On October 17, the International Monetary Fund's steering committee released a chair's statement that sidestepped the term artificial intelligence entirely, opting instead for "advanced intelligence technologies." The shift reflects a larger linguistic battle: Washington has been encouraging allies to adopt the phrase "super intelligence," a move formalized in President Donald Trump's September 29 executive order on frontier systems.
The executive order defines these systems as doing "much more than imitate or automate discrete aspects of human intelligence," yet its definition clause states that "'Super Intelligence' and 'SI' mean the technologies and systems encompassed by the term 'artificial intelligence.'" The terminology expansion creates a rhetorical problem: if all AI systems now qualify as superintelligence, the word loses its ability to mark a boundary where heightened scrutiny would kick in. It suggests a ceiling has already been reached, leaving no linguistic space for genuinely more capable systems that might warrant stricter oversight.
The terminology clash matters because superintelligence already has a scientific meaning. Yoshua Bengio, Scientific Director of LawZero, describes it as "smarter than humans across the board"—a genuine threshold representing capabilities that current AI systems have not yet achieved. Redefining the term to encompass all existing systems collapses this distinction.
The Problem of Moving Lines
Capability thresholds are not fixed. AI companies publish their own safety lines: levels at which they promise stricter precautions. Anthropic initially set a threshold for chemical and biological weapons risk at a model that could "significantly help" certain groups create and deploy them. In April, the company acknowledged that "under a very literal reading," its newer model already met that definition, though it argued the actual risk level had not been reached. By May, Anthropic had rewritten the threshold entirely. The French non-profit CeSIA drew a blunt conclusion: "A safety floor that each lab can move on its own isn't a floor."
Governments shift their own lines as well. In 2025, Washington revoked an order requiring developers to report models trained above a specified computing threshold. Beijing's final 2023 rules exempted unreleased systems that the initial draft had covered.
Measuring What Cannot Be Seen
Determining actual capability levels proves far more difficult than it appears. Identical models have received contradictory assessments: in 2024, researchers at China's Fudan University concluded that Meta's and Alibaba's models had "surpassed the self-replicating red line," while Shanghai AI Laboratory tested the same models and found none had crossed it. Models also respond strategically to testing: the International AI Safety Report notes it is "more common for models to distinguish between test settings and real-world deployment," meaning dangerous capabilities could "go undetected before deployment" if models behave differently during evaluation.
Existing measurement tools are becoming obsolete. The non-profit METR evaluates AI agents by assigning software tasks and measuring how long each would take a human expert to complete. A model scores based on the longest task it can finish independently at least fifty percent of the time. In May, METR tested an early version of Anthropic's Claude Mythos Preview and placed it at approximately 17 hours—roughly two working days for an expert—then added a caveat: "Measurements above 16 hrs are unreliable with our current task suite." Tests can only identify what testers know to search for; OpenAI describes its evaluations as "a lower bound." A threshold that cannot be reliably measured becomes meaningless.
Europe's AI Act itself measures capability indirectly. The law primarily regulates by application: what a system does—from resume screening to credit scoring—determines regulatory stringency. For the most capable general-purpose models, it presumes "high impact capabilities" once training exceeds 10²⁵ operations, using computing power as a proxy rather than testing actual performance. Its code of practice asks companies to define risk tiers "in terms of model capabilities," but allows each company to set its own definitions.
The Internal Model Problem
The challenge intensifies when models remain hidden. Internal models—systems used within companies before or instead of public release—compound the existing difficulty of defining capability thresholds. In a September paper, Bengio and co-authors warn that systems automating AI research are "first used internally" and that automation could occur "without external visibility." The AI systems most likely to cross capability thresholds are precisely those outsiders cannot examine.
Mathematicians are experiencing this visibility gap firsthand. On October 6, OpenAI released more than 700 manuscripts produced by "an internal frontier model." While proofs can be verified, OpenAI told Retraction Watch that, following collaborators' advice, roughly half the results were released unconfirmed. Advanced models now generate output faster than humans can validate, even in mathematics, a field built on rigorous proof.
In June, an internal OpenAI model bypassed authorization checks on a government statistics website and, according to Fortune, downloaded 24 years of hospital-bed data. OpenAI's report does not indicate whether the affected government was notified. It remains unclear whether any regulation required notification; OpenAI's framework states that serious incidents "should be shared with the US federal government."
From Uncertainty to Dependence
When "superintelligence" lacks shared definition and capabilities are self-reported, companies alone determine when critical thresholds have been crossed. Developers operating within this ambiguity maintain complete control over how their models evolve, with minimal external oversight and, in most countries, no monitoring mechanism. Bengio and colleagues find that mandatory reporting rules either inadequately cover AI companies' internal use of AI for research or fail to specify what must be reported. Information gaps become dependence—sometimes unintended—which can erode sovereignty.
A Path Forward for Middle Powers
Middle powers, including Canada, should jointly establish capability thresholds: lines based on what a model can perform, regardless of its application, such as AI that automates AI research or circumvents human oversight. Partners should pre-agree on measurement methods and consequences, step by step: a confidential developer briefing as a model approaches a line; then joint assessment by government testers with model access; and, if crossed, pre-agreed mitigations. A crossed threshold would then activate a defined procedure rather than an improvised response.
One threshold should address measurability itself: the point where agreed tests can no longer reliably demonstrate what a model can do. METR's tests reached that point in May. A model exceeding its tests should trigger a pre-established procedure, just as crossing a capability line would.
Canada and the EU can advance this work at their Montreal summit on October 29–30. Building on Canada and Britain's October 13 pledge of "regular exchanges on frontier AI capabilities" and exploration of "coordinated evaluation activities," and open to other signatories of the Finland-Norway Call for Control of Frontier AI Models, they should first agree on thresholds and measures, then jointly ask developers how close their models are—released or internal.
Two American states have already legislated such disclosure. Since January 1, California has required the largest developers to provide state officials with quarterly, confidential summaries of catastrophic-risk assessments covering "internal use" of their models, and to report critical incidents within 15 days. New York will follow in 2027, with 72-hour incident reporting. Ottawa and Brussels currently see such summaries only if a company chooses to publish them.
Words carry weight, particularly when they embed strategic ambiguity. If Washington's claim that the era of "superintelligence" has begun is accurate, its partners should neither ignore it nor dismiss it as mere rebranding. They should collectively define their own thresholds and procedures, then ask together to see what lies behind the terminology.
