228 – Eileen Yam on how we're completely out of touch with what the public thinks about AI

1) Core thesis

AI experts and the general public live in radically different epistemic worlds — experts see AI as empowering and productive, the public sees it as an erosion of human agency and connection — and this gap, combined with the public’s lack of confidence in either industry or government to govern AI, sets the stage for a collision between the industry’s deployment ambitions and the public’s appetite for control.

2) Claim and Evidence

• Claim: The expert-public gap on AI’s benefits is staggering — and experts are wildly overestimating public exposure.

  • Evidence: 73% of experts see positive impact on jobs vs. 23% of the public. 69% vs. 21% on the economy. 74% vs. 17% on productivity. Meanwhile, 80% of experts think the public interacts with AI “almost constantly or several times a day” vs. 27% who actually say they do. Only 34% of US adults have ever used ChatGPT [00:16:25, 00:51:49].
  • Strength: strong. These are Pew data with large samples (N≈5,000 for public, N≈1,000 for experts), probability-based sampling. The gaps are far outside any margin of error.

• Claim: The public’s #1 AI concern is erosion of human abilities and social connections — not job loss or existential risk.

  • Evidence: When asked in free-text responses, 27% of those seeing high risk cited erosion of human competencies. A teacher respondent captured the core anxiety: “children are digital natives, [but] adults who understand a world without AI need to still pass the torch to children for developing these human qualities with our own human brains” [00:20:03].
  • Strength: strong. Free-text responses avoid priming effects. The consistency of this theme across demographic groups (including the young) suggests it’s not an artifact of unfamiliarity.

• Claim: Younger people — the most exposed demographic — are the most concerned about AI degrading human capabilities.

  • Evidence: 61% of adults under 30 say AI will worsen creative thinking vs. 42% of those 65+. 58% of under-30s say AI will worsen ability to form meaningful relationships vs. 40% of 65+. This runs counter to the intuitive expectation that exposure breeds comfort [01:28:58].
  • Strength: strong. Counterintuitive finding that holds across multiple measures. Suggests familiarity with AI doesn’t resolve concerns — it may amplify them.

• Claim: Public concern about AI is not comparable to early concerns about the internet or smartphones — it is fundamentally different in character.

  • Evidence: Pew’s 1999 polling on the internet showed 62% liked having access to information; concerns were about connection speed and privacy, not about human agency. AI is “ambient” — people interact with it unknowingly, unlike smartphones or browsers which require affirmative opt-in. Yam also notes the cultural priming effect: respondents invoked Terminator and 2001: A Space Odyssey as touchpoints [00:44:52].
  • Strength: moderate. The historical comparison is informative but asymmetric — the internet didn’t promise to automate cognition. The “ambient” quality is a real distinguishing factor.

• Claim: The public is not polarized on AI along partisan lines — yet.

  • Evidence: Both Republicans and Democrats are skeptical of government or industry handling AI effectively. The only partisan gap is degree: 70% of Republicans have little/no confidence in government regulation vs. 54% of Democrats. No significant differences in AI use, excitement, or concern by party [01:28:58].
  • Strength: strong. This is notable given the extreme polarization on nearly every other policy domain. It implies AI attitudes are not yet crystallized into partisan identities.

3) Mechanisms

Yam’s data reveals two distinct mechanisms driving public attitudes. First, domain sensitivity: the public distinguishes sharply between AI in impersonal, high-stakes domains (weather forecasting, financial crime detection, drug discovery — broadly accepted) and personal, identity-linked domains (matchmaking, religion, political speech — strongly rejected). This is not a simple “pro-AI” vs. “anti-AI” binary but a domain-specific permission structure. Second, control asymmetry: the public’s anxiety is amplified by the perception that AI is ambient and inescapable (“I don’t even know when I’m engaging with AI”) — unlike smartphones or the internet, which required deliberate opt-in. This creates a baseline distrust that colors responses even to beneficial applications.

The expert-public gap is explained in part by selection effects (people who expect AI to benefit them enter the field) and in part by informational asymmetry (experts see their own productivity gains and project them onto everyone, vastly overestimating public usage). The fact that exposure correlates with concern rather than comfort among young people suggests a third mechanism: experiential learning — using AI may reveal its limitations and risks in ways that distant observation does not.

4) Concrete actions

• For AI companies: recognize that the public’s desire for “more control” (61%) is the dominant sentiment. Build products that visibly preserve human agency and decision rights rather than automating them away. The data shows people want AI as a tool, not as a replacement — design for augmentation, not substitution.
• For policymakers: the public wants regulation but trusts neither industry nor government to do it. This is a legitimacy vacuum that independent, transparent auditing bodies could fill — but only if they are perceived as genuinely independent from both industry capture and political agendas.
• For AI communicators: address the erosion-of-human-capabilities concern directly. The framing “AI empowers you” is contradicted by the public’s lived intuition that delegating cognitive tasks atrophies those skills. Acknowledge this tension rather than dismissing it.
• Track the same individuals longitudinally: as AI exposure increases, do people become more comfortable or more anxious? The cross-sectional data (young people = more exposed = more concerned) weakly suggests the latter, but panel data would be decisive.
• Invest in AI literacy efforts targeted at the 66% of US adults who have never used ChatGPT. The “not sure” third of respondents represents an undecided bloc whose views will be shaped by first experiences and salient news events.

5) Delta vs prior episodes

The Ezra Klein episode (2023-07-24) on this same channel argued that policy action requires a crisis — a “punctuation” of the equilibrium. Yam’s data updates this in two ways. First, it identifies what the crisis might be: not a technical catastrophe (the AI risk community’s focus) but a social one — a mass experience of disempowerment or eroded relationships that crystallizes public anger. Second, Yam’s finding that younger, more-exposed people are more concerned suggests the crisis may not require a discrete “event” at all — it could build gradually as exposure spreads, making the punctuated equilibrium model less predictive.

6) Red flags

• The expert sample (N≈1,000, drawn from conference presenters/authors) skews young (80% under 45) and includes 40% current students. This is not a representative sample of “AI experts” — it overweights junior researchers and underweights industry leadership. The expert-public gap may be overstated for this reason; more senior experts might hold views closer to the public’s on certain dimensions.
• The survey defines AI with a single sentence (“designed to learn tasks that humans typically do”), which respondents may interpret very differently. Someone thinking of a Terminator-style autonomous agent will answer differently from someone thinking of a spam filter.
• The “erosion of human abilities” concern, while genuinely held, may be partially a status-quo bias effect. Similar concerns were raised about writing (Plato), calculators, and search engines. The question is whether AI is different in kind or merely the latest target of a recurring anxiety pattern.
• Yam emphasizes that a third of the public answers “not sure” on key questions. This means many of the headline percentages are drawn from the ~67% who have formed opinions — potentially overrepresenting those with stronger priors.

7) Open questions

• Is the age gradient (younger = more concerned) a cohort effect or a lifecycle effect? Will today’s concerned 25-year-olds become comfortable 45-year-olds, or will their concerns intensify with further exposure?
• What happens when the 66% of Americans who have never used ChatGPT have their first meaningful interaction? Will it shift them toward expert optimism or toward public concern?
• The public is not politically polarized on AI — yet. What event or framing would change this? A major labor displacement? A deepfake election scandal?
• How much of the expert-public gap is explained by the expert sample’s youth/student composition vs. a genuine informational or experiential divide?