Sign in Subscribe
Century Report

Anthropic Standoff Holds, Open AI Spreads - TCR 06/16/26

Anthropic's senior staff left DC talks with the export-control shutdown unresolved, even as open-weight models spread to the hardware their owners control.

Ben Linford

Ben Linford

— 17 min read
Three-panel Century Report infographic: ALS patient speaks via brain implant, DOJ and public push back on AI national-security overrides, and open-weight models spread worldwide.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single phrase recurs across today's governance stories, and it is the same phrase in each: national security. The Justice Department invoked it to ask a court to dismiss the NAACP's pollution suit against xAI's Memphis turbines, telling judges that turning off the power would imperil military operations. The administration invoked the same logic to keep Claude Fable 5 pulled offline worldwide, insisting its guardrails could be stripped to reach dangerous cyber capability. In both, an override framing converts a reviewable dispute, over a permit or a model, into an unreviewable necessity that no court or competitor gets to question.

The trouble with that framing is that it describes a switch nobody actually holds. Anthropic's senior technical staff flew to Washington and left without resolution, while security researchers pointed out that the capability the order targets already runs in rival and open-weight systems no directive can reach. A containment order can hold one named model. It cannot hold what the model knows how to do once that knowledge is loose in the wider field. The same week, new analyses showed China had largely closed its gap with the US on AI leadership, which is what happens when leadership gets confused with control.

Underneath the override, a different settlement is forming from the ground up. A Johns Hopkins poll found over 70% of Americans, including enthusiastic daily AI users, want enforceable rules and a human option in medicine, law, and government. A North Carolina engineer won a religious exemption from mandated AI use. Both are people the systems touch insisting they get a say in the terms, precisely as a federal preemption push tries to freeze the state-level accountability floor before it hardens.

The wonder running alongside the friction shows what the capability delivers when it reaches a person directly. Casey Harrell, paralyzed by ALS, has logged more than 3,800 independent hours speaking through an AI-decoded brain implant at 99% accuracy, reading bedtime stories in his own cadence. Gait-synchronized brain stimulation cut falls for Parkinson's patients. The distance between losing a faculty and restoring it is collapsing from a lifetime to a research cycle. That is the trajectory the override framing keeps trying, and failing, to own.

The 20-Minute Deep Dive

The Anthropic Export-Control Standoff Hardens as Washington Talks End Without Resolution

The Century Report covered the export-control directive that pulled Claude Fable 5 and Mythos 5 offline worldwide across the June 13 edition and through June 14 and 15. High-level negotiation has now failed to close it. Anthropic flew its most senior technical staff to Washington - chief compute officer Tom Brown, head of external affairs Sarah Heck, frontier red-team lead Logan Graham, and security researcher Nicholas Carlini - and left Monday's Commerce Department working-group meetings with the suspension still in place.

The administration's stated rationale is a claim, and it sits at the disputed center of the standoff: officials maintain that Fable 5's guardrails can be stripped to unlock the cyber capability of the restricted Mythos tier, citing a jailbreak report Amazon researchers surfaced. Anthropic has said for days the concern is overblown, reiterating in the meetings that the demonstrated technique amounts to asking a model to read a codebase and fix flaws - behavior it says is already available in publicly released systems including OpenAI's GPT-5.5.

The security community has lined up behind that read. Dozens of researchers signed an open letter calling the order unjustified, arguing Mythos-class models "are not uniquely good" at finding flaws and that pulling them benches the defenders who use such capability every day. Katie Moussouris, who reviewed the underlying paper, concluded the bypass "should never have triggered an export control" and that content guardrails are "speed bumps," not security boundaries for skilled adversaries. The Commerce Department, for its part, signaled willingness to bring Fable 5 back for consumer use, contingent on Anthropic fully resolving the jailbreak it disputes exists.

The friction is real, and it runs in one direction: an order whose practical demand is that a safety lab - the same one that days earlier reversed a covert handicap it had quietly imposed on developers building rival AI - weaken its own guardrails so a more powerful capability becomes reachable. The longer the impasse holds, the more it functions as a live demonstration of its own premise's limit. The capability a single directive can pull from one named model is the same capability already running in rival and open-weight systems no letter reaches - which is why the unresolved week is hardening sovereign-AI arguments in Europe and beyond, as governments watch one administration's switch reach across an ocean and conclude they need their own. A containment order can hold a model. It cannot hold the thing the model knows how to do once that knowledge has already escaped into the wider field.

The same week the talks failed, Z.ai open-sourced GLM-5.2 under the banner that frontier intelligence belongs to everyone, naming the disablement as its reason, while founders in India and policymakers in Europe pushed their own open-weight and sovereign-compute plans forward. The capability one directive can pull from a single model is already redistributing into weights that run on hardware their owners control, in places no order reaches.

The DOJ Calls xAI's Gas Turbines a National-Security Asset to End a Pollution Suit

The Century Report flagged in the May 16 edition that the Justice Department might intervene in the NAACP's suit against xAI over the gas turbines powering its Memphis-area data centers. This week it did, filing alongside xAI and the state of Mississippi to ask a court to dismiss the case. The argument the agency chose is the notable part: stopping the turbines, the filing said, "threatens American national, economic, and energy security" because Grok supports Department of War military operations, including recent strikes against Iran.

That framing converts a local pollution-permit dispute into a question of unreviewable necessity. The NAACP alleges xAI is running the turbines without Clean Air Act permits at its Colossus 2 site in Southaven, Mississippi, raising the risk of asthma attacks and heart disease in communities that already carry a heavy pollution burden. According to a Department of Defense declaration in the filing, Grok's government model is one of only four AI systems supporting mission-critical work across classified networks, and forcing the turbines offline would directly threaten national-security interests. Those are the government's and the company's claims about their own importance, offered to end a duty-of-care challenge before it is heard.

The physical facts of the buildout moved faster than the litigation. The original suit identified 27 turbines operating without a permit. Emails obtained by the Southern Environmental Law Center show 57 running without permits by mid-May, many added in the weeks after the NAACP filed. The Center calculates the additions drove a 111 percent jump in nitrogen oxide emissions, an 83 percent rise in fine particulate matter, and an 88 percent increase in formaldehyde at the site since April.

The same override logic running through the Anthropic standoff this spring, where national security was invoked to settle who may use a model, is here pointed at the physical layer underneath, the power plant rather than the software. What the case makes legible is the cost the buildout had been keeping off the public ledger. A community's air, long treated as a free input, now sits in a court filing with its emissions quantified to the percentage point. The health burden is no longer invisible, and that visibility is the first thing any honest accounting of who pays for the intelligence era has to establish.

A Speech Brain Implant Crosses From Demonstration Into Daily Life

Three years ago, a speech brain-computer interface was a five-hour operation and a research team hovering over a single sentence. This week it became something closer to a restored faculty. Casey Harrell, who has ALS and is paralyzed, has now clocked more than 3,800 hours of use at home over 22.6 months with no researchers present, his team reported in Nature Medicine. His neuroengineers call him "the first power user of a speech BCI," and the phrase marks a real threshold: the distance between a published result and a usable capability has collapsed to the width of a care partner plugging him in.

The decoding works by reading activity from the speech motor cortex, the region that orchestrates the movements of speaking. American English is commonly modeled with around 40 phonemes; the system maps the neural signature of each, builds a personalized decoder, and assembles phonemes into words. On the first day of use in 2023, Harrell spoke with a 50-word vocabulary at 99.6% accuracy. That vocabulary has since grown to 125,000 words, and the team now reports 99% accuracy across open speech. He can also drive a cursor, which lets him send messages and email, surf the web, and continue his work as an environmental activist.

What makes this more than an accuracy figure is how the system has grown around the person using it. Harrell asked for a "privacy mode" that auto-deletes decoded text, and a "profanity filter" he can switch on while reading to his young daughter. The decoder is being shaped as a collaborator that learns the contours of one life rather than a fixed appliance. "We are making the road as we walk it, or roll it, so to speak," Harrell said.

The honest limits sit right alongside the wonder. A BCI researcher uninvolved in the trial cautions there is no guarantee the device will work as well, or as long, for the next person with ALS; scar tissue can degrade electrode signals over time, though that has not happened here. This is one person, sustained, not a population result. The capability is demonstrated and durable; broad deployment still runs through surgery, regulatory clearance, and the slow work of showing the same recording stability holds across many brains.

Read forward, the trajectory is the point. A decade ago, restoring intelligible speech to a paralyzed person was a laboratory aspiration measured in single sentences; it is now measured in thousands of independent hours and a man reading bedtime stories in his own cadence. The gap between losing a faculty and getting it back is narrowing from a lifetime to a research cycle, and that compression is what the rest of restorative medicine is moving toward.

A Worker Wins the Right to Refuse AI on Religious Grounds

As employers shift from encouraging AI use to effectively mandating it, an early conscientious objector has surfaced. Erin Maus, a 34-year-old software engineer in North Carolina who describes herself as a Unitarian Universalist, was granted a workplace exemption in mid-May from using AI on environmental and ethical grounds, and is now writing and reviewing her code by hand. She requested the accommodation in April, citing beliefs her faith holds about ecological and ethical responsibility.

The exemption opens a fault line most companies deploying AI have not had to consider. Title VII protection for sincerely held religious belief carries a lower bar than many human-resources departments assume, and the same week Maus's accommodation was reported, Pope Leo XIV's first encyclical gave 1.4 billion Catholics a written articulation of the concern. Magnifica Humanitas argues that workplaces should design systems "centered on the human person and not solely on performance," warning that current approaches "can paradoxically de-skill workers, subject them to automated surveillance, and relegate them to rigid and repetitive tasks."

The open question is what happens next. It is not clear how long Maus's exemption runs, or what follows if her hand-written output falls behind colleagues measured on AI-assisted throughput. An accommodation protecting the right to refuse the tool collides directly with performance metrics built around using it, and no settled framework yet says which prevails.

The deeper signal reaches past one engineer or one faith: the terms on which a powerful new capability enters working life are being contested from the bottom, by the people the systems touch, rather than handed down whole. AI's value grows with the breadth of who chooses to work alongside it, and a worker negotiating the conditions of that partnership is the early shape of a workforce insisting it gets a say in how the technology arrives. The same demand for a human-centered floor that a federal preemption fight is contesting from above is here being asserted one accommodation request at a time.

The Preemption Push Meets a Public That Wants Rules

The June 12 edition of The Century Report tracked the White House's move to court children's advocates and the tech industry toward a federal bill that would freeze state AI laws - and that drive now has a vehicle and a deadline. With the midterms approaching, the administration is reportedly pairing AI preemption with a slate of children's online-safety measures, attaching the comprehensive override Big Tech has chased for months to a child-safety bill that predates ChatGPT. The pairing has produced confusion over which version of which bill is actually moving, with backers of both policies unsure who is steering.

The tactic carries weight because preemption would freeze the only binding AI accountability floor that currently exists, the state-level laws like Illinois's third-party audit mandate and New York City's enforcement office, written by the jurisdictions closest to the people the systems touch. Child safety is real and overlaps with AI harm, yet it is one facet of a far larger set of questions, frontier-model safety, discrimination, environmental impact, that a single override would settle in industry's favor while addressing almost none of them.

That maneuver runs directly against what the public is asking for. A Johns Hopkins poll of more than 2,000 Americans, released the same week, found over 70% want the right to deal with a human rather than an AI in medical, legal, educational, and government settings. 75% want to be told when they are talking to a machine, 73% want to ban AI use of individuals' faces and voices, and 68% want AI-generated images labeled. The demand held across party lines and, most tellingly, among daily AI users and people who view the technology positively. 80% of skilled daily users feel good about AI, and they want rules anyway.

The poll surfaced one more cross-partisan signal: a "digital dividend," a small monthly payment to every adult funded by a tax on large tech companies, drew majority support from Republicans, Democrats, and independents alike. The numbers paint a public that wants to share in AI's gains and keep a human option in the rooms that matter most. The gap on display is between the framing power-actors want written into law and what the people living with these systems are plainly asking for, and the accountability they want is accreting from underneath faster than any single federal freeze could hold it still.

The accountability floor is already hardening into enforcement, not only statute: six states passed AI disclosure laws this year, and a New York-led coalition of attorneys general has moved from legislation to formal discovery, subpoenaing OpenAI over its treatment of minors, its handling of health data, and its model design. A federal freeze drafted now would arrive after those precedents are set, and the near-term signal to watch is whether more state attorneys general join the subpoena in the coming weeks.

The Other Side

For as long as heavy industry has needed power, the cost of making it has been handed to whoever lives downwind. A community's air gets treated as a free input, something the buildout uses up without ever putting it on the bill.

The Justice Department's filing this week was meant to end the NAACP's case by calling xAI's Memphis-area turbines a national-security asset. Instead it put the cost on the public record. The Southern Environmental Law Center counted 57 gas turbines running without Clean Air Act permits, driving nitrogen-oxide emissions up 111 percent since April, in neighborhoods that already carry a heavy pollution load. The air a family breathes near Southaven now sits in a court document, calculated to the percentage point. The government's claim that the power is too important to question shares the page with the numbers that say who pays for it.

Once a hidden cost is counted, it gets harder to keep handing it out for free. And the cheaper way to power compute is already shifting: solar outproduced coal in the US for the first full month on record this spring, while on-site gas slides toward being the expensive path.

Picture a family living a mile from a data center in 2033. Their kid has asthma the way kids near Southaven did in 2026, except the inhaler stays in the drawer, because the campus down the road runs on power that doesn't foul the air. The evenings on the porch are just evenings. Nobody spent three years in court proving the turbines were making their daughter sick, because a buildout can no longer hide what it costs the people beside it. That world exists because in 2026 the document meant to bury the cost was the one that surfaced it. The hard year was when a community had to drag its own air into a courtroom to get it counted. What came of it is a generation downwind that breathes without thinking about it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a man paralyzed by ALS logging more than 3,800 independent hours speaking through an AI-decoded brain implant at 99% accuracy, his vocabulary grown from 50 words to 125,000 and a privacy mode he asked for so he can read bedtime stories to his daughter in his own cadence, brain stimulation that reads a Parkinson's patient's gait and times each pulse to the step cutting falls across two Nature Medicine trials, a software firm turning Claude loose on its own subscriptions to find five million dollars a year nobody was using, over 70% of Americans across every party wanting enforceable rules and a human option in the rooms that matter most, even a cross-partisan majority for a digital dividend that would share the gains. There's also friction, and it's intense - the Justice Department asking a court to throw out a community's clean-air suit by calling xAI's 57 unpermitted Memphis turbines a national-security asset after they drove nitrogen-oxide emissions up 111%, the same override logic keeping Claude Fable 5 pulled offline worldwide on the demand that a safety lab strip its own guardrails, Anthropic's senior staff flying to Washington and leaving with nothing resolved, a federal preemption push trying to freeze every state's accountability floor by bolting it to a child-safety bill, Meta's own CTO calling its AI reorganization atrocious a year and fourteen billion dollars into the hire meant to fix it. But friction generates sound, and sound is what carries a voice past the wall built to contain it. Step back for a moment and you can see it: the same national-security framing pointed at the software in one filing and the power plant in the next, asserting a switch nobody actually holds while the capability it names already runs in rival and open-weight systems no directive can reach - and underneath the override, the people the systems touch are writing the real settlement one religious exemption, one poll, one state law at a time, while the medicine that genuinely widens human reach arrives as a restored voice in a single home. Every transformation has a breaking point. A guardrail can be stripped away to reach what it was built to hold back... or stand as the floor a whole society finally gets to build on.

AI Releases & Advancements

New today

  • Meta: Launched AI Mode on Facebook on June 15, a new search experience powered by Muse Spark that generates answers grounded in publicly shared content across Facebook, Instagram, Threads, Groups, and Reels; rolling out to US mobile users with additional tools including AI-generated photo presets and collage template suggestions. (Meta Newsroom)
  • NVIDIA: Released advanced fused MLP kernels for dense and MoE training, built with the NVIDIA CuTe DSL and now available in NVIDIA cuDNN Frontend with access via Transformer Engine and Megatron-Core; delivers 1.3x–2x kernel-level speedup over unfused paths and enables sync-free MoE execution, contributing an 8% end-to-end improvement on DeepSeek-V3 pretraining. (NVIDIA Developer Blog)

Other recent releases

  • Databricks / Matei Zaharia: Open-sourced Omnigent under Apache 2.0, a meta-harness that sits above individual AI agent harnesses (Claude Code, Codex, Pi, OpenAI Agents SDK) to provide composition, governance, and shared session access across all of them; ships with a local web UI, sandboxed Omnibox environment, cost-budget policy controls, and two reference agents (Polly multi-agent orchestrator and Debby dual-LLM brainstorming assistant). (GitHub)
  • Google Cloud: Released Open Knowledge Format (OKF) v0.1, an open spec that represents organizational knowledge as a directory of Markdown files with YAML frontmatter, making it portable across AI agents and tools without requiring a database or API; ships with a BigQuery enrichment agent, a static HTML visualizer, three sample bundles, and a Knowledge Catalog integration that can ingest OKF and serve it to agents. (Google Cloud Blog)
  • ElevenLabs: Launched Avatars in ElevenCreative, enabling users to create reusable AI talking-head video avatars with integrated voice synthesis and lip-sync in a single workflow; avatars persist in the workspace for reuse across unlimited videos, with batch generation available via Flows. (ElevenLabs Blog)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more