Superpersuasion

Happy Friday!

A big study landed this week, from Oxford, Stanford, the LSE, and the UK AI Security Institute, and the finding is: AI is now better at persuading people than expert humans are.

AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses.

They ran four experiments, 18,978 conversations across 6,923 people, on real UK policy positions and actual charity donations. The AI beat random laypeople, beat tournament-selected debaters, and beat elite debaters. In the study that involved real money it wasn't close: the models were nearly 3x more effective than professional fundraisers at getting people to donate to Save the Children.

I remember, in the late 1990s BJ Fogg started the Persuasive Technology Lab at Stanford and coined "captology," the study of computers as tools of persuasion. That lab trained a lot of the people who went on to build the engagement machinery of social media.

I'm not sure where it ends up, but this is all honestly really concerning.

👀 Interesting this week

Vibe Architects: Agentic Vibe Coders, Nielsen Norman Group.

The Prototype Is Done. What's Left to Design?, Ileana Marcut. The argument is that a vibe-coded prototype is now the brief itself, arriving polished and already settled in the requester's mind.

Still Holds: Gall's Law, Jorge Arango. Restates John Gall's rule that "a complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works".

Your design system runs on one person's judgment, AI is about to prove it,

The organizational cost of low taste, UX Collective.

Taste is the standard. Judgment is the act of applying it.

An Interview with Figma CEO Dylan Field About Design and AI, Stratechery. Figma's market cap fell from a $56B valuation at IPO to under $10B on the narrative that it's an "AI loser."

The CEO of AWS on why Amazon is hiring 11,000 interns and junior employees, Platformer. Matt Garman calls replacing junior staff with AI "one of the dumbest things I've ever heard" and a "non-starter for anyone trying to build a long-term company."

Product eats the AI company, Exponential View, on a Harvard/INSEAD working paper:

At similar funding and growth, AI natives are 25% smaller; they have more engineers, denser expertise and fewer managers.

Make AI Boring Again, Charity Majors.

AI is just technology.

Her argument against treating AI as uniquely evil or uniquely magic: every technology surfaces risks before we know how to govern them, and the job is ordinary engineering discipline, not panic or hype.

There Is No Such Thing as a Representative Sample, Saeideh Bakhshi. Demographic quotas work in political polling because age, region, and education correlate with the outcome (vote choice) and population margins are known. In product research the outcomes (adoption, retention, trust after a failure) are driven by behavioral and contextual variables instead, so matching a sample on demographics buys legitimacy, not validity.

AI and Liability, via Simon Willison. A German court ruled that Google's AI Overviews are Google's own words and that Google is liable for false answers in them. Schneier's framing:

AI agents are agents of the person or organization that deploys them, and should be treated by the law as such.

His point is that letting companies blame "faulty AI" would create an incentive to replace accountable humans with unaccountable systems.

Prompt Injection as Role Confusion. Research from Ye, Cui and Hadfield-Menell finds models can't reliably separate privileged text (in system/assistant role tags) from untrusted user input, and weight the style of text over its content.

destyling causes average attack success in our dataset to plunge from 61% to 10%.

The Magic 8-Ball vs. Gen AI,

Better search, worse web,

The State of AI, 2026, The Algorithmic Bridge. Reports the demand side turning: per The Information, customers are cutting OpenAI and Anthropic bills as unaffordable; Microsoft canceled internal Claude Code licenses; Uber capped monthly token spend at $1,500. Every engineering leader I talk to is trying to figure out what to do about token cost exploding.

The state of the AI economy. A bottom-up estimate of AI spending across consumer and enterprise: $110 billion in sales over the past 12 months, a run rate above $175 billion.

My Vibe Coding Adventure. Ben Thompson built a personal app himself by vibe coding and writes up ten takeaways on what the experience and the resulting app were actually like.

Slow down to speed up, Gergely Orosz. A conference keynote, now written up: AI coding tools are now used by essentially every engineer, and his case is that teams moving fastest with agents are the ones who slow down at the planning and review steps rather than letting agents run unchecked.

The new inner game, Lenny's. Joe Hudson, who coaches OpenAI's research team, argues the skills that matter as AI commoditizes effort and knowledge are emotional, not technical, and names four: discernment (decisions are fundamentally emotional, so emotional clarity improves them), "in conflict we trust," willingness to fail, and positive self-talk.

Building the most AI-pilled engineering team in the world, Lenny's. Podcast with Fiona Fung, who runs the Claude Code and Cowork teams at Anthropic (ex-Microsoft VS/TypeScript, ex-Meta Marketplace). She describes a team shipping roughly 8x more code than before, Claude "routines" changing how she manages, and the still-unsolved context-switching problem.

Structural Risk, Jorge Arango, links round up.

Two modernisms, UX Collective. Uses three examples (Mies van der Rohe hiding the steel, Harry Beck removing geography from the London tube map, Radix stripping component styles) to argue that abstraction is a recurring practical move designers make, later reframed as philosophy.

A reading list for different thinking, Saeideh Bakhshi. A grouped list for thinking in systems rather than objects, leading with Meadows' Thinking in Systems, Capra's The Systems View of Life, and Dörner's The Logic of Failure.

Why women need community during this AI shift.

Are Halland on how to do content with AI: pair writing.

Everyone is trying to document their design patterns into markdown files and such to feed the AI, and naturally there are interesting attempts to structure and standardize that stuff too.

Jason Cyr on Design Clarity.

Claude Code's lead designer on their design workflow. (Video)

Really good this week: "I Fed the People Building the Metaverse - On AI, male ego, and other reasons I no longer believe technology is magic". Got some killer quotes:

The engineers loved steak day. They lined up for steak day with the excited commitment of people waiting for concert tickets in 1987. They also loved cheese to a degree I can only describe as extensively Midwestern.

And this one

I made cheesecake for the metaverse while men discussed artificial intelligence across a counter from me like they were debating fantasy football statistics. I watched mediocre executives fail upward. I watched women get ignored, patronized, screamed at, and passed over. I watched corporations reward ego and punish humanity. I watched an industry convince itself it was inventing the future while reproducing every flaw of the present.

It's a fun read.