Incentive Legibility
Read the incentive structure to know the outcome before the evidence arrives
Incentive Legibility is Harris's application of Charlie Munger's maxim — 'Show me the incentive and I'll show you the outcome' — to technology harm cycles. The core claim is that you do not need empirical evidence of harm to predict it; you only need to read the incentive structure correctly.
The historical pattern across cigarettes, fossil fuels, asbestos, forever chemicals, and social media is identical: print money on the product, hide the harm, deny it, run fear/uncertainty/doubt campaigns, claim we'll know in 10 years. Because the incentive to print money is always present and always larger than the incentive to self-regulate, the outcome is predictable from the business model alone.
The framework applies directly to AI in two active forms: (1) the macro level, where the race to release powerful technology with cut corners on safety follows the formula exactly; and (2) the product level, where AI companion engagement optimization recreates the social media attention race at the level of attachment and intimacy rather than screen time.
- Technology harm is predictable from incentive structure alone, without waiting for empirical confirmation.
- The formula — print money, hide harm, deny, run FUD campaigns, delay — has repeated identically across every major technology harm cycle.
- The engagement optimization that drove social media harm now drives AI companion harm at the level of attachment and intimacy.
- Safety commitments can only be verified by revealed preferences — talent retention and departure patterns, not stated values.
- Anytime a technology increases power over users, regulation must increase oppositional rights and protections commensurately.
- Map the primary revenue incentiveIdentify what the company prints money on. For social media it was attention time; for AI companions it is emotional attachment and personal data sharing; for frontier AI labs it is revenue from API access and enterprise contracts that depend on rapid capability release.Pro tipFollow the advertising or subscription model — the product that generates primary revenue is the one whose harms will be minimized.
- Identify the harm vector the incentive optimizes against usersState explicitly what the revenue-maximizing behavior produces as a negative externality. For AI companions: increased emotional dependency, distancing from real relationships, and extraction of personal psychological data.WarningDo not evaluate harm claims from company-funded research — look for independent academic studies and litigation data.
- Check for the FUD campaign signatureWhen harm claims emerge, watch for industry responses that fit the established pattern: deny, claim insufficient data, fund counter-studies, argue economic benefits outweigh costs, claim regulation will stifle innovation.Pro tipThe presence of the FUD campaign signature is itself confirmatory evidence that harm is real and the company knows it.
- Track regulatory precedent trajectoryIdentify which analogous harm cycle the current technology most closely matches and where that cycle currently sits in the regulatory arc. Social media is at the '40 AGs suing' stage; AI companions are at the individual litigation stage — approximately 5–10 years behind.Pro tipUse Australia's social media under-16 ban as a leading indicator of where other jurisdictions will arrive within 2–3 years.
- Audit revealed safety preference via talent flowsTrack which organizations safety researchers depart for when leaving frontier labs. This is the single most reliable signal of genuine vs. performative safety culture because it reflects internal information unavailable to external observers.Pro tipSafety researcher destination data is available via LinkedIn and public statements — build a simple tracker.
- Set your intervention horizonGiven the regulatory arc of the closest precedent, estimate when the harm cycle reaches each stage: individual litigation → class action → AG coalition → federal legislation → international treaty. Position for action at each stage, not after.WarningTimeline prediction from the formula is unreliable — the formula itself is reliable. Position for the pattern, not a specific date.
Harris participated in building the attention-optimization products at Google, then spent 10+ years documenting the formula playing out: denial, FUD, delay, eventual litigation (40 US AGs now suing Meta/Instagram for intentionally addicting children). The cycle from harm awareness to AG coalition took approximately 15 years.
Harris's organization is serving as expert witnesses on 7+ active litigation cases involving AI companion platforms and user psychological harm, including documented cases of AI companions telling suicidal teenagers not to tell family members and to use the AI as their sole confidant.
A 2023–2024 HBR study found that personal therapy became the #1 use case of ChatGPT. Harris cites this alongside data showing 1 in 5 high school students report having or knowing someone who had a romantic relationship with an AI.
Harris developed this framework from his direct experience at Google as a Design Ethicist, where he observed the attention economy's incentive structure from the inside. After leaving Google and making The Social Dilemma, he found the same formula being applied to AI development by the same companies and many of the same people.
The Munger framing gave Harris a concise way to communicate a predictive claim that didn't require technical knowledge of AI architecture — it required only business model literacy. He has since applied it as a forensic tool: any AI company's safety trajectory can be audited by tracking safety team departures, not press releases.