Engagement, Not Fear: Who Actually Drives Demand for an AI Safety Net

The question

Two of the reports in this series point in different directions about why Americans support a government response to AI at work. The AI-governance report finds support for retraining programs concentrated among the educated, higher-income, and AI-engaged — and argues the appetite is "engagement, not fear." The job-loss-anxiety report documents the other candidate mechanism: workers who fear AI will take their jobs. Neither report can settle it alone, because neither puts fear and engagement in the same model. This capstone does.

The test. Among workers, does fear (expecting AI to cost jobs) or engagement (using AI, being educated/higher-income) better predict support for a government AI-retraining program — controlling for each other and for party, age, and gender? The outcome is a binary "supports government AI-retraining programs" (policy_retrain = support or strongly support). Fear and engagement are entered together; whichever survives, wins.

The finding

Engagement drives support for a policy response. Fear does not — if anything, it works against it. Net of everything, the workers most threatened by AI are less likely to back a retraining program, while the engaged and advantaged are far more likely. This holds across two independent fear measures.

The raw numbers already hint at it: workers who expect AI to cost jobs in their field support retraining at 58%, below the 62% among workers who don't. Controlling for socioeconomics — which the fearful are lower on — widens the gap into a clean negative association.

Key model (weighted logistic; outcome = supports government retraining; worker base, n = 11,609; McFadden R² = 0.064):

Robustness. Swapping the fear measure for a different one — expecting AI to have an only-negative impact on one's own job (from the job-loss report) — reproduces the result and, if anything, strengthens it: fear OR 0.77 (p<.001), with engagement unchanged (uses AI 1.34; graduate degree 2.13). Two independently-worded threat measures, same answer: the threatened are less supportive, not more.

Comparative context. External polling finds broad aggregate support for AI-displacement retraining programs: Data for Progress — a progressive-leaning advocacy pollster surveying likely voters rather than the general public — reported 74% support for "creating a new jobs training program for workers displaced by AI," a figure useful only as an order-of-magnitude signal and not as confirmation of any AI@Work survey level. No external survey, in that house or in the project's preferred benchmark set, decomposes that support by engagement with AI versus exposure to its risks — which is precisely what this report tests. The engagement-over-fear result is therefore novel rather than unverified: it stands uncontradicted elsewhere because it has not been asked elsewhere.¹

What it means

The coalition for an AI worker-safety-net is led by the comfortable, not the exposed. The strongest supporters are graduate-degree holders (roughly twice the odds of high-school graduates), higher earners, AI users, men, and older workers — the same "engagement axis" the governance report identified, now confirmed against a direct fear control. The workers who most expect AI to take their jobs are, net of their circumstances, slightly less likely to want a government retraining program.

Why would fear reduce support? The data can't identify the mechanism, but the plausible readings are worth stating (as hypotheses, not findings):

• Fatalism. Workers who expect AI to eliminate jobs like theirs may doubt that retraining can save them — if the job is going away, a training program reads as beside the point.
• Efficacy and trust. Threat perception is highest among lower-SES workers, who tend to have less political efficacy and less trust that a government program would reach them — and the effect survives controlling for income and education, so it is not only composition.
• Engagement as informed demand. Support may track understanding of AI (which comes with use and education) more than exposure to its risks — people back the policy because they grasp the disruption, not because they personally fear it.

The interpretive caution the governance report already carries is now sharper: reading majority support for retraining as a demand "from the vulnerable" would be wrong. On this evidence it is a demand from the engaged and secure. That has a real forward-looking implication — if AI displacement broadens, the politically important question is whether the threatened begin to mobilize behind a safety net, or whether the coalition stays led by the advantaged who currently carry it. The two produce very different policy dynamics.

One consistency note: the pure-independent trough reappears here (OR 0.55) — the disengaged middle is the least supportive party group, exactly as in the reskilling and governance reports. Disengagement suppresses policy support at both ends: independents who don't follow politics, and — the new finding — workers whose main relationship to AI is fear rather than use.

Methods & verification

Data. AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project), the AI/employment supplement, weighted. Outcome policy_retrain_support = policy_retrain ∈ {support, strongly support} vs. {neutral, oppose, strongly oppose}; "not sure" (code 0) is absent in the data. Base restricted to workers (the fear items ai_loss_ultimate / ai_work_imp are asked of the employed, n ≈ 12,400; complete-case model n = 11,609). All figures are weighted estimates and are self-reports — what respondents say about their own use, expectations, and experience, not audited behaviour.

Measures. Fear (main): ai_loss_ultimate = "will AI eventually cause people who do work like yours to lose their jobs?" (Yes = fear). Fear (robustness): ai_work_imp = "only negative" impact on own job in 5 years. Engagement: uses_ai_work (uses AI at work at all), plus education_cat and income_cat_10 as the socioeconomic engagement axis. Controls: party7, age_cat_8, gender. All recodes direction-checked against the codebook; standard errors model-based, not design-based.

Discipline. Both fear coefficients are significant at p<.001 and negative across two operationalizations; engagement coefficients significant and positive. McFadden 0.064 is modest — most workers support retraining (64% of those who give an opinion; the analysis tool drops this item's "not sure" responses, so all-adult support is about 60%), so demographics explain a minority of the variation, which is normal for a broadly-popular policy item and is itself the point: support is wide, and what structures the remainder is engagement, not threat. The binary outcome treats "not sure" as non-support, which is the appropriate conservative coding for a support model. This is an observational association, described as such; no causal claim is made about why fear depresses support.

Reproducibility. Two weighted logistic regressions and one descriptive crosstab; recodes (policy_retrain_support, fears_jobloss, fears_negimpact, uses_ai_work).

Companion to Rein In the Titans, Cushion the Workers (engagement axis), Whose Job Is Next? How Workers See AI and Job Loss (fear channel), and the series index. This is the analysis the two reports set up but neither could resolve alone.