gender, income, community type, and age. (race_cat_5 is not available as a grouping variable in the Wave 38.6 extract — only race indicator flags are — so a race breakdown is not included.) The five attitudes and the summary measure reported for each are:

Coding and direction checks. pol_trust_ai (1 = not at all … 4 = a lot) confirmed via the AI-use gradient (mean 1.89 → 3.00). policy_export (1 = strongly promote … 5 = strongly oppose) and policy_influence (5 = strongly agree … 1 = strongly disagree) were direction-checked against AI trust: export-opposition falls and elite-concern rises as trust drops. policy_retrain (5 = support) and policy_lilas (1 = support) run in opposite numeric directions — verified previously — so each "% support" is computed in its own direction. Income (income_cat_10) is collapsed into lower (brackets 1–3), middle (4–7), and upper (8–10) thirds for readability; full-precision tables are in Appendix B.

Key Takeaways

Introduction

As AI moves from novelty to infrastructure, the politics of governing it are still forming — and public opinion will help decide them. The national picture is one of wariness. A Pew Research Center study fielded in 2024 found 51% of U.S. adults more concerned than excited about AI (just 11% more excited), 59% with little or no confidence in U.S. companies to develop and use AI responsibly, and bipartisan majorities worried government will not go far enough in regulating it¹.

Wave 38.6, fielded in mid-2026, extends that picture into specific governance questions and — the focus of this report — into how those attitudes vary across the American public. The headline posture is populist-protective: skepticism of AI and its corporate elite, reluctance to export it, and a clear, bipartisan appetite for cushioning the workers it disrupts. But the demographic detail tells a subtler story about who holds these views, and it complicates the simple "fearful public" reading: the Americans most at ease with AI are frequently the same ones most insistent that it be governed.

The five attitudes at a glance

Across all adults, the five attitudes stand as follows (weighted estimates):

• Trust AI: 41% trust AI somewhat or a lot; 59% little or not at all (only 10% "a lot").
• Elites too powerful: 55% agree tech elites have too much influence over AI policy; 9% disagree; 37% neutral.
• Oppose AI exports: 40% oppose promoting the export of advanced AI; 21% favor it; 40% neutral.
• Gov retraining: 65% of adults with an opinion support government AI-retraining programs (≈60% of all adults — the analysis tool drops this item's "not sure," unlike the Lifelong Learning Accounts item); 13% oppose.
• Lifelong accounts: 53% support portable Lifelong Learning Accounts; 10% oppose.

Comparative context. External polling corroborates the skeptical core: Pew's finding that 59% lack confidence in AI companies maps closely onto Wave 38.6's 55% who say tech elites have too much influence, and the bipartisan appetite for regulation matches the cross-party support for a worker safety net documented below. The rest of this report examines how each attitude varies across the population.

Findings by demographic

Each subsection reports the five summary measures (defined in the Cover Memo) across one demographic. All figures are weighted percentages; all items were asked of the full adult sample. "Oppose AI exports" is the share opposed; the other four are the share expressing the labeled view.

Party

The 7-point scale rewrites the party story the 3-point measure tells, in two ways the coarser split flattens.

First, pure independents are the trough on every single item — least trusting of AI (32%), least likely to say elites are too powerful (42%), least opposed to exports (34%), and least supportive of both retraining (49%) and Lifelong Learning Accounts (37%). This is not opposition but disengagement: on every item pure independents are the group most likely to sit at "neutral" (50% are neutral on the elite-influence question, 54% on exports). The 3-point measure understates this because its "Independent/Other" bucket blends these disengaged pure independents with party-leaning independents, who think and answer much like partisans — so the coarse "Independent 34%/50%/59%…" line is really an average of a disengaged middle and two partisan-like wings.

Second, Republican trust in AI is concentrated among the strongest Republicans. The much-noted fact that Republicans are the most trusting of AI is really a base phenomenon: trust runs from 54% among Strong Republicans down to 47% of Republicans and 44% of Lean Republicans, and Strong Republicans are also the group most likely to trust AI "a lot" (18%). It is the committed Republican base, not the party's establishment-leaning wing, that drives the counterintuitive topline. This same partisan trust gradient echoes in behavior: in the lead report, controlling for trust in AI cuts the partisan gaps in AI use roughly in half (adoption report, Table 10), so the disengaged-middle trough in use is, to a substantial degree, this trust trough carried into behavior.

Set the disengaged middle aside, and the striking thing about the rest of the spectrum is how little the protective attitudes divide it. Concern that tech elites have too much influence runs 53–63% across every partisan group — if anything slightly higher at the engaged poles (Strong Republicans 58%, Lean and Strong Democrats 63% and 61%) — so the "bipartisan distrust of the industry" holds, but as a partisan consensus that the unaffiliated middle abstains from rather than joins. The worker safety net is the same shape: retraining draws 64–70% from every partisan group and Lifelong Learning Accounts 53–60%, with pure independents alone well below (49% and 37%).

The 3-point breakdown (Democrat 40/58/39/67/55, Independent/Other 34/50/39/59/46, Republican 51/56/41/68/58 on the five measures) is retained in Appendix B for continuity; it collapses both the Strong-Republican trust gradient and the pure-independent trough shown above.

Education

Education produces the steepest and most consistent gradients in the report — and they point in a telling direction. The more educated a person is, the more they trust AI (56% of graduate-degree holders vs. 37% of the least-educated), the more they think tech elites have too much power (67% vs. 40%), the less they oppose exporting AI (30% vs. 38%), and the far more they support government retraining (82% vs. 45%) and Lifelong Learning Accounts (72% vs. 33%). In other words, the most educated are simultaneously the most comfortable with AI and the most insistent on governing it and protecting workers from it — an engaged, not a fearful, profile. The less-educated are not hostile so much as unformed: markedly more likely to sit neutral or unsure on every item.

Gender

Men give more directional answers than women on almost every item: they are more trusting of AI (47% vs. 35%), more likely to say tech elites are too powerful (59% vs. 50%), more supportive of both retraining (68% vs. 61%) and Lifelong Learning Accounts (57% vs. 48%), and slightly more open to exporting AI. Women are correspondingly more likely to be neutral or "not sure" across the board (for example, 13% of women are unsure about Lifelong Learning Accounts versus 7% of men). The gender gap here is less about opposing views than about men holding firmer opinions in every direction.

Income

Income tracks education almost exactly, reinforcing the same "engaged" pattern. Higher-income Americans trust AI more (55% in the upper third vs. 36% in the lower), are more convinced tech elites hold too much power (65% vs. 46%), are less opposed to exporting AI (33% vs. 40%), and are substantially more supportive of both worker-policy measures (retraining 76% vs. 55%; Lifelong Learning Accounts 67% vs. 42%). That upper-income Americans are the most supportive of government retraining programs is notable — it runs against the usual assumption that higher earners resist government spending, and fits the broader picture in which economic and educational advantage predicts engagement with AI policy in every direction at once.

Community type

Community type produces gentler gradients. Urban residents are somewhat more trusting of AI (45% vs. 35% rural) and more open to exporting it, while rural residents are the most export-skeptical (43% opposed). Concern about tech elites is broadly flat (52–56%), and support for the worker safety net barely moves across the urban–rural divide (retraining 62–65%) — one of the few places where geography is nearly irrelevant, underscoring how uniformly the protective agenda is held.

Age

Age reshapes two of the five attitudes sharply while leaving the others flat. Support for government retraining rises steeply and monotonically with age — from 50% among adults under 21 to 78% in their seventies and 86% past 80 — the clearest age gradient in the report. Trust in AI is lowest among the very youngest adults (just 24% of 18–20-year-olds, a heavily student group, trust it somewhat or a lot), then peaks in the thirties and forties before easing. Export skepticism is highest at the age extremes — the youngest (54% opposed) and the oldest — and lowest among prime-working-age adults. Concern that tech elites have too much power, by contrast, is remarkably constant across every age band (52–59%): whatever their generation, a majority of Americans distrust the industry's grip on AI's rules.

What the demographics add up to

Three patterns organize the demographic detail. First, a single "engagement" axis — running through education, income, gender, and to a lesser degree urbanicity — predicts nearly every attitude at once. The educated, higher-income, male, and urban are more trusting of AI and more willing to export it, and more convinced tech elites are too powerful, and far more supportive of retraining and Lifelong Learning Accounts. Comfort and control move together; the people most at ease with AI are the same ones most determined to govern it and to protect workers from it. This is the opposite of a fear-driven politics, in which the most anxious would demand the most control.

Second, the worker safety net is where age and class matter most. Support for government retraining climbs steeply with age, education, and income, even as it stays essentially flat across party and geography. The protective agenda's strongest constituency is older, better-educated, and higher-income Americans — not the workers most directly exposed to displacement, a gap worth noting for anyone reading these numbers as a demand from the vulnerable.

Third, distrust of the industry is nearly universal — with one telling exception. Concern that tech elites have too much influence over AI policy varies little by age or region, and almost no one actively disagrees (just 9% overall). But the 7-point party scale shows it is better described as a partisan near-consensus than a truly universal one: 53–63% of every party group agree, while pure independents are markedly lower (42%) — not because they disagree but because half of them sit neutral. The engaged, of every party, distrust the industry's grip; the disengaged middle mostly withholds judgment. If there is a mandate in these data, it is anchored there: across the partisan spectrum, a majority do not want the rules written by the industry alone.

Conclusion

American opinion on AI governance rests on two reinforcing feelings — distrust of AI as a technology and sharper distrust of the corporate elite steering it — from which follows an appetite for external checks and, decisively, for a government-backed safety net for displaced workers. The demographic detail deepens rather than complicates that story. The appetite for governing AI is strongest not among the fearful and marginal but among the engaged and advantaged: the educated, higher-income Americans who are simultaneously the most comfortable with the technology and the most insistent that it be regulated and its costs cushioned. The worker safety net commands majorities everywhere and rises with age and class; the conviction that tech elites are too powerful spans every divide.

If there is a coalition to be built for AI policy, these numbers suggest its shape: broad and bipartisan on protecting workers, near-universal on curbing the industry's influence, and — importantly — led as much by the confident as by the anxious. Rein in the titans, cushion the workers: across the American public, in nearly every demographic, that is where opinion currently stands.

Appendix A — Methods

Data source. AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project), the AI/employment supplement. Total unweighted n = 23,462; weighted n = 23,485. Fielded June 21–July 13, 2026, with an oversample of students and adults 18–22; it is nonetheless a broad adult sample. Estimates are weighted to U.S. Census population targets. Data collection was ongoing at the time of this extract — figures reflect responses through July 13, 2026 and are preliminary.

Measures and bases. All five items were asked of all adults: pol_trust_ai (n ≈ 23,301), policy_influence (n ≈ 23,443), policy_export (n ≈ 23,436), policy_retrain (n ≈ 21,565 excluding "not sure"), policy_lilas (n ≈ 23,450). Breakdowns use the 7-point party scale (party7), education_cat, gender, income_cat_10, urban_type, and age_cat_8. Party is shown on the 7-point scale because the 3-point measure hides the report's central partisan findings — the pure-independent disengagement trough on every item, and the within-Republican gradient in AI trust (Strong Republicans most trusting). The 3-point breakdown is retained in Appendix B for continuity. race_cat_5 is not available as a grouping variable in this extract (only race indicator flags), so no race breakdown is reported. Data-integrity note: the analysis tool drops policy_retrain's "not sure" responses (base 21,565) while retaining policy_lilas's (base 23,450), so retraining-support shares are computed among opinion-holders and run ~5 points above the all-adult figure — retaining "not sure," support is ≈60%, the figure comparable to the 53% for Lifelong Learning Accounts. The demographic breakdowns share the opinion-holder base, so contrasts within the item (party, age, education) are unaffected in shape.

Summary measures. For readability the Findings tables report a single share per cell (defined in the Cover Memo): trust "somewhat/a lot"; agree elites too powerful; oppose exports; support retraining; support Lifelong Learning Accounts. Full response distributions underlie each and are available in the topline tables (Appendix B) and the source distributions.

Coding and direction checks. pol_trust_ai (1 = not at all … 4 = a lot) — direction confirmed via the AI-use gradient (weighted mean 1.89 among non-users → 3.00 among daily users). policy_export (1 = strongly promote … 5 = strongly oppose) — confirmed: its mean falls as AI trust rises (3.88 among those who trust AI "not at all" → 2.49 among those who trust it "a lot"). policy_influence (5 = strongly agree … 1 = strongly disagree that tech elites are too powerful) — confirmed: agreement is highest among those who trust AI least. policy_retrain (5 = support) and policy_lilas (1 = support) run in opposite numeric directions (verified previously); each "% support" is computed in its own direction.

Income collapse. income_cat_10 is grouped into lower (brackets 1–3), middle (4–7), and upper (8–10) thirds; band percentages are weighted-n–pooled across brackets (verified programmatically).

Estimation. Cross-tabulations are weighted shares. All demographic patterns are descriptive weighted breakdowns, not modeled effects, and are described as such; no causal or "independent effect" claims are made. Cells with unweighted n < 10 are suppressed.

Not available. policy_energy and policy_titans (rest of the AI-policy battery) and race_cat_5 are not available in the released data and are omitted.

Reproducibility. Figures map to Appendix B; values copied verbatim from the underlying tabulations. External benchmark fetched and verified from Pew Research Center (full citation in the footnote at the end of Appendix B).

Appendix B — Data tables

Toplines (all adults, weighted %)

Table 1. Trust in AI (pol_trust_ai; n ≈ 23,301): Not at all 31.0 · Not much 27.8 · Somewhat 31.2 · A lot 10.0. (Somewhat/a lot = 41.2%.)

Table 2. Tech elites have too much influence (policy_influence; n ≈ 23,443): Strongly agree 23.6 · Agree 31.1 · Neutral 36.5 · Disagree 5.5 · Strongly disagree 3.3. (Agree = 54.7%.)

Table 3. Promote vs. oppose exporting advanced AI (policy_export; n ≈ 23,436): Strongly promote 6.5 · Promote 14.5 · Neutral 39.5 · Oppose 19.3 · Strongly oppose 20.3. (Oppose = 39.6%; promote = 21.0%.)

Table 4. Support for government AI-retraining programs (policy_retrain; n ≈ 21,565): Strongly support 35.2 · Support 29.3 · Neutral 22.3 · Oppose 6.3 · Strongly oppose 6.9. (Support = 64.5% among opinion-holders; the analysis tool drops this item's "not sure," so all-adult support is ≈60% — the comparable figure to the "not sure"-inclusive 53% for Lifelong Learning Accounts.)

Table 5. Support for portable Lifelong Learning Accounts (policy_lilas; n ≈ 23,450): Strongly support 21.7 · Support 31.0 · Neutral 27.3 · Oppose 5.6 · Strongly oppose 4.6 · Not sure 9.9. (Support = 52.6%.)

Summary-measure matrix by demographic (weighted %)

The tables in the Findings section reproduce, verbatim, the five summary measures for every demographic group (Party [7-point], Education, Gender, Income [lower/middle/upper thirds], Community type, Age). Each cell is the weighted share expressing the labeled view; unweighted group ns are large (thousands per group for the coarse demographics; the smallest age cell, 80+, has unweighted n ≈ 200+ per item, except suppressed sub-cells noted in source).

Direction-check gradients (weighted means, confirming coding): pol_trust_ai by AI-use 1.89→3.00; policy_export by AI-trust 3.88 (no trust) → 2.49 (high trust); policy_influence highest among AI-distrusters. Income-band pooling verified: e.g., trust "somewhat/a lot" 35.7 (lower) / 40.2 (middle) / 54.9 (upper).

Appendix C — Question wording (verbatim / mapped from codebook)

• pol_trust_ai — trust in AI, four-point scale (Not at all; Not much; Somewhat; A lot).
• policy_influence — "How much do you agree or disagree: tech elites have too much influence over AI policy." (Strongly agree … Strongly disagree)
• policy_export — "Should the U.S. promote or oppose the export of advanced AI?" (Strongly promote … Strongly oppose)
• policy_retrain — "Do you support or oppose government job-retraining programs for workers affected by AI or automation?" (Strongly support … Strongly oppose; Not sure)
• policy_lilas — support for portable "Lifelong Learning Accounts" workers could use for training throughout their careers (Strongly support … Strongly oppose; Not sure)
• Not available: policy_energy, policy_titans, race_cat_5.