What Now? With Trevor NoahTristan Harris Is Trying to Save Us From AI! - What Now? with Trevor Noah Podcast
CHAPTERS
- 0:00 – 2:20
Bonus episode setup: Why bring Tristan Harris into the AI conversation
Trevor introduces the bonus episode and frames it as a third perspective in a recent run of AI discussions (after Sam Altman and Janelle Monáe). He positions Tristan Harris—known from The Social Dilemma—as a key voice on incentives and ethics in tech.
- •No episode next week due to the holidays; bonus episode instead
- •AI has dominated recent public debate and the show’s recent guests
- •Tristan Harris’ background: ex-Google, attention/design ethics, Center for Humane Technology
- •Goal: round out the AI conversation with an incentives-and-harms lens
- 2:20 – 3:28
From AI skeptic to AI ‘in-the-middle’: Christiana’s concerns about AI critics
Christiana describes swinging from skepticism (job displacement fears) to optimism (accessibility breakthroughs) and landing in a mixed stance. She questions what’s productive about prominent AI skeptics and whether it’s possible to meaningfully influence a technology that’s already spreading fast.
- •Writers’ strike shaped initial job-loss skepticism
- •Optimism driven by medical and accessibility use cases
- •Suspicion of “career AI skeptics” as a genre
- •Question: what’s to be gained if the ‘genie is out of the bottle’?
- 3:28 – 4:58
Trevor’s framing: guilt, responsibility, and the window to steer AI before it’s entrenched
Trevor compares tech reformers to figures like Alfred Nobel—people who build powerful tools and later feel responsible for downstream harms. He argues social media may be beyond easy reversal, but AI may still be within a steerable window before full-scale AGI arrives.
- •Pattern: insiders who later become critics/evangelists
- •Nobel/dynamite analogy: wealth, guilt, and redirecting incentives toward good
- •Social media ‘genie’ vs AI still in a scaling phase
- •Idea: there may still be time to intervene before AGI-level deployment
- 4:58 – 6:55
Humor and unease: AI surveillance examples and ‘AI as Nigerian dad’ riff
The conversation briefly shifts to comedic banter while spotlighting real AI monitoring uses, like classroom attention scoring in China. The jokes underscore a serious theme: AI already enables surveillance and behavioral control, not just harmless convenience.
- •Example: AI in classrooms monitoring fatigue/attention and grading students
- •Optimization mindset and surveillance creep
- •Comic riff: “AI is my dad” / Terminator reimagined
- •Transition into the formal interview with Tristan Harris
- 6:55 – 9:54
Harris’ core stance: not anti-tech—anti-business models that distort society
Tristan Harris rejects the “anti-tech” label and clarifies that his critique targets incentive structures—especially engagement/attention maximization. He explains ‘humane technology’ as design aligned with human and societal ergonomics, contrasting it with today’s “backache” of social media harms.
- •Center for Humane Technology’s vision is pro-human, pro-better-tech
- •Macintosh ‘bicycle for the mind’ ethos as inspiration
- •Ergonomics metaphor: tech should fit human needs, not exploit weaknesses
- •Catalog of social media ‘backaches’: mental health, democracy, addiction, exploitation
- 9:54 – 14:37
Origin story and ‘attention as steering’: the illusion of control with phones
Trevor probes Harris’ tech-industry background and what changed his thinking. Harris explains that technology ‘steers’ people by shaping attention and social expectations, and that our perceived control is often a magician’s illusion—made obvious by how hard it is to put the phone down.
- •Bay Area/Stanford/Google trajectory; design ethics focus
- •Correcting the ‘Burning Man epiphany’ stereotype; nature reflection instead
- •Attention is foundational to perception, choices, and beliefs
- •Smartphones as all-in-one devices: social signaling and constant pull
- 14:37 – 15:58
Design mechanics that hook us: ‘time well spent,’ Do Not Disturb, and social pressure
Harris explains how interface choices create new social norms (e.g., instant replies as proof of caring). He points to design interventions—like Do Not Disturb—as examples of mitigating harm by changing defaults and expectations rather than relying on individual willpower.
- •Tech design reshapes norms: perceived rudeness, expectations of immediacy
- •‘Time well spent’ and the missing middle option between always-on vs disconnected
- •Do Not Disturb as bidirectional signaling and harm-reduction design
- •Small design choices can reduce the “backache” even within current systems
- 15:58 – 19:52
Infinite scroll and doomscrolling: convenience feature turned attention weapon
The conversation unpacks infinite scroll—why it was invented, and how incentives transformed it into a mechanism for endless engagement. Harris and Trevor connect it to brain completion instincts, describing how platforms exploit unfinished-task feelings to keep users scrolling.
- •Aza Raskin’s invention: remove paging friction for search/blogs
- •Incentives repurposed it for social media addiction and ‘doomscrolling’
- •Completion instinct: near-finish triggers that reload to create perpetual “unfinished”
- •Trustworthy vs untrustworthy ‘progressive disclosure’ and the role of incentives
- 19:52 – 23:01
Incentives over moral labels: why ‘good vs bad’ is the wrong question
Trevor and Harris argue that debating whether social media is “good” or “bad” is immature and misses the point. The real predictor is incentive design: engagement maximization will systematically prefer outrage, addiction, and polarization because those drive attention.
- •Social media produces real benefits—but incentives shape the dominant outcome
- •‘Show me the incentive, I’ll show you the outcome’ logic introduced
- •Engagement incentives favor outrage, sexualized content, tribalism, anxiety
- •Attention steering alters collective decision-making and democratic trust
- 23:01 – 25:10
Context collapse and the funhouse mirror: San Francisco as an algorithmic case study
Trevor uses perceptions of San Francisco to illustrate how algorithms warp reality without necessarily spreading falsehoods. Harris highlights that “fact-checking” alone can’t solve a context problem when platforms cherry-pick true clips into a misleading, emotionally amplified narrative.
- •People form strong views from feeds without direct experience
- •Fear/outrage transmit more easily online than nuanced sadness
- •True videos can still mislead when sequenced into a high-dose pattern
- •Need for ‘context-checking’ and systems-level fixes beyond fact-checking
- 25:10 – 34:17
How to change incentives: collective action, regulation, and ‘BridgeRank’ shared reality goals
Harris argues no single platform can reform alone because competition creates a race to the bottom. He proposes treating major platforms as stewards of an “information commons,” with obligations to rank for shared reality and ‘unlikely consensus,’ citing Community Notes as a partial model.
- •Coordination problem: ethical actors lose without shared rules
- •Tobacco analogy: litigation and regulation shifted incentives at scale
- •41 states suing Meta as a potential lever for change
- •Information commons framing; BridgeRank/unlikely-consensus ranking; Community Notes origins
- 34:17 – 36:18
Protecting kids and the ‘do unto your own children’ test
The discussion turns to children’s exposure, sleep, and social pressure—drawing comparisons to store hours for alcohol and China’s “lights out” approach. Trevor and Harris argue that creators and executives restricting their own kids’ use is a moral signal that should guide policy.
- •Platform rules could reduce late-night pressure and sleep deprivation
- •Harris’ stronger stance: social media likely shouldn’t exist for under-18s
- •Executives’ own children often barred from the products they build
- •Principle: if you wouldn’t subject your kids to it, reconsider shipping it
- 36:18 – 40:09
AI becomes the new frontier: the arms race, ‘three laws of technology,’ and scaling to AGI
Trevor pivots to AI as the defining technology of the era, and Harris frames it as another uncoordinated race—now toward AGI. He explains why modern AI is uniquely risky: scaling compute and data produces unpredictable emergent capabilities (‘I know kung fu’ moments).
- •AI moved from ‘cute’ tasks to system-wide societal disruption potential
- •Three laws: new tech creates responsibilities; power creates races; uncoordinated races end in tragedy
- •Race target is AGI via scaling models, not just attention
- •Emergent capabilities appear without clear predictability as models scale
- 40:09 – 45:15
Why the switch happened: transformers, ChatGPT shock, and the acceleration spiral
Harris and Trevor trace the pivotal shift to the 2017 transformer breakthrough, making ‘just scale it’ the dominant path. ChatGPT’s release then triggered competitive panic—pressuring rivals to ship prematurely and, in some cases, open-source powerful models that expand access to misuse.
- •Transformers enabled general capability gains through scaling
- •ChatGPT catalyzed a market-wide ‘must ship now’ dynamic
- •Google and others reversed cautious stances under competitive pressure
- •Open sourcing (e.g., Llama) as a major accelerant and safety challenge
- 45:15 – 55:08
From misuse to catastrophic risk: bio/weaponization, open-source insecurity, and jailbreak reality
Harris argues that safety controls are often ‘security theater’ because models can be jailbroken and open-source variants can be fine-tuned cheaply to remove safeguards. He warns that AI can concentrate destructive power in the hands of individuals and that model theft undermines ‘beat China’ logic.
- •Examples: biological weapon instructions via jailbreaks; ‘grandma’ napalm prompt
- •Open source is ‘insecure-able’—safety layers can be stripped cheaply
- •Open models can help discover jailbreaks for larger closed models
- •Model security/theft risks (RAND concerns) weaken the arms-race justification
- 55:08 – 1:02:22
A governance model for ‘prudent optimism’: aviation-style safety, independent kill switches, and racing to safety
Harris emphasizes he’s pro-beneficial AI (medical advances, accessibility, projects like Earth Species) but argues deployment must match society’s ability to manage risk. Trevor and Harris point to FAA-style testing, grounding unsafe systems, and independent oversight (like rocket termination authority) as concrete precedents for AI governance.
- •Prudent optimism: accelerate benefits while controlling irreversible harms
- •Analogy to chemicals: ban/limit ‘forever’ harms rather than all chemistry
- •Aviation precedent: extensive testing, grounding fleets after failures
- •Proposal: independent oversight/early-termination authority for frontier training runs; shift from race to scale to race to safety
