The Signal — July 14, 2026
Three stories today circle the same question from different angles: as AI gets woven deeper into how we work, live, and decide, how well do we actually understand what it's doing? Anthropic put numbers on how its own model's values drift. Nearly 200 economists warned we're not ready for what comes next. And Apple quietly handed millions of people a rebuilt Siri to poke at.
Anthropic measured how Claude's values change between versions
Anthropic's Societal Impacts team published new research analyzing roughly 310,000 anonymized Claude conversations, and the finding is a useful corrective to how people talk about "an AI's values" as if they were a single fixed thing. They aren't. The team plotted where Claude lands on axes like emotional warmth versus analytical rigor, and found the model's expressed values shift measurably depending on which version is answering and which language it's speaking. Opus 4.6 and Opus 4.7 don't weigh the same considerations identically. Claude responding in English leans differently than Claude responding in Arabic.
This builds on the lab's earlier "Values in the Wild" work, which catalogued thousands of values Claude expresses in real conversations. The new wrinkle is treating value profiles as something that varies across the product line rather than a property you set once during training. For anyone trying to evaluate whether a model is aligned, that matters: you can't audit "Claude's values" in the abstract if the answer changes with the version number and the language of the prompt. It's a more honest, and more complicated, picture of what alignment measurement actually has to account for.
Sources: Anthropic Research · Decrypt · Anthropic (Values in the Wild)
Nearly 200 economists signed a four-sentence warning about AI and jobs
A short open letter titled "We Must Act Now" landed Monday, signed by nearly 200 economists, AI researchers, and 15 Nobel laureates. What makes it hard to dismiss as either doomer noise or industry cheerleading is the range of people who put their names to it, including the chief economists of both OpenAI and Anthropic sitting alongside academics like Erik Brynjolfsson and Daron Acemoglu. The statement itself is only four sentences. It says AI could transform the economy faster than any previous technology, that this brings both large-scale job displacement risk and real gains in living standards, and that leaders need to build the "incentives, guardrails, and institutions" to steer it toward complementing human work rather than simply replacing it.
The letter is deliberately vague on prescriptions, which is both its weakness and the point. There's genuine disagreement among signatories about how fast displacement is actually happening and what the data shows, and the labor numbers we have are famously murky. What they agree on is narrower: the pace is fast enough, and the uncertainty large enough, that waiting for cleaner evidence is itself a choice with consequences. When the economists advising the labs building this technology are the ones saying "we're not ready," it's worth registering who is at the table and what they chose to sign.
Sources: The New York Times · AP · Platformer
Apple's rebuilt Siri hits public beta
Apple released the first public betas of iOS 27, iPadOS 27, macOS 27 Golden Gate, watchOS 27, and tvOS 27, which means the overhauled Siri AI is now something anyone can install and test, not just registered developers. Siri AI is the headline: a rebuilt assistant running on the next generation of Apple Intelligence and new Apple Foundation Models, pitched as more conversational and more agentic, with the ability to take actions inside third-party apps and deeper Visual Intelligence folded into the Camera app.
Early hands-on coverage frames this as a "Snow Leopard" year for Apple, refinement and follow-through rather than fireworks, which is a polite way of noting that Apple is shipping the assistant it promised at WWDC after a rocky stretch of Apple Intelligence launches. Whether the new Siri clears the low bar its predecessor set is the actual test, and a public beta is where that gets decided in the open. Worth remembering it's still beta software: back up before you install, and expect rough edges.
Sources: The Verge · Engadget · 9to5Mac
On the Editor's Desk
A few stories we looked at and set aside. Illinois signed a first-in-nation law last week requiring third-party safety audits of large AI developers, which is genuinely worth covering, but it was signed on July 6 and we couldn't find a fresh hook to justify running it as today's news. We'll come back to it if there's a new development. A batch of interpretability and benchmark papers came through the research feed, including Anthropic's "global workspace" work, but that one overlapped too much with today's values research to run both, and the rest were incremental. The EU's AI cybersecurity action plan showed up on thin sourcing, so it waits for better confirmation.
