Video BreakdownGeek12 April 2026

The AI Dilemma: Social Media's Architects Warn History Is Repeating

The guys who called social media's harms before Congress now argue AI is repeating the same pattern — faster, with higher stakes and fewer guardrails.

Tristan Harris & Aza RaskinCenter for Humane Technology1h 15m5.1M viewsWatch original

Top Claims — Verdict Check

AI deployment is following the exact same "deploy first, fix later" pattern as social media

🟢 Real

“Representative of their position: social media was the first contact with AI and we lost — now the same race dynamics are playing out with large language models, but at compressed timescales.”

AI has already surpassed 50% of humans at persuasion tasks

🟡 Partially True

“Representative of their position: AI systems can now generate persuasive text that outperforms the majority of human writers in controlled studies, and this capability is accelerating.”

We have a narrow window — months, not years — to establish AI safety norms

🔴 Hype

“Representative of their position: the window for meaningful intervention is closing fast because once capabilities are deployed and integrated into business models, pulling them back becomes nearly impossible.”

AI will make the cost of disinformation effectively zero

🟢 Real

“Representative of their position: generating personalized disinformation at scale used to require state-level resources — now it costs pennies per message with LLM APIs.”

Existing institutions and regulatory bodies are structurally incapable of governing AI at the speed required

🟡 Partially True

“Representative of their position: the gap between AI development speed and government response time is not a policy failure — it is a structural mismatch that existing institutions cannot bridge.”

What's Real

The social media parallel is the strongest argument in the talk, and it holds up. Facebook knew its algorithm amplified polarisation by 2018 (internal "divisiveness" research, leaked in the 2021 Frances Haugen disclosures) and shipped it anyway because engagement metrics demanded it. The same incentive structure — deploy capability, capture market share, address harms later — is playing out at OpenAI, Google, Meta, and xAI in real time. The disinformation cost argument is also concrete: generating 1,000 unique, targeted phishing emails cost roughly $50,000 in human labour in 2020. By March 2023, when this talk was recorded, GPT-3.5 could produce equivalent output for under $5 in API calls. That cost has dropped further since. The 2024 election cycles in the US, India, and Indonesia all featured documented AI-generated disinformation at scale — this prediction landed.

What's Hype

The "months, not years" urgency framing was recorded in March 2023. It is now April 2026. The window did not close. Meaningful AI governance work has happened since — the EU AI Act passed, executive orders were signed (and some revoked), and voluntary commitments were extracted from labs. Imperfect, yes. Absent, no. Harris and Raskin compress the timeline to create urgency, which is effective advocacy but poor forecasting. The "50% persuasion" claim references a specific study on GPT-3 persuasive writing, but strips the context: the study measured short-form text persuasiveness, not real-world behaviour change. Generating convincing text is not the same as changing minds at scale. The leap from "AI writes persuasive paragraphs" to "AI will manipulate populations" is plausible but unproven at the scale they imply.

What They Missed

The economic incentive layer for AI safety is absent. Unlike social media — where safety and engagement genuinely conflict — some AI safety measures (hallucination reduction, output accuracy, content filtering) are commercially valuable because enterprise customers demand reliable outputs. Anthropic, for example, has built a business model where safety is the product differentiator, not the cost center. The open-source AI movement gets no airtime. Meta releasing Llama weights to 700 million+ downloads represents a structural counterweight to concentration of AI power that does not fit the "small group controls everything" narrative. And the Global South perspective is entirely missing — AI governance framed exclusively through a US/EU lens ignores that countries like India, Brazil, and Nigeria are deploying AI in governance, agriculture, and healthcare with different risk-benefit calculations.

The One Thing

The social media playbook — deploy fast, capture market, apologise later — is the default template for AI deployment, and recognising that pattern is worth more than any specific technical prediction.

So What?

If your product uses AI-generated content facing users, you already have a disinformation surface — audit it before a journalist or regulator does
The EU AI Act risk categories are the best available framework for assessing your own AI compliance exposure, even outside Europe — read the high-risk classification list
Social media's lesson is that voluntary safety commitments from labs are marketing until enforced — build your product assuming regulation will arrive, not hoping it won't

Action Items

1Run a "bad actor audit" on your AI features: spend 30 minutes trying to make your AI product generate misleading, harmful, or reputation-damaging content. Document what you find. If you can do it, someone else already has.
2Read the Center for Humane Technology's "AI policy recommendations" page (humanetech.com/policy) — it is a 15-minute read that provides a structured framework for thinking about AI harms that applies directly to product decisions.
3Map your AI product's output chain: from prompt to user-visible result, identify every point where false or misleading content could be generated, cached, or amplified. Prioritise the highest-reach touchpoint for a guardrail.

Tools Mentioned

ChatGPT

Referenced as primary example of rapid public AI deployment and persuasion capability

GPT-4

Cited as the capability frontier at the time of the talk — benchmarked against human performance

Bing Chat

Microsoft's AI search — referenced as example of premature deployment causing public trust damage

Workflow Idea

Build a quarterly "AI trust audit" for your product. Take your three highest-traffic AI-powered features. For each: (1) attempt five adversarial prompts designed to produce harmful or misleading output, (2) log results and severity, (3) compare against the previous quarter. Track your failure rate over time. This gives you a defensible safety record, surfaces real vulnerabilities before users or regulators find them, and takes about two hours per quarter.