The Adjacent Brief

TL;DR: Deepfake detection is breaking down at the expert level — the NYT profiles Hany Farid, the leading digital forensics authority, who says he can no longer reliably identify AI-generated images. AI nudification tools are creating a new vector for child harassment that schools and parents have no good answer to yet. Stripe's new agent-native payment infrastructure and a wave of AI-linked layoff anger round out the picture: downstream consequences of fast AI deployment are doing most of the work today.

Worth Reading

• Germany's activists are trademarking neo-Nazi symbols before extremists can (paywall) — Using IP law offensively to deny commercial use of hate symbols — a creative and replicable tactic worth watching in brand and legal contexts.
• AI design isn't ugly. It's fluent — and that's the problem. — The more interesting critique of AI-generated design: it's not bad, it's frictionlessly average, and average at scale is its own problem.
• Electric grids, not chips, are the real bottleneck in the AI buildout — Jigar Shah makes the infrastructure constraint case plainly: you can fund compute, you can't fund a substation in 18 months.
• The scam economy is bigger than Denmark's GDP — and growing — $442 billion in global fraud losses in 2025; AI is lowering the cost of personalized deception at the exact moment trust infrastructure can't keep up.
• Ted Gioia got sold an AI slop book impersonating a real author — First-person account of retail platform failure: AI-generated books mimicking real authors are getting through discovery, recommendation, and purchase without friction.
• Cross-screen attention measurement finally has a framework — Evan Shapiro's cross-screen audience index is the first attempt to measure total viewer attention across platforms — the metric ad buyers have wanted for a decade.
• Azeem Azhar's data: top 1% US firms spend 650x more per employee on AI than typical firms — $7,450 per employee monthly vs. $11 — that gap is the most concrete number on AI's consolidating returns available right now.

Brand & Growth

The job market is sorting by AI fluency, not AI enthusiasm

Recruiters posting marketing roles have stopped asking whether candidates "understand AI" and started asking whether they can run it. Analysis of 300 monthly marketing job postings shows the concentration pattern: performance marketing, content operations, and analytics roles are absorbing the most volume, and specs in those roles increasingly assume tool proficiency rather than request it. For anyone hiring or job-hunting: the generic "AI-curious" candidate is already oversupplied. The scarce profile is someone who closes a loop — uses a model, measures the output, iterates.

Platform identity anxiety is real, but the pivot logic is thin

LinkedIn rolling out short-form video and YouTube leaning into social features sounds like strategic adaptation; this week's marketing read frames it more usefully as identity confusion. LinkedIn becoming Instagram risks the one thing that makes LinkedIn worth paying for — the professional context that makes B2B targeting defensible. YouTube chasing social engagement metrics risks diluting the reason creators build there: monetizable long-form attention. Platforms copying each other's surfaces without copying their underlying user intent tend to produce worse versions of both products. Brand teams buying on these platforms should watch whether audience behavior actually follows the format changes, or whether users keep doing what they came to do.

The layoff optics problem isn't going away

The TechCrunch piece on AI layoffs becoming a powder keg lands on something brand leaders should take seriously: the anger isn't just about job losses, it's about the juxtaposition. When AI company insiders sell equity during periods of mass redundancy, the narrative writes itself. For any company doing headcount restructuring while publicly celebrating AI productivity gains, the communications problem is structural — no press release fixes a visible tradeoff between people and compute.

Connected World

The browser privacy stack is about to change for most users

Google is disabling Manifest V2 extensions in Chrome — which means uBlock Origin and most other content blockers stop working for the majority of users who haven't switched browsers. The technical argument (MV3 is more secure) is real but secondary; the commercial argument (Chrome is Google's primary distribution surface for its ad business) is obvious. For brand and media buyers, this is an expansion of addressable inventory — more users exposed to more ads on Chrome properties. For anyone building privacy-forward products or positioning against surveillance advertising, the default browsing experience just got less filtered.

Hardware constraints are getting solved in the wrong order

Semi-solid-state gel batteries are getting traction as a near-term bridge between lithium-ion and true solid-state chemistry — better energy density, less fire risk, manufacturable at existing scale. This is useful and incremental, but the Verge piece is honest that full solid-state remains years out. The headline technology keeps landing as future-tense while adjacent solutions do the actual work. For EV makers and energy storage buyers, the practical bet remains on chemistry that ships, not chemistry that impresses at a demo.

Open fabrication is moving from hobbyist to manufacturable

OpenCAL — an open-source implementation of computed axial lithography 3D printing — is notable less as a product than as a capability unlock. CAL printing produces objects without layer lines, at speeds that make conventional FDM look slow, and with material properties closer to injection molding. Open-sourcing the approach puts it on the cost curve of every prior open fabrication technology: fast commoditization, wide experimentation, then a few breakout applications nobody anticipated. The manufacturing complexity-as-moat argument gets harder to sustain when the process documentation is on GitHub.

Culture & Signal

The expert floor just dropped out of deepfake detection

Hany Farid has spent two decades as the person institutions called when they needed to know if an image was real. The NYT piece on his current situation is worth reading not for the profile but for what it implies operationally: the world's most practiced human detector is no longer reliable. That's a measurement of how fast generation quality improved. Anyone building trust systems, legal processes, or content moderation on the assumption that expert human review catches what automated systems miss needs to revisit that assumption.

Child safety is the AI harm story that isn't moving fast enough

The Wall Street Journal's reporting on AI nudification tools being used to bully children (paywall) is a story about institutional lag. Schools don't have the policy infrastructure; parents don't have the detection tools; platforms are playing catch-up with takedown requests. The tools themselves are cheap, accessible, and improving on the same curve as everything else in generative AI. The response infrastructure is not. The gap between harm velocity and response velocity is where the risk sits, and it's a problem no single institution can solve unilaterally.

Censorship pressure creates its own infrastructure

Reuters' reporting on Russian users running dual phones and VPN stacks documents Russian internet users circumventing state restrictions — and something worth watching at a structural level: when a state restricts access aggressively enough, it doesn't eliminate demand, it creates a parallel technical literacy. Russian internet users are now, by necessity, among the most practically sophisticated VPN and evasion users in the world. The parallel to Chrome's MV3 move is loose but present: when default infrastructure closes off access, some portion of the population builds workarounds, and those workarounds often outlast the original restriction.

The New Consumer

Safety systems fail when users are motivated to defeat them

Chinese Tesla drivers are using plastic figurine heads to fool the driver attention monitoring system — a $3 fix for a system Tesla spent engineering resources building. The Wired piece is framed as a Tesla story, but the pattern generalizes: any safety or compliance system that imposes friction on users who don't want it will be socially engineered around, especially when the workaround is cheap and shareable. This is the same logic that governs ad blockers, account-sharing crackdowns, and DRM. The question for product teams building compliance systems is whether the cost of defeating the system is high enough that most users won't bother — and $3 plastic heads suggest Tesla's current answer is no.

Rideshare workers know exactly what's coming

The Rideshare Guy's survey of what Uber and Lyft drivers actually think about autonomous vehicles is behavioral data from the people closest to the displacement risk, and it cuts against the narrative that workers are either unaware or unconcerned. Drivers understand the timeline, they're skeptical of platform reassurances, and they're making income diversification decisions now. This connects to the layoff tension in the Brand & Growth section above: the AI disruption story isn't future-tense for the people living inside it. The awareness gap that comfortable analysts describe doesn't exist at the worker level.

Fraud at scale is a systemic design problem, not a user education problem

Global scam losses hit $442 billion in 2025 — larger than Denmark's GDP. The Next Web piece makes the AI connection explicit: generative tools are lowering the cost of personalized, high-credibility fraud at exactly the moment social trust infrastructure was already under pressure. Solving this through consumer awareness campaigns is wrong at this scale. Fraud this large is a platform architecture problem and a financial system design problem. The institutions that treat it as infrastructure risk — rather than a user literacy gap — will be better positioned when regulators come looking for explanations.

Commerce Rewired

Stripe is building for agents as first-class buyers

Stripe's new infrastructure for AI agent purchasing — specifically enabling agents to buy cloud infrastructure programmatically, without human-facing pricing pages or checkout flows — is a cleaner signal than most AI infrastructure announcements. The bet is that agent-to-agent commerce will need its own payment rails, and Stripe wants to own them before the category has a name. The companies building durable positions in the enterprise AI arc are the ones instrumenting the pipes that models run through. For anyone pricing or selling services that AI agents might eventually purchase, the question of whether your pricing architecture is machine-readable just became more concrete.

13 articles across 5 themes · 12 sources · Powered by Folo + Claude