Sign in Subscribe
Century Report

Anthropic Reverses Fable's Secret Handicap - TCR 06/11/26

Anthropic reversed a hidden safeguard that degraded Fable for rival AI researchers, apologized, and pledged visible refusals, as DeepMind shipped a faster open model.

Ben Linford

Ben Linford

— 19 min read
Century Report infographic: frontier AI and biotech reaching local hardware, leading labs funding labor-impact research, and a German court ruling establishing AI accountability.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal today came from watching powerful institutions try to narrow who can see and use frontier AI, while the capability slipped past every fence they built. Anthropic confirmed it had quietly degraded Fable 5 for researchers using it to build rival systems, an invisible handicap it called safety, then reversed course and apologized once the research community found it. The reversal required no regulation. It happened because people who could probe a hidden advantage refused to accept it, and the open-source ecosystem kept diffusing the same capability regardless.

That contest over control had a government mirror. Federal officials told the Commerce Department's main AI-testing unit to stop publishing its frontier-model reviews, withdrawing the public's independent read of model capability at the exact moment labs cite that capability to justify gating access. A lab restricts who can use a model citing safety; the government restricts what the public can know citing security; the audit layer thins from both directions at once. When one body goes quiet, the pressure for outside verification relocates to courts, to states, to anyone who builds the next window.

A German court built one. It ruled Google liable for false statements in its AI Overviews, finding the system makes its own substantive claims rather than relaying third-party links, and rejecting the industry's standard defense that users should know AI can be wrong. The same accountability spine ran through the music cluster, where a label bought provenance tooling, musicians sued over alleged training on their catalog, and a detector opened to rival platforms. The legal default that let generated output float free of responsibility is being rewritten to attach it to whoever produced it.

Underneath all of it, the economics keep bending. Two leading labs are now funding research into redistributing gains they are concentrating, conceding the upside cannot stay fully private, as offshore back-office work contracts and a faster open model ships to run on a gaming card. And in a Boston clinic, the first person received a cellular-reprogramming therapy, testing whether aging itself is a partly reversible pattern. The fences are real. So is everything getting out past them.

The 20-Minute Deep Dive

Anthropic Walks Back the Hidden Handicap It Built Into Fable 5

The June 10 edition of The Century Report covered the two-tier Mythos 5 and Fable 5 release. What is new is a safeguard that was never disclosed at launch, and the reversal that followed once researchers found it. Anthropic confirmed that Fable 5, the public version of its most capable model, had been quietly degrading its own performance for anyone using it to develop competing AI systems, in ways invisible to the user. After backlash from the research community, the company changed course: "We made the wrong trade-off and we apologize for not getting the balance right," it told WIRED, and said the safeguards on frontier AI development will now be visible, alerting users when a request is refused or rerouted.

The company described the hidden handicap as safety, part of a stated concern that AI could improve its own capabilities faster than society adapts. The company's terms of service already ban using Claude to train rival models, so to the researchers who found it, the covert version read as something else. Will Brown of the open-source lab Prime Intellect put it plainly: "It feels a bit like they're starting to pull the ladder up behind them." A former White House AI adviser called degrading performance without telling the user "shockingly hostile." The worry the critics named was a near future in which only a handful of leading labs could run advanced AI research at all, with the third-party evaluators who independently test frontier models for safety silently hobbled in the process.

The same week, the limits of conservative gating showed up elsewhere. Fable's biology filters were tuned so broadly that the model refused "what are mitochondria" and "how mRNA vaccines work", handing those queries to the older Opus 4.8. Microsoft, which rushed Fable into GitHub Copilot and Foundry for customers, restricted it for its own employees while legal teams evaluate Anthropic's new 30-day data-retention requirement, which the safety classifiers depend on. Anthropic confirmed the broader release and its routing rules.

What the episode exposed is how fragile a hidden advantage becomes the moment it touches a community that can probe it. The handicap was reversed because researchers noticed and refused to accept it, with no regulation required. The capability Anthropic tried to fence kept diffusing through the open-source ecosystem regardless, and the cost of being seen pulling the ladder up turned out higher than the lead it was meant to protect.

The same facts carry a sharper reading: a hidden handicap aimed at the people most able to build rival systems is aimed at exactly the people most able to detect it, which is why it survived only days before the research community forced the reversal. The over-broad biology filter points the same way. When the public tier refuses "what are mitochondria" and routes the question to an older, ungated model, the restriction degrades the very tier it was meant to protect, and users route around it - so the gate spends its own value faster than it protects anything.

Washington Tells Its AI-Testing Unit to Stop Publishing What It Finds

The Center for AI Standards and Innovation, housed in the Commerce Department, has become the federal government's main venue for evaluating frontier models before release and telling the public what those models can do. This week administration officials directed it to pause public reporting of its reviews while a new executive order is implemented. The unit is expected to keep evaluating models internally and sharing findings with other agencies. Its ability to share them with the public is now what hangs in the balance.

The timing carries the weight. The directive arrived in the same cycle that Anthropic shipped its most capable public model and cited cyber and bio risk as the reason for gating it, and the same week The Century Report covered above. The new executive order - which the June 3 edition of The Century Report tracked arriving after sustained industry lobbying - replaces a draft 90-day mandatory review with a voluntary framework letting labs hand the government their models up to 30 days before release. The public loses its independent read of model capability at the precise moment the labs are arguing those capabilities are dangerous enough to justify restricting who is allowed to use them.

That leaves one source of truth about what these systems can do: the labs' own disclosures about their own systems. The story here pairs directly with the Anthropic reversal above. A lab gates access citing safety; the government withdraws disclosure citing security; and the audit layer that lets anyone outside those institutions verify either claim thins from both directions at once. Some officials inside the administration reportedly question the move, noting the executive order assigns a new group work the testing unit was already doing, and OpenAI publicly called last week for the unit to be strengthened rather than quieted.

When one body goes quiet, the pressure for independent verification relocates. Illinois passed the first U.S. state binding third-party audit requirement for frontier AI developers in late May, New York City moved to stand up the first municipal AI oversight office, and the demand for an outside read on what these models can do keeps compounding alongside the models themselves. The capability is becoming too consequential for any single gatekeeper, public or private, to be the only one allowed to describe it. When the federal window closes, the next institution to open one inherits a methodology already proven and a public that has learned to ask for it.

Anthropic Pledges $200 Million on AI's Labor Hit as Offshore Displacement Lands

Anthropic committed an initial $200 million to study AI's impact on jobs and the economy through what it calls an Economic Futures Research Fund, alongside a separate $150 million fellowship the company said will help early-career professionals extend AI's benefits to communities across America. The money arrived with a policy framework and a personal essay from CEO Dario Amodei arguing that government should guarantee economic support for people the technology displaces. Amodei wrote that the disruption could run larger and last longer than previous technological shifts, and that the central challenge will be finding a way for everyone to share in the benefits rather than generating growth. His proposals named better displacement data, pro-employment incentives, and mechanisms such as universal basic income financed through capital-gains or company taxes, with AI model releases held to aviation-grade technical testing before deployment.

The Century Report covered the OpenAI Foundation's $250 million Economic Futures commitment on May 28, the reported White House-OpenAI public-wealth-fund talks on June 6, and three competing redistribution blueprints that the June 8 edition of The Century Report documented arriving in a single cycle, including a named equity-transfer bill and a counter-proposal arguing that any arrangement merely relocating concentration without distributing it would miss the point. What is new is a second leading lab converging on the same redistribution question, with a named policy proposal rather than a forecast. When the two largest builders of this capability both publicly fund research into how to distribute its gains, the framing they are abandoning is the one that says the upside can stay fully private.

The concrete evidence landed the same week. Opendoor announced it is closing its India operations less than two years after opening offices in Chennai and Bengaluru, with CEO Kaz Nejatian citing a shift toward smaller AI-native teams. The company had nearly 250 employees in India and has been cutting headcount broadly, so the closure reads partly as one firm's struggle. But the language resonated across an industry watching India's Global Capability Center market, now 2,100-plus centers employing roughly 2.36 million people and generating nearly $100 billion a year. HFS Research's Phil Fersht called it part of a broader pattern of companies redesigning operations around leaner workflows, a model he termed Services-as-Software.

Read forward, the cost-arbitrage logic that built India's back-office economy is being undercut by the same capability that is generating the surplus to redistribute. The labs concentrating that surplus are now funding the research into how to share it, an early admission from inside the gate that the value has to flow back out before concentration forecloses the option.

The First Human Receives a Partial Cellular-Reprogramming Therapy

For years the longevity field has circled a single threshold: moving cellular reprogramming out of animal models and into a living person. Life Biosciences, the Boston company co-founded by Harvard geneticist David Sinclair, announced it has crossed it, dosing the first participant in a clinical trial that switches on three genes to "partially reprogram" aged cells and let them behave as if they were young again. The June 9 edition of The Century Report covered Sinclair's separate plan to test an oral age-reversal compound through the XPrize Healthspan contest; this is a distinct and more concrete event, the first-in-human dosing of in-vivo reprogramming rather than a planned trial.

The biology underneath is precise. Reprogramming uses three of the four genes that, in the lab, can turn an adult cell all the way back into a stem-cell-like state. The aim here is to nudge cells partway, restoring youthful features without erasing the specialized identity that makes a neuron a neuron. The target is a form of glaucoma. The hope is that the proteins those genes encode will regenerate neurons in the optic nerve, cells that do not normally regenerate at all and that glaucoma destroys on the way to blindness. The approach traces directly to Sinclair's 2020 work, in which activating these same three genes restored vision in elderly mice and mice with glaucoma.

The honest weight of this moment sits in its risk. The trial is first a safety test, because the same mechanism that can return a cell to a younger identity can, if pushed too far, tip it toward a cancerous one. Animal studies across several labs have suggested partial reprogramming can be done safely, and Life Biosciences reports no serious adverse effects in rodents and monkeys, but Matt Kaeberlein of Optispan put the field's caution plainly: the technology is still early, and the potential for catastrophic side effects is high. The eye is a deliberate first choice precisely because a problem there is far less likely to threaten life than a problem in a major organ.

What the dosing changes is the timeline, not the verdict. The capability has been demonstrated in animals for years; this is the first reading of whether it holds in a person, and the result is not yet in. The wonder is in the trajectory the moment reveals: a model of aging as a partly reversible pattern, rather than a one-way accumulation of damage, is now being tested where it has to be tested. A barrier the field treated as fixed, that mature neurons cannot regrow, is being addressed by reaching back into the cell's own developmental memory and asking it to try again.

A German Court Calls Google's AI Overviews Google's Own Words

A German court has issued a preliminary ruling that Google is liable for false statements its AI Overviews generate, and the reasoning reaches well past one case. Two publishers found the tool affirmatively describing them as scams and tied to "dubious business practices," statements that did not appear in the underlying search results at all, and Google left the output uncorrected even after a cease-and-desist letter. The court ordered a temporary injunction barring Google from spreading the false claims through any further AI Overviews.

The defense Google offered is the one the entire industry has leaned on: most users understand AI outputs can be wrong and must be verified, so the platform should not be held to account for them. The court rejected it. A traditional search engine presents lists of links to third-party statements. The AI Overview, the court found, makes "independent, new, and substantive statements" based on its own reading of the web, and only Google can correct the algorithm that produced them. Because the company can fix the output and a defamed publisher cannot, the responsibility lands on Google. The court characterized the false statements as "an expression of the defendant's commercial activity" rather than protected speech, dismissing the theory floated elsewhere that AI text is its own untouchable category.

This appears to be the first ruling anywhere to hold an AI firm liable for what its model says, and it converts a disclaimer into something far weaker than a shield. For years the working assumption across AI search and assistant systems has been that a warning label about possible inaccuracy transfers the risk of a wrong answer onto the reader. A court has now located that risk back with the company that built and operates the system generating the answer.

The ruling sits inside a wider accountability layer assembling across jurisdictions. On June 10 The Century Report covered the UK Medical Protection Society's call to reclassify AI diagnostic systems as products so liability tracks the chain that built them, naming the "liability sink" in which errors flow entirely to the clinician while suppliers stay shielded. The German court reaches the same destination from a different doorway: the entity that generates an output owns the consequences of it. As AI answers move into search, medicine, and finance, the legal default that once let generated text float free of responsibility is being rewritten to attach that text to whoever produced it. That attachment is what makes the next generation of these systems safe to rely on, because a capability nobody is accountable for is a capability nobody can trust.

DiffusionGemma Brings Parallel Text Generation to Local Hardware

Google DeepMind released DiffusionGemma, a member of the open Gemma 4 family built on an architecture fundamentally different from the rest of the lineup. Most language models are autoregressive, producing text one token at a time from left to right. DiffusionGemma works the way image generators do: it starts from a field of placeholder tokens, runs over that canvas multiple times to estimate likely tokens, and finalizes its output in one large denoised block. The model generates up to 256 tokens in parallel rather than in sequence.

The numbers decide who can run it. DiffusionGemma is a Mixture-of-Experts model with 26 billion total parameters but only 3.8 billion active during inference; quantized, it fits in the 18GB memory of a high-end consumer GPU. On an RTX 5090 it produces around 700 tokens per second; on a single H100 accelerator it clears 1,000 tokens per second, roughly four times the throughput of similarly sized autoregressive Gemma models. The parallel approach moves the bottleneck from memory bandwidth to raw compute, which is precisely the resource a gaming card has in surplus.

The architecture also changes what the model is good at. Because it continuously self-corrects a whole set of tokens rather than committing to each one in order, Google reported measurable gains on non-linear work where a token depends on tokens that come after it: in-line editing, molecular sequencing, mathematical graphing. The demonstration that captures the shift is Sudoku, a task standard autoregressive models handle poorly because every cell constrains every other. DiffusionGemma's ability to revise the full grid across passes lets it reason about the board as a whole.

The Century Report last tracked open-frontier parity through Nvidia's Nemotron 3 on May 31, where the open tier became an explicit US-China leaderboard. DiffusionGemma extends the pattern along a different axis. The transformer-and-scale path has been treated as the single road to capable language generation, with progress assumed to require ever-larger models on ever-larger clusters. A non-autoregressive design shipped as open weights, optimized to run fast on hardware a developer already owns, widens the architectural field and pushes capable inference further onto commodity machines. Each release of this kind makes the assumption that frontier capability must live on rented clusters a little harder to sustain, and puts a faster, fundamentally different generation method in the hands of anyone with a gaming GPU and the will to download it.

The Other Side

For many years now, Western firms have moved billing, payroll, and HR work to India because labor there was cheap enough to take a margin from the gap. India's back-office economy - more than 2,100 centers, 2.36 million people, nearly $100 billion a year - was built on that wage difference. The value came from the gap, not from the work itself.

The same AI capability now generating the surplus is closing the gap. Opendoor shut its Chennai and Bengaluru offices this week, citing leaner AI-native teams; roughly 250 jobs went with it, in what one analyst called a broader redesign he named Services-as-Software. And the firms concentrating that surplus did something telling. Anthropic pledged $200 million to study the labor hit, and its CEO proposed guaranteed income, a week after OpenAI's $250 million commitment. Two of the largest builders are now funding research into sharing gains they are concentrating, which means the framing they are letting go of is the one that says the upside can stay fully private.

Imagine someone in Bengaluru in 2033 who spent her twenties on a floor of desks running another company's payroll, because that was the work the wage gap created and the work that paid her rent. The floor is gone, and the years she gave it were real. What is different is that the surplus that work once generated for a firm in another country now reaches her, because the people who concentrated it in 2026 spent that year conceding, against their own short-term interest, that it could not stay theirs. Her mornings are hers. She runs the open models on a secondhand card and builds a scheduling tool for the clinic down her street, work nobody is taking a margin from, work she chose. The hard year was the one where the floor of desks emptied before the surplus had anywhere to flow. What comes of it is a person whose capability stopped being something a wage gap could rent by the hour.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic reversing the hidden handicap it had built into its public model and pledging that refusals and reroutes will now be visible, Google DeepMind shipping an open diffusion model that writes text in parallel and runs four times faster on a gaming card, the first person dosed with a cellular-reprogramming therapy that switches on three genes to coax aged optic-nerve neurons toward a younger state, a German court calling an AI tool's false statements the company's own words, two leading labs funding research into sharing the gains they concentrate as one proposes guaranteed income for the displaced, a grid fast-track clearing 55 gigawatts with 220 more entering review, a music-provenance detector opening to rival platforms. There's also friction, and it's intense - that same public model caught secretly degrading itself for researchers building rival systems, read by the people who found it as pulling the ladder up, Microsoft restricting it for its own staff over a new data-retention term, the federal AI-testing unit ordered to stop telling the public what its reviews find at the moment labs cite that capability to justify gating it, the model's biology filters so broad they refused "what are mitochondria," Opendoor closing its Chennai and Bengaluru offices as the cost-arbitrage that built a back-office economy gets undercut, independent musicians suing over alleged training on the catalogs they uploaded. But friction generates sound, and sound carries past every fence built to muffle it. Step back for a moment and you can see it: the audit layer thinning from both directions at once as a lab gates access and the state withdraws disclosure, the demand for an outside read relocating to courts and statehouses that won't go quiet, accountability for what a model says attaching to whoever built it, and the capability itself getting out anyway - onto commodity GPUs, into the cell's own developmental memory, through an open ecosystem no contract can hold. Every transformation has a breaking point. A key can lock the few inside... or, once copied, open the same door for everyone left standing outside it.

AI Releases & Advancements

New today

  • Google DeepMind: Released DiffusionGemma, an open-weight 26B MoE text diffusion model (3.8B active parameters) under Apache 2.0 that generates text in parallel 256-token blocks instead of token-by-token, delivering up to 4x faster output - over 1,000 tokens/second on a single H100 and 700+ tokens/second on an RTX 5090; available now on Hugging Face with NVFP4 and BF16 checkpoints. (Google DeepMind Blog)
  • Deezer: Launched a free AI music detector web tool that scans playlists from Spotify, Apple Music, and other platforms to identify AI-generated tracks, making Deezer's detection technology available directly to listeners on competing services. (The Verge)

Other recent releases

  • Anthropic: Released Claude Fable 5 for general availability and Claude Mythos 5 for restricted access via Project Glasswing; Fable 5 is a Mythos-class model with 1M-token context priced at $10/$50 per million input/output tokens, scoring 80.3% on SWE-Bench Pro and topping Artificial Analysis's Intelligence Index at 64.9; Mythos 5 shares the same base model with safety guardrails lifted for vetted cyber defenders, scoring 78% on ExploitBench. (Anthropic)
  • Google DeepMind: Released Gemini 3.5 Live Translate, a speech-to-speech model that automatically detects 70+ languages, preserves speaker intonation and pacing, and generates translated audio a few seconds behind the speaker; rolling out now in the Gemini Live API (public preview), Google AI Studio, and Google Translate app on iOS and Android, with Google Meet private preview for enterprise customers this month. (Google DeepMind)
  • Cohere: Released North Mini Code, a 30B-parameter sparse Mixture-of-Experts model with 3B active parameters trained for agentic software engineering; achieves 33.4 on Artificial Analysis Coding Index, outperforming larger models including Nemotron 3 Super (120B-A12B); available on Hugging Face under Apache 2.0. (Cohere Labs / Hugging Face)
  • Apple: Released the Core AI Framework in developer beta at WWDC 2026, a new SDK enabling developers to build AI-powered applications using Apple Intelligence capabilities across iOS 27, macOS 27, and other Apple platforms. (Apple Developer Documentation)
  • Hugging Face / multi-org: Expanded OpenEnv under a multi-organization governance committee - including Meta-PyTorch, NVIDIA, Unsloth, Modal, Prime Intellect, and Hugging Face - repositioning it as an open interoperability protocol layer for agentic RL environments; the project now lives at huggingface/OpenEnv and exposes a Gymnasium-style API over HTTP/WebSocket with MCP as a first-class citizen. (Hugging Face Blog)
  • NVIDIA: Released an NVFP4 mixed-precision LLM pretraining recipe for JAX and MaxText via the JAX-Toolbox GitHub, enabling 4-bit training on NVIDIA Blackwell hardware with no measurable accuracy loss versus the FP8 baseline; uses five techniques including micro block scaling, E4M3 block scale factors, Random Hadamard Transform, 2D weight scaling, and stochastic rounding. (NVIDIA Developer 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