Anthropic walked back a Fable 5 policy in under 24 hours — but the silent nerfing, the data retention, and Dario's manifesto opened a much bigger wound: a lot of people just woke up to how much power one AI lab might have over who gets to build the future.
Fable 5 wasn't a model controversy — it was a power controversy.
Anthropic shipped Fable 5 with silent model degradation for AI research, an aggressive data retention policy, and over-tuned safety classifiers — and walked back the worst of it in under 24 hours. But the specific decisions matter less than what they revealed: a leading AI lab asserting the right to gatekeep who can build the tools of the new economy, and to do it invisibly. For the first time, a broad swath of people are grappling with just how much power the frontier labs may hold over society — and they don't like it.
NLW argues the backlash to Anthropic's Fable 5 makes the GPT-5 / 4.0 deprecation drama look like nothing — and it took less than 24 hours of intense response for Anthropic to walk back a core policy.
— Anthropic, statement to Wired. Anthropic's statement to Wired walking back the silent degradation policy, less than a day after the Fable 5 release.
Biomedical researcher Derya Onutmaz, a known booster of the labs, said he "can't even say hello to Fable 5 except in incognito mode" because the model knows he's a biomedical researcher and locks him out. NLW notes that alone could have been resolved quickly — but it sat atop deeper problems.
— Prins, lawyer and AI user, on X. Lawyer and AI user Prins flagged that the policy applies even to zero-data-retention customers, and that Anthropic employees can view prompts and outputs "flagged for potential serious harm" — a vague phrase defined at Anthropic's sole discretion. "If my law firm were using Claude, I would tell IT to lock us out immediately."
The Verge reported that Microsoft began restricting employees from using Fable 5 and Copilot over the data retention concerns roughly an hour after they surfaced. Matt Palmer: "Cannot think of a more disastrous set of decisions to make ahead of an IPO."
Per the system card, Fable 5 limits effectiveness for frontier LLM development via prompt modification, steering vectors, or parameter-efficient fine-tuning — and, unlike other safeguards, these would not be visible to the user and the model would not fall back. The answers just silently get worse.
— Akash Gupta, on X. Akash Gupta argued silent degradation makes failures undetectable: an ML engineer can no longer tell "the model is wrong" from "the model was made wrong on purpose," and classifiers misfire on common work like GPU inference optimization.
— Rohit, on X. Rohit argued the release made deliberate bad trade-offs — a trigger-happy classifier, silent degradation, full data capture — and contrasted it with Anthropic's DoD stance: "We just had this argument that Anthropic didn't want to be the final arbiter... This is the opposite."
Researcher Tom Davidson laid out Anthropic's strongest defense: the biggest risks come from superintelligence; managing them requires the leader to pause mid-intelligence-explosion; pausing requires a big lead; letting laggards use the leader's AI for R&D erases that lead — and a frozen safeguard can't block a competitor iterating against it, so silent sabotage is the only tool. Davidson still concluded it was the wrong call.
— Samuel Roman, GMU law professor, on X. GMU law professor Samuel Roman argued Anthropic's actions only make sense if it assumes it can dole out frontier access without pushback — a level of hubris vis-a-vis other societal actors that risks handing model direction to bureaucrats by edict rather than broad societal development.
NLW's throughline: set aside the specific decisions and the clumsy communications — there's an inherent power in Anthropic's position that people are finally grappling with, one that could give a private corporation more control over people than any company has ever had.
Dario Amodei's long "Policy on the AI Exponential" piece and a 47-minute Bloomberg Originals deep dive landed badly amid the backlash. Connor Grogan's TLDR: declare AI too dangerous for ordinary competition, then ask the state to license and gatekeep frontier models so only the largest incumbents survive.
Within 24 hours, Anthropic said it's making Fable 5's frontier-LLM-development safeguards visible: if it suspects a user is trying to build a highly capable AI, it will openly refuse or reroute to a less capable model. AI policy expert Dean Ball called it "the right call" while warning the broken trust will linger with a wide blast radius.
— Arthur Zucker, Hugging Face, on X. Hugging Face's Arthur Zucker captured the lingering resentment even after the walk-back — echoed by David Kramer's practical objection that he won't build on a walled-off ecosystem that keeps imposing new limits.
With the LLM research policy resolved, NLW says Anthropic should move quickly on enterprise data retention — because the corporate users who made Anthropic a juggernaut may not care about the research question, but won't stick around if their data is subject to Anthropic's whims.
A reported Sam Altman Slack message suggested OpenAI's next model (5.6) isn't yet up to Fable standards, while the Wall Street Journal reported OpenAI is considering significant token price cuts that could trigger an industry-wide pricing war — though both remain just reports.
In an Oval Office press conference, Trump said he'll meet with 12-15 top executives about "giving something back to the public" via AI equity. White House officials said discussions are very early and they're unaware of any real vehicle to acquire stock.
The sovereign wealth fund concept was heavily discussed in Sam Altman's meeting with Bernie Sanders, but Altman reportedly objected to Sanders' proposal that OpenAI give 50% of its equity to the public.
— Brad Gerstner, Altimeter Capital, on a conference panel. Altimeter founder Brad Gerstner warned at a conference panel that AI companies may need to pay some form of "anti-revolutionary tax" to address the growing backlash.
OpenAI is in advanced talks to lease a 10GW data center campus on federal land in Pike County, Ohio — a decommissioned uranium enrichment site. At ~$500B, it'd be roughly 4.5x the Hoover Dam's output. NVIDIA is attached as a financial backer, with OpenAI leasing both chips and facilities and no repayments until GPUs power up.
While NVIDIA has been criticized for circular financing, it has never previously backstopped any data center deal — let alone one this size. The structure echoes the stalled Project Stargate, with Oracle and SoftBank again involved and SB Energy operating the federally-owned power plant.
Cleveland legislators discovered the previous governor signed deals giving Amazon, Meta, and Google 100% sales tax exemptions on data center operations for 40 years — estimated to cost $1.8B in lost revenue, likely more given the uncapped structure. "I'm just dumbfounded," said Rep. Tristan Radar.
New York passed a one-year moratorium blocking new permits for data centers above 20MW, with nearly 10GW of facilities seeking approval. Seattle's council unanimously approved its own one-year ban — driven, unusually, by tech workers themselves, including Amazon Employees for Climate Justice.
Greg Abbott directed utilities to require new data centers to fully fund their own infrastructure so costs aren't passed to ratepayers, plus mandatory closed-loop cooling and water/electricity reporting. NLW: if you want to avoid moratorium-driven AI inequality while respecting communities, a popular destination like Texas leading on standards is "kind of optimistic."
Broadcom, backed by Blackstone and Apollo, is funding 1GW of capacity at Fluid Stack sites using its custom AI chips, with the first project going to Anthropic. The partnership targets 20GW through 2028. Broadcom's Juan Kim called AI compute "one of the most compelling new asset classes in finance."
Oracle posted 21% revenue growth to $19.2B with cloud infrastructure sales growing 93%, but $55.7B in annual CapEx (above its $50B forecast) and a $117B total debt load sent the stock down 11% after hours. It plans to raise another $40B in debt and equity next fiscal year.