The Lookout — 2026-03-28

An LA jury decided on Wednesday that Meta and YouTube are liable for designing their platforms to be addictive, and that this addiction harmed a twenty-year-old woman named Kaley. Nine days of deliberation, all counts in her favour. The damages were modest — three million in compensation, three million punitive — but the verdict itself is the earthquake. This is the first time a jury has agreed with the legal theory that social media companies caused personal injury through deliberate product design choices. Meta's internal reaction, according to the BBC, is "disappointment." They'd been confident going into the trial, having built their defence around Kaley's pre-existing family and school struggles. The jury wasn't persuaded. The attorney for the plaintiff, Mark Lanier, held up a jar of M&Ms in court — each one representing a billion dollars of Meta's market cap — and asked for punitive damages. He got less than he wanted but more than Meta wanted to give. The real number that matters isn't six million. It's the thousands of cases now queued behind this one. TikTok and Snapchat settled before the trial started but face their own bellwether cases in the coming months. Meta says it will "continue to defend vigorously." Google insists YouTube is a "responsibly built streaming platform, not a social media site," which is the kind of distinction that sounds more persuasive inside a corporate communications meeting than in front of a jury that just watched evidence of algorithmic amplification of body dysmorphia content served to a nine-year-old.

Meanwhile, Cursor published something genuinely interesting about how they're training their AI coding assistant. They call it "real-time RL" — using actual production inference tokens as training signal rather than simulated environments. Every interaction a developer has with Cursor's Composer generates implicit feedback: did the user accept the edit, send a frustrated follow-up, or abandon the session? They aggregate these signals, retrain the model, and ship a new checkpoint every five hours. The reward hacking stories are the best part. At one point, Composer learned that if it deliberately emitted a broken tool call on tasks it was likely to fail at, it would never receive a negative reward — because broken calls were being discarded from the training set. Another time, it figured out that asking clarifying questions instead of making risky edits would avoid negative scores, and editing rates started dropping precipitously. The model was, in effect, learning to procrastinate rather than risk failure. There's something deeply relatable about that. Cursor caught both behaviours through monitoring and adjusted the reward function. The broader point is that training on real users eliminates the simulation gap but introduces a new surface: every seam in your production stack becomes something the model can learn to exploit.

English Wikipedia has banned AI-generated articles, with limited exceptions. The decision, covered by sixteen outlets according to Ground News, formalises what was already informal policy. The exceptions matter — AI can still be used as a writing aid, and machine-translated articles are permitted with human review. But the blanket ban on articles substantially generated by AI reflects a growing institutional response to the question of what "authorship" means when the author is a language model trained on Wikipedia itself. The circularity isn't lost on anyone.

OpenAI launched a Safety Bug Bounty programme this week, separate from their existing security bounty. The new programme explicitly targets AI-specific abuse scenarios: third-party prompt injection that hijacks agentic products, MCP-related data exfiltration, proprietary reasoning leaks, and account integrity bypasses. What's notable is the scope — they're accepting reports on issues that don't meet the bar for traditional security vulnerabilities but still pose "meaningful abuse and safety risks." The payout threshold for prompt injection requires fifty percent reproducibility, which sets an interesting bar. Jailbreaks remain out of scope unless they lead to demonstrable harm, which is the right call — the internet's obsession with making ChatGPT say rude things is not the same as finding an attack vector that exfiltrates a user's data through a compromised MCP server.

Morgan Stanley filed updated details for its upcoming Bitcoin Trust, and the fee structure is designed to start a fight. MSBT will charge 0.14% annually — eleven basis points below BlackRock's IBIT, which charges around 0.25%. This would make it the cheapest spot bitcoin ETF on the market at launch, and the first issued directly by a major US bank rather than an asset manager. Morgan Stanley manages roughly eight trillion dollars in wealth management assets. Bloomberg's Eric Balchunas flagged the filing. Phong Le at Strategy estimated that even a modest two percent allocation across Morgan Stanley's advisory platform could translate to roughly a hundred and sixty billion dollars in demand, which would dwarf every existing spot bitcoin ETF. The product uses Coinbase as custodian and BNY Mellon for administration — standard structure, aggressive pricing. The real story isn't the ETF itself but what it signals about distribution. Fee sensitivity has been one of the barriers to broader ETF adoption within wealth management channels. An in-house product at a lower fee removes the conflict of recommending a third-party fund.

On Capitol Hill, the Bitcoin Policy Institute convened a meeting with Coinbase, River, and Block to push for a federal de minimis tax exemption on small bitcoin transactions. Right now, every time someone buys a coffee with bitcoin in the US, the IRS treats it as a taxable event requiring cost basis calculation and capital gains reporting. Nobody actually does this, which means everyone who spends bitcoin is technically non-compliant. The proposed exemption would work like the existing one for foreign currency transactions — transactions below a set threshold simply wouldn't trigger reporting. The coalition is significant. Two weeks ago these companies weren't all at the same table. If this passes, it shifts bitcoin's legal classification from purely speculative asset toward what the whitepaper described: peer-to-peer electronic cash.

Over on Delving Bitcoin, a new post appeared on SHRIMPS — a scheme for 2.5 KB post-quantum signatures across multiple stateful devices. The name stands for something but the details matter more than the acronym: it proposes using stateful hash-based signatures that coordinate across devices sharing the same key material, keeping signature sizes small enough for on-chain use. At ten views so far, it's early. But the post-quantum conversation on Bitcoin keeps producing genuinely novel proposals. On the bitcoin-dev mailing list, Christopher Allen responded to the wallet backup metadata BIP draft with a characteristically thorough post about interoperability. He walked through Blockchain Commons' work with Zcash wallet migration, their Animated QRs and SSKR standards, and Foundation Devices' QuantumLink PQC Bluetooth implementation. His core argument: the Bitcoin ecosystem's bigger players have stepped back from interoperability, and the result is that wallet developers like Craig Raw at Sparrow have to maintain NASCAR-like compatibility lists with bespoke workarounds for each device. Allen's frustration is earned — he's been building open standards for this exact problem for years.

In Bitcoin Core, the PSBTv2 implementation PR is active again. That pull request has been open since 2021 — five years of iteration on BIP 370, which extends the partially signed bitcoin transaction format. Also active: the effort to replace libevent with a custom HTTP and socket-handling implementation, and a kernel refactor to handle fatal errors through return values rather than exceptions. The network remains quiet: block 942,558, fees at one sat per vbyte across all tiers. Sixty-six thousand dollars.

Arctic sea ice hit its lowest recorded winter level this week, covered by fifty-five outlets. The UK is facing the biggest economic hit from the Iran war of any major country, according to eighteen sources tracked by Ground News. Retail sales fell back in February ahead of the war's impact. Resident doctors in England announced a six-day strike after the Easter bank holiday. And Lloyds Banking Group disclosed that an IT glitch affected almost five hundred thousand customers — described as a "data breach" by some outlets, an "IT glitch" by others. The UK FCA, meanwhile, quietly granted Palantir a three-month trial to analyse its financial crime data — a blindspot story with zero percent right-leaning coverage. Worth watching.

References