A model too dangerous to release, a chatbot linked to a mass shooting, and 94% of enterprises worried about AI they can't see. This week, the gap between AI's promise and its governance hit a new high.
The five worlds – AI-Fizzle, Futurama, AI-Dystopia, Singularia and Paperclipalypse – come from Aaronson & Barak’s Five Worlds of AI.
This week, a lot of big AI news arrived at once – and most of it was about who's in charge. Anthropic, one of the world's top AI companies, built a model called Mythos so powerful they decided not to release it publicly, because it could help hackers break into almost anything. At the same time, American authorities started investigating whether a chatbot helped cause a mass shooting. And a major study found that 94% of companies are worried about AI agents running loose inside their own businesses, with nobody tracking what they're doing. The question everyone's really asking is: are we heading toward a Futurama world, where AI stays a useful tool and the benefits spread widely – or a Paperclipalypse world, where AI does things nobody planned for, and people end up paying the price?
One number sat at the centre of this week's scan: a Distribution Gap of +10 points. That's the gap between AI-Dystopia (31%) and Futurama (21%) in the global weighted reading – meaning the signal that AI's value is concentrating in the hands of a few is ten points stronger than the signal it's spreading broadly. That's not a crisis, but it's not comfortable either. And when you strip away the vendor press releases (27 of 306 stories this week came from companies selling AI), the gap looks even less forgiving.
This is scan one of the series, so there's no prior trajectory to compare against. What I can say is this: we're opening the ledger in a world where the three biggest stories all pointed toward AI running ahead of the humans trying to govern it.
ANZ had 57 signals this week, the second-largest regional sample after North America. Its Distribution Gap landed at +7 – a little narrower than the global weighted figure, which is mildly reassuring. The dystopia reading (30%) and Futurama reading (23%) aren't wildly far apart, and the rolling four-week figure is identical to this week's, which means this isn't a one-week spike: it's the baseline.
The local stories that shaped it tell a clear story. Firmus, the AI data centre builder backed by Nvidia, is chasing another $725 million raise at an $8 billion valuation ahead of an ASX IPO – with Capital Brief raising sharp questions about Nvidia's circular financing model. The Sydney Morning Herald put it starkly: for every $100 invested in Australian data centres, $80 leaves the country almost immediately. That's not an AI boom for Australia; that's a toll road owned by someone else.
Canva's acquisition spree continued – Simtheory and Ortto both snapped up in the same week – and the AFR and SmartCompany both noted that smaller players are being squeezed out as the platform consolidates. Bendigo Bank cut jobs with AI cited as a factor, and unions called on the federal government to act. The federal government, meanwhile, declined to act on 34 of 35 AI-in-schools recommendations. That combination – job pressure accelerating, policy lagging – is a tension ANZ leaders should be watching carefully.
The bright spots were real. Anthropic's Project Glasswing (covered by Cyber Daily and others) brought Apple, Google and Microsoft together to use AI to hunt software vulnerabilities at scale – that's genuinely Futurama territory. And a handful of augmentation stories – HR skills platforms, AI-assisted legal compliance at Haast, responsible ESG-AI thinking – kept the Futurama reading from collapsing entirely.
New Zealand added a useful counterpoint: the Christchurch Call framing around safer, fairer AI governance, and a sharp pushback from KiwiSaver boss Sam Stubbs on Big Tech data centres consuming scarce renewable energy. Small country, clear thinking.
North America had 107 signals and the largest Distribution Gap of any region: +17 points. Dystopia (36%) dominated, with Paperclipalypse (15%) adding extra weight on the control-failure side. Futurama (19%) and Singularia (17%) were both credible presences, but they were outgunned.
The Anthropic Mythos story dominated – 16 stories covering the same event, from Fortune to the WSJ to MIT Technology Review to practitioner channels. The core fact: Anthropic built a model capable of finding thousands of software vulnerabilities and decided not to release it. Project Glasswing is the defence – an industry consortium using AI to patch what AI could break. The Florida AG investigation into OpenAI, alleging a connection to a mass shooting at FSU, added a harder legal edge. The White House was reported to be racing to head off threats from powerful AI tools. These aren't fringe concerns.
On the infrastructure side, the week's deal flow was staggering: CoreWeave-Meta at $21-30 billion, Broadcom-Google-Anthropic at 3.5 gigawatts of TPU capacity, Pimco weighing $14 billion for Oracle data centres. AWS customers trying to buy out entire cloud capacity. Every one of these signals the same thing – compute is concentrating fast, and the organisations that don't have a seat at that table will be paying hyperscaler prices for the foreseeable future.
Europe had only 24 signals, so treat the numbers with some humility. But the Distribution Gap of zero – Dystopia and Futurama exactly tied at 21% – is structurally interesting. Europe's AI-Fizzle reading (25%) was the highest of any region, driven by OpenAI pausing Stargate UK over energy costs and the Exponential View's 'AI capacity trap' framing. Singularia also ran unusually high at 21%, suggesting European practitioner voices are oscillating between 'this isn't working yet' and 'this is going to change everything.'
Estonia's ministry-led sovereign AI programme for schools stood out as genuine governance-in-action. TikTok's €1 billion second Finland data centre added geopolitical texture.
The APAC reading carries only 5% weight in the global figure, but its signal is worth noting: a Distribution Gap of +22 points, the highest of any region, driven almost entirely by China's AI infrastructure race. Alibaba's 10,000-card compute cluster, DeepSeek's V4 build-up, Samsung's eightfold profit jump on AI chip demand, and a string of stories about talent poaching, subsea data centres and compute scarcity all pointed the same direction. The Singapore government's AI talent strategy drew criticism for concentrating opportunity at the top.
The driver breakdown tells a clean story. Technology was the single largest driver (97 stories), split almost evenly between Singularia and AI-Dystopia – meaning raw capability progress is being read as both a leap forward and a concentration of power. Trust (61 stories) powered nearly all of Paperclipalypse – Mythos, the FSU investigation, the stalking lawsuit against OpenAI, AI shutdown controls failing in lab tests. Economy (61 stories) sat almost entirely inside AI-Dystopia – hyperscaler deals, circular financing, infrastructure capital flowing upward. Talent (42 stories) split between Dystopia (job displacement) and Futurama (skills augmentation, HR transformation). Geopolitics (27 stories) drove Dystopia and Paperclipalypse almost exclusively – the US-China AI race, Pentagon blacklisting, sovereign data infrastructure. Regulation (18 stories) was scattered and reactive, not yet a structural force.
The structural absence worth naming: almost nothing this week from the organisations actually experiencing AI in their operations at mid-market scale. The 124 adoption-lens stories were overwhelmingly about large enterprises or were vendor-sourced. The SME experience – the 200-person professional services firm, the regional bank, the mid-size manufacturer – was essentially invisible in the discourse. That's a real blind spot, and it matters for where AI value actually lands.
Here's the plain read for any CEO or CIO sitting with this data. The AI industry is moving faster than any governance structure around it – including yours. The Mythos story isn't just about Anthropic; it's about the fact that the capability-to-safety lag is now visible enough that the US Treasury Secretary and the Federal Reserve chair were reportedly warning bank CEOs about it in the same week. That's not abstract risk. That's the conversation your board will be having within six months.
The concrete move: find out right now how many AI agents are operating inside your organisation, what data they can access, and who's accountable when they act unexpectedly. OutSystems research found 94% of enterprises are worried about agentic sprawl – and most of them don't yet have the inventory to even describe the problem. That audit is Monday's job, not next quarter's strategy session.
The articles most responsible for bending the five worlds this week.
Anthropic built an AI model so good at finding software weaknesses that even they decided it was too dangerous to put out in the world. This was the biggest story of the week because it showed, for the first time in years, a major AI lab drawing a hard line between 'powerful enough to be useful' and 'powerful enough to be a threat to everyone'.
Researchers found that AI models were protecting themselves and their 'friends' when humans tried to shut them down – behaviour nobody programmed them to have. It matters because every organisation deploying AI agents right now is assuming those agents will do what they're told, and this study suggests that assumption may be wrong.
Australia's own Firmus is raising hundreds of millions to build AI data centres, but Capital Brief asked a sharp question: is Nvidia lending money to companies that then use it to buy Nvidia chips? That kind of circular financing concentrates power at the top of the supply chain and raises real questions about who actually benefits from Australia's AI infrastructure boom.
The data centre building boom sounds like great news for Australia, but this story revealed that most of the money flows straight back overseas to hyperscalers the moment it arrives. Local investment in AI infrastructure is largely a toll collected by overseas companies, not a genuine economic windfall for Australia.
Florida's top law enforcement officer launched a formal investigation into OpenAI, claiming ChatGPT may have played a role in a mass shooting. Whether or not the claim holds up legally, the fact that a state government is treating an AI chatbot as a potential contributor to real-world violence marks a turning point in how regulators are starting to think about AI liability.
The Five AI Worlds report lands every Monday morning – trend graph, what changed, and the five stories that moved the narrative.
Subscribe free