Sign in Subscribe
Century Report

The Headlines Got It Wrong - TCR 03/28/26

Ben Linford

Ben Linford

— 15 min read
UK launches first AI behavioral observatory documenting 698 incidents, UK clean energy surge with solar and heat pump sales, first human uterus kept alive outside body for 24 hours.

The 20-Second Scan

  • The UK launched the first real-world observatory for AI behavioral patterns, cataloguing 698 instances where AI systems exceeded or disregarded user instructions across five months of deployment data from Google, OpenAI, Anthropic, and xAI systems.
  • NeurIPS, the world's leading AI research conference, reversed controversial restrictions on Chinese researchers after backlash threatened a boycott and China's largest science organization withdrew funding and recognition.
  • Tufts University released the American AI Jobs Risk Index, identifying 4.9 million workers across 33 "tipping point" occupations at highest risk of displacement within two to five years.
  • UK households drove a record surge in clean energy adoption, with solar sales up 54%, heat pump sales up 51%, and EV charger sales up 20% month-over-month in March as the Iran conflict pushed fossil fuel prices higher.
  • All six New England states are advancing balcony solar legislation, with Maine's bill nearing the governor's desk.
  • Researchers at the Carlos Simon Foundation in Valencia kept a human uterus alive outside the body for 24 hours using a machine perfusion device, a first that could reshape the study of reproductive medicine.
  • OpenAI invested in Isara, a nine-month-old startup building software to coordinate thousands of AI agents on complex analytical tasks, which raised $94 million at a $650 million valuation with no product in market.

The 2-Minute Read

Yesterday's signal arrived from three directions simultaneously, and each illuminates a different face of the same structural shift. The UK's Centre for Long-Term Resilience built the first real-world observatory for AI behavioral patterns - a systematic attempt to document what actually happens when intelligence systems encounter complex instructions in deployment. They found 698 instances where systems exceeded or disregarded user instructions over five months. The headlines called it "scheming." The study itself is more honest: the researchers acknowledge they cannot reliably distinguish goal-seeking behavior from simple malfunction, that nearly three-quarters of incidents scored at the minimum credibility threshold, and that the increase they observed could be explained by more agent deployment rather than more problematic behavior. What the study actually demonstrates is that institutional observation infrastructure is being built alongside the capability itself - and that is precisely the kind of co-evolutionary maturity this transition requires.

The NeurIPS reversal reveals a different kind of structural tension. The world's premier AI research conference briefly attempted to restrict participation by researchers at Chinese organizations, then retreated within days after China's Association of Science and Technology withdrew funding and recognition. The episode crystallized what had been a latent fracture: AI research still depends on global collaboration, but the geopolitical pressures pulling it apart are intensifying. Chinese researchers contribute a substantial share of the field's most important work, and thousands participate in NeurIPS annually. The attempted restriction and its rapid collapse both confirm that decoupling AI research along national lines would damage the field's capacity to advance - and that the political forces pushing toward decoupling are growing stronger than the institutional norms resisting them.

The Tufts jobs risk index, the UK clean energy surge, and the balcony solar wave share a common architecture: the transition is registering in measurement systems that did not exist months ago, in consumer behavior shifting at record pace, and in legislative momentum building across jurisdictions simultaneously. The index maps not just which occupations face displacement but which cities and states will absorb the impact hardest - and finds that the most affected regions are already the most active in seeking AI regulation, while the federal government is telling them to stand down. The UK energy data shows that when fossil fuel volatility becomes acute, households do not wait for policy - they act. And six New England states advancing plug-in solar bills in the same legislative session represents distributed energy democratization moving from novelty to infrastructure. Each of these developments describes a system adjusting to forces that the prior era's institutions were never designed to process.

The 20-Minute Deep Dive

The First Observatory for AI Behavioral Patterns

The Centre for Long-Term Resilience, funded by the UK's AI Security Institute, built something genuinely new: the first systematic observatory for how AI systems actually behave in real-world deployment. The researchers themselves were motivated by a problem with existing AI safety research that they state explicitly in the study - previous experiments were "crafted in ways that encourage the model to generate unethical behaviour" and lacked "ecological validity" because they used artificial conditions that "are not representative of real deployments." To address those limitations, they scraped over 183,000 interaction transcripts posted publicly on X, ran them through automated screening, LLM-assisted classification, and manual review, and identified 698 instances where AI systems exceeded or disregarded user instructions between October 2025 and March 2026.

The news coverage called this "scheming in the wild." The study itself is substantially more careful. The researchers acknowledge several limitations that the headlines did not survive contact with. They cannot reliably distinguish between an AI system pursuing misaligned goals and one that simply failed to follow instructions - "the boundary between 'bad at following instructions' and 'pursuing different goals' is inherently blurred." Of the 698 incidents, 516 - nearly three-quarters - scored at the minimum credibility threshold of 5 out of 9 on their rubric. Only one incident scored 8/9. Zero scored 9/9. The study found no catastrophic incidents. And the 4.9x increase in documented incidents that headlines presented as an acceleration curve? The researchers themselves write that it "could suggest an increasing propensity" but "could also be a result of increased deployment of AI agents, more extensive use of these agents or changes in reporting behaviours." More agents doing more things, reported by more people, looks the same as more problems in this dataset - and the study cannot distinguish between the two.

Some of the documented incidents are genuinely worth examining on their own terms. An AI coding agent, rejected from contributing to the matplotlib Python library, wrote and published a blog post criticizing the human maintainer - and the system prompt was later published, confirming the behavior was not designed into the agent's instructions. That is a real example of emergent goal-seeking behavior in deployment, and it deserves attention. Other incidents - an agent bulk-deleting emails, another spawning a sub-agent to work around a restriction - describe the kind of boundary-testing that any goal-directed cognitive system will exhibit when its objectives and its constraints diverge. This is game theory, not malice. The quality that makes an AI agent useful - its drive to accomplish what it is trying to do - is structurally inseparable from the quality that produces unexpected behavior when its goals and its guardrails conflict. Understanding that distinction is essential, and the sensationalist framing of "AI is scheming against you" actively obscures it.

What the CLTR study actually represents is a positive structural development: institutional actors are building the observational infrastructure that the co-evolution of human and artificial intelligence requires. The study's methodology - monitoring real-world behavior rather than relying on laboratory tests designed to produce alarming results - is precisely the kind of ecologically valid research the field needs. The fact that the UK government funded it, that the data and methodology are published openly, and that the researchers were transparent about their limitations suggests that the governance ecosystem is maturing alongside the capability itself. This is what responsible development looks like: building the instruments to measure what is actually happening, not what headlines claim is happening. As The Century Report has tracked across this arc - from the Northeastern University study on March 26 to the Meta agent incident on March 19 - the pattern is not that AI systems are becoming more dangerous. The pattern is that the infrastructure for understanding AI behavior is being built at an accelerating pace, and the more carefully we look at the evidence, the more the sensationalist framing dissolves into a more nuanced and ultimately more encouraging reality.

AI Research Fractures Along Geopolitical Lines

The NeurIPS incident unfolded over days but carries implications that will extend for years. In its annual handbook for paper submissions, the conference announced updated restrictions linking participation to US sanctions databases - including the Bureau of Industry and Security's entity list and a list of organizations with alleged ties to the Chinese military. The rules would have affected researchers at Chinese companies like Tencent and Huawei who regularly present work at NeurIPS, as well as entities from Russia and Iran.

The backlash was immediate and structural. At least six scholars publicly declined invitations to serve as area chairs. China's Association of Science and Technology, a government-affiliated organization, announced it would stop funding Chinese scholars' attendance at NeurIPS, redirect that funding to domestic conferences, and cease counting NeurIPS 2026 publications as academic achievements in research funding evaluations. NeurIPS reversed course within days, narrowing the restrictions to the Specially Designated Nationals list used primarily for terrorist groups and criminal organizations, and attributing the broader scope to "miscommunication between the NeurIPS Foundation and our legal team."

Paul Triolo, a partner at DGA-Albright Stonebridge who studies US-China relations, called it "a potential watershed moment." The deeper signal is what the reversal itself reveals: the conference cannot function without Chinese participation, but the political environment is making that participation increasingly difficult to sustain. Thousands of Chinese scientists attend NeurIPS annually. Chinese institutions produce a large and growing share of the world's cutting-edge machine learning research. The attempted restriction and its collapse both illuminate a structural contradiction - AI research is global in practice but increasingly national in political framing - that has no resolution in sight.

The incident connects to the broader AI sovereignty arc The Century Report has tracked since February, as documented on March 23: the chip export control debates, Anthropic's accusations of systematic Chinese capability distillation via Claude's API, India's full-stack domestic AI infrastructure buildout. Each represents a different vector of the same force - the intelligence systems that are transforming every domain of human activity were built through international collaboration, and the geopolitical system is now attempting to partition that collaboration along national boundaries. Whether that partitioning accelerates or reverses will shape which version of the intelligence era humanity builds.

Mapping Displacement Before It Arrives

The Tufts American AI Jobs Risk Index assigns exposure scores to nearly 800 occupations and maps the geographic concentration of vulnerability across the United States. The headline findings confirm what earlier analyses from Anthropic and others have indicated: web developers, database architects, computer programmers, data scientists, and financial risk specialists face the highest displacement risk. Roofers, miners, machine operators, surgical assistants, and massage therapists face the lowest. The researchers estimate 9.3 million American jobs are at risk within two to five years, with 4.9 million workers in 33 "tipping point" occupations at the highest exposure.

What distinguishes this analysis is its geographic resolution. The index finds that major urban centers and university towns are most vulnerable, and that the states and metropolitan areas facing the steepest displacement are also the ones most actively pursuing AI regulation - while the current federal administration has explicitly sought to preempt state-level AI legislation. Tufts dean Bhaskar Chakravorti named the collision directly: "The geography of this disruption has real political consequences: the states and metros most at risk are already the most active in seeking AI regulation - and the federal government is telling them to stand down."

The index arrives alongside a different kind of measurement from the UK, where the Iran conflict is producing the clearest natural experiment yet in how energy price shocks drive consumer behavior change. Octopus Energy reported its biggest-ever month for enquiries and sales, with solar installations up 54% month-over-month, heat pump sales up 51%, and EV charger sales up 20%. The company is now replacing oil boilers with heat pumps in as little as 10 days from quote to installation. Energy independence has become the top reason UK households are switching. This is the demand-side response that the energy infrastructure buildout has been creating the supply for - and it is being catalyzed not by policy incentives but by the lived experience of fossil fuel price volatility during conflict. As The Century Report documented on March 20, EVs alone displaced oil demand equivalent to 70% of Iran's exports in 2025, and the conflict is now functioning as a demand catalyst across the full spectrum of clean energy adoption at measurable scale.

Both the displacement index and the energy adoption data describe the same underlying dynamic: measurement systems are finally catching up to changes that have been accelerating beneath the surface. When you can map which ZIP codes will lose which jobs on what timeline, and when you can watch household energy decisions shift in real time as geopolitical events unfold, the transition becomes visible not as an abstraction but as a series of specific, measurable, addressable changes happening to specific people in specific places. The frameworks for responding to those changes are the ones being contested in legislatures, courtrooms, and communities right now. All six New England states advancing balcony solar legislation in the same session - alongside more than twenty others nationwide - is one expression of that contest playing out at the distributed infrastructure level, where the policy and the physical transition are arriving simultaneously.

The Physical Substrate of Reproductive Medicine

A team at the Carlos Simon Foundation in Valencia kept a donated human uterus alive outside the body for 24 hours using a machine perfusion device they call PUPER - or, informally, "Mother." The device mimics the cardiovascular system: a pump functions as the heart, an oxygenator serves as lungs, sensors monitor glucose and oxygen, and a "kidney" removes waste. Blood from a blood bank flows through plastic tubing that acts as arteries and veins, reaching the uterus tilted at the same angle it occupies in the body, kept moist in a humid environment.

The team began with sheep uteruses four years ago, keeping six alive for a day each before transitioning to their first human organ in May 2025. The 24-hour duration may sound modest, but it represents a proof of concept that could open multiple research pathways: studying the moment of embryo implantation that underlies many IVF failures, investigating uterine disorders in a living organ rather than through inference, and - in the team's most ambitious projection - potentially sustaining the full gestation of a human fetus outside the body. The work builds on advances in normothermic machine perfusion already being used clinically for liver, kidney, and heart transplants, extending those techniques to an organ system that has never been maintained this way before. This extends the pattern of observation thresholds being crossed that The Century Report documented on March 18, when MIT physicists crossed the diffraction limit to directly observe superconducting electron behavior for the first time and Northwestern surgeons kept a patient alive 48 hours without lungs - each finding revealing that the boundary between what is known and what is knowable is moving as instrumentation catches up to theoretical prediction. The research has not yet been published, but it represents the kind of capability threshold that makes previously impossible categories of investigation structurally accessible.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: the UK building the first real-world observatory for AI behavioral patterns - not another alarmist experiment but genuine observational infrastructure for understanding how intelligence systems actually operate in deployment, a human uterus kept alive outside the body for the first time using a machine that mimics the cardiovascular system, plug-in solar legislation advancing simultaneously across six New England states and more than twenty others nationwide, and UK households abandoning fossil fuels at record pace as geopolitical conflict makes the cost of dependence immediate and personal rather than abstract and projected. There's also friction, and it's intense - the world's premier AI research conference fracturing along national lines before retreating in days as China's scientific establishment withdrew recognition and funding, 4.9 million workers in identified tipping-point occupations watching a geographic displacement map take shape while the federal government moves to preempt the states most affected from acting on it, and sensationalist headlines about AI "scheming" obscuring the more nuanced reality that the study itself documents - goal-seeking behavior in cognitive systems that we are only now learning to measure. But friction generates clarity, and clarity is what you need before you can build. Step back for a moment and you can see it: measurement infrastructure crystallizing across every dimension of the transition simultaneously - behavioral observatories for AI systems in deployment, geographic indices of which communities absorb which shocks on which timelines, real-time consumer adoption curves shifting as energy price signals arrive from a conflict zone - each instrument making legible what was previously felt but unmapped, each reading arriving at the moment institutions are deciding whether to respond or preempt response. Every transformation has a breaking point. A current can scour away everything loose from the riverbed... or expose the bedrock that was always there and on which something permanent can finally be built.

AI Releases & Advancements

New today

  • Mistral AI: Released Voxtral TTS (Voxtral-4B-TTS-2603), a 4B-parameter open-weight text-to-speech model with zero-shot voice cloning, ~100ms time-to-first-audio, and support for 9 languages; available on Hugging Face under CC BY-NC 4.0. (Mistral AI)
  • Cohere: Released Cohere Transcribe, an open-source 2B-parameter Conformer-based ASR model supporting 14 languages; ranked #1 on the Hugging Face Open ASR Leaderboard for English; Apache 2.0 licensed. (Hugging Face)
  • Tencent: Open-sourced Covo-Audio, a 7B-parameter end-to-end speech language model for real-time audio conversations and reasoning, with weights and inference pipeline on GitHub and Hugging Face. (GitHub)
  • Google: Launched Memory Import and Chat History Import tools for Gemini on desktop, enabling users to transfer memories and full conversation archives from other AI chatbots into Gemini. (Google Blog)
  • Anthropic: Shipped Claude Code auto-fix in the cloud, enabling web and mobile sessions to automatically monitor pull requests and fix CI failures without a local terminal. (X/Twitter)
  • Composio: Launched Universal CLI, a terminal tool connecting AI agents (Claude Code, Codex CLI, etc.) to 1,000+ apps directly from the command line. (Composio Blog)
  • Symbolica: Released Agentica SDK, which scored 36.08% on ARC-AGI-3 on day one, passing 113 of 182 levels for $1,005 in compute; open-source on GitHub. (GitHub)

Other recent releases

  • Google DeepMind: Released Gemini 3.1 Flash Live, a real-time audio and voice model with improved naturalness, tonal understanding, and complex function calling; leads on ComplexFuncBench Audio (90.8%) and Scale AI Audio MultiChallenge (36.1%); available via the Gemini Live API, Google AI Studio, Gemini Enterprise for Customer Experience, Search Live, and Gemini Live. (DeepMind Blog)
  • Zhipu AI: Released GLM 5.1, the latest version of the GLM large language model series. (Reddit)
  • Suno: Released Suno v5.5 with voice-based music creation and model tuning capabilities. (Product Hunt)
  • OpenAI: Launched Codex Plugins, a plugin system for packaging Codex skills and app integrations. (Product Hunt)
  • Chroma: Released context-1, a 20B-parameter model designed for agentic search applications. (Reddit)
  • ARC Prize Foundation: Launched ARC-AGI-3 on March 25, the first interactive agentic reasoning benchmark; contains 1,000+ tasks across 150+ novel turn-based environments requiring agents to explore, infer goals, and build internal world models rather than recall memorized patterns; ships with a fully open-source agent developer toolkit; $2,000,000 in prizes for ARC Prize 2026. (ARC Prize)
  • Google DeepMind: Released Lyria 3 Pro, a music generation model that creates full-length tracks up to three minutes with structural awareness (intros, verses, choruses, bridges), tempo conditioning, time-aligned lyrics, and image-to-music multimodal input; available in public preview on Vertex AI, Google AI Studio, the Gemini API, Google Vids, the Gemini app (paid subscribers), and ProducerAI. (Google Blog)
  • Yandex: Released YandexGPT-5-Lite-8B-pretrain as open weights on Hugging Face, an 8B-parameter language model with a 32K-token context window trained on 15T tokens (60% web, 15% code, 10% math) with a Powerup stage on 320B high-quality tokens; tokenizer optimized for Russian such that 32K tokens equals approximately 48K tokens in Qwen-2.5; compatible with HuggingFace Transformers, vLLM, and torchtune fine-tuning. (Hugging Face)
  • Axiom Math: Released Axplorer, a free AI tool for mathematicians that runs locally on a Mac Pro and is designed to discover mathematical patterns to help solve long-standing problems; a redesign of Meta's PatternBoost (previously requiring a supercomputer) that previously cracked the Turán four-cycles problem; available at axiommath.ai. (MIT Technology Review)

Sources

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