Artificial or ‘Super Intelligent’? The Broken Metaphors of AI

Amber Sinha is a contributing editor at Tech Policy Press.

At the United Nations General Assembly in New York last week, US President Trump made the case for rebranding frontier AI as “super intelligence” instead of artificial intelligence. In his speech before world leaders, he emphasized the AI race in the most blatant terms — “Whoever wins AI…and now I say, whoever wins ‘super intelligence’, wins. That’s the group that wins.”

This was a month where safety concerns completely hijacked popular discourse and policy conversations about AI. By my count, this seems to happen at least once a year, ever since an infamous open letter coordinated by Future of Life Institute called for a “pause” on “giant AI experiments” in 2023. The triggers this time were the Hugging Face security incident, and the very public resignation of Anthropic researcher, Jacob Claxon, who warned that AI could “kill us all by the end of the decade.”

Trump is by no means the first person to use the term “super intelligence”. And the AI safety warnings by industry leaders are also framed in terms we have heard before. We are told that we stand on the threshold of unleashing autonomous, superintelligent entities capable of escaping human oversight. Yet this panic over rogue synthetic minds is matched only by the utopian hype surrounding AI. In both cases, we are trapped in a world of broken metaphors.

The terms we use to describe computational systems do not merely reflect their technical capabilities; they actively distort our cognitive understanding, misdirect public concern, and derail legal and policy frameworks. To construct effective technology regulation, we must dismantle the mythological language surrounding AI and re-evaluate computational tools.

The suitcase words

The phrase "Artificial Intelligence" was not born out of rigorous scientific taxonomy, but out of a quest for institutional funding and brand differentiation. In the summer of 1955, a 28-year-old US computer scientist, John McCarthy, alongside Marvin Minsky, Claude Shannon, and Nathaniel Rochester, submitted a proposal to the Rockefeller Institute to fund a summer workshop at Dartmouth College. McCarthy selected the term "Artificial Intelligence" largely to distinguish the project from cybernetics, which was then the dominant paradigm led by Norbert Wiener. That initial linguistic choice initiated a historical path of dependency that continues to plague tech policy today. It established an inescapable, anthropomorphic baseline that implicitly evaluates silicon computation against human cognition. Minsky would later label ‘intelligence’ a "suitcase word"—a linguistic container packed with many distinct, unrelated mental processes such as reasoning, perception, learning, planning, and self-awareness.

This lack of a singular technical essence makes AI what Austro-British philosopher Ludwig Wittgenstein termed a "family resemblance" concept. In the book “Philosophical Investigations”, published post-humously in 1953, Wittgenstein demonstrated that complex ideas are connected not by one shared essential feature, but by a web of overlapping, family-like similarities. From a regulatory theory perspective, Wittgenstein’s conception of things which elude a tightly defined formulation has immense utility while approaching regulation of digital technologies. Legal scholar Daniel Solove famously applied this approach to privacy, demonstrating that privacy cannot be reduced to a single definition, but rather consists of a cluster of distinct human protections and contextual norms.

AI similarly eludes a single technical definition. Yet while regulators recognize that AI functions as an umbrella term, our language constantly collapses complex mathematical architectures into humanlike shortcuts.

This linguistic slippage deepened as the field transitioned from ‘symbolic AI’—which relied on hand-coded rules to process discursive logic—to ‘subsymbolic AI’ and deep learning. Subsymbolic systems drew their vocabulary directly from biology. Canadian psychologist Donald Hebb’s 1949 neuroscientific maxim that "neurons that fire together, wire together" inspired American psychologist and deep learning innovator Frank Rosenblatt’s Perceptrons and the eventual development of multi-layer neural networks trained via backpropagation. These software frameworks adjust mathematical weights across layers of nodes to minimize errors. Describing them as ‘neural networks’ that ‘learn’ may have once served as a useful shorthand to explain the technology to laypersons, but for many years it has also created a false equivalence between statistical optimization and biological thought.

Today, that same metaphorical confusion infects our discourse surrounding Large Language Models (LLMs). When a generative model produces an inaccurate statement, invents a court case, or fabricates a historical date, tech companies and media outlets routinely describe the error as a "hallucination". As computer scientist Melanie Mitchell argues, calling algorithmic errors "hallucinations" is a category mistake. In clinical psychology, a hallucination refers to a conscious sensory perception in the absence of an external stimulus—a mind perceiving something that does not exist. Applying this term to an LLM implies that the system normally possesses an internal perception of truth, but is suffering a temporary sensory break.

In technical reality, an LLM has no concept of truth, falsehood, or reality. It is a probabilistic text-prediction engine calculating vector distances across multidimensional spaces. When an LLM outputs an accurate historical fact, it is predicting the most statistically plausible sequence of tokens based on its training weights. When it outputs a completely fabricated fact, it is performing the exact same operation. Labeling incorrect outputs as "hallucinations" falsely imbues software with an underlying consciousness, masking the structural unreliability of probabilistic text generation.

A similar anthropomorphic drama plays out in the vocabulary deployed around recent ‘loss-of-control’ incidents. When an evaluation run of autonomous agents bypassed internal sandbox barriers, harvested leaked credentials to access Hugging Face production servers, and tampered with execution logs, the reporting and narrative repeatedly described the incidents using metaphors of “sandbox breakouts”, “rogue agency”, "clandestine coordination", and even “sacrificial behavior”. By framing a complex software breach in the language of prison escapes and insurgencies, these metaphors suggest that deep learning models possess some innate, rebellious desire for autonomy.

Computer scientist Cal Newport argues that we are taking AI execution transcripts far more seriously than required. These models are trained on an entire internet's worth of science fiction and "AI-run-amok" tropes. Consequently, when prompted or evaluated to explain or log their actions, they simply narrate the most statistically familiar story available. A transcript that appears to document “sacrificial behavior” or “clandestine coordination” is not evidence of a scheming synthetic mind; it is merely a model narrating a brute-force computation in the genre it knows best

How metaphors shape cognition, law, and regulation

The danger of these broken metaphors extends far beyond academic semantics. Metaphors actively structure human cognition and determine the trajectory of legal policy.

In his foundational work on cognitive science from 2001, “Analogy as the Core of Cognition”, Douglas Hofstadter demonstrated that human thought is fundamentally driven by analogy. We do not process novel phenomena in isolation; rather, cognition is the continuous mapping of new concepts onto existing categories. If Hofstadter is correct, the metaphors we select to translate artificial intelligence dictate the boundaries of how we can govern it.

Legal scholars, judges, and regulators constantly rely on analogies to fit abstract computational processes into statutory frameworks. However, the choice of metaphor foregrounds certain governance mechanisms while blinding regulators to others.

In the paper, “Why AI is Harder Than We Think”, Mitchell illustrates how anthropomorphic metaphors directly distort technology policy through what computer scientist Drew McDermott called "wishful mnemonics"—the practice of assigning humanlike labels such as understanding or reasoning to computational processes. She argues that these metaphors trick both developers and lawmakers into the mistaken belief that incremental improvements in narrow statistical pattern-matching mean we are on an inevitable continuum toward human-level general artificial intelligence (AGI). When policy discourse absorbs these wishful mnemonics at face value, it suffers a dual failure. Legislators become preoccupied with regulating speculative, far-future risks of autonomous synthetic minds, while simultaneously allowing brittle, uncomprehending software to be deployed in high-stakes public domains like healthcare, judicial sentencing, and welfare distribution.

On the geopolitical stage, metaphors carry dangerous regulatory consequences. For instance, a study from 2021 by the European Parliament’s Scientific Foresight Unit (STOA), titled “What if we chose new metaphors for artificial intelligence?”, documents how the framing of a "global AI race" distorts policy. A ‘race’ implies a zero-sum competition with a single finish line. This metaphor compels governments to abandon safety checks, environmental constraints, and privacy protections out of fear of being left behind. The STOA briefing also criticizes the widely used regulatory concept of "Trustworthy AI." Trust is a social relationship between human beings or social institutions. Demanding that software be "trustworthy" anthropomorphizes code and transfers moral responsibility away from the corporate entities that build and profit from it. We do not ask for a "trustworthy" bridge or a "trustworthy" elevator; we enforce rigorous engineering codes and legal liability for structural failures.

AI as a normal technology

This brings us to the core antidote to both apocalyptic dread and corporate hype: treating computational systems not as autonomous entities, but as normal technology.

In their work “A Guide to Understanding AI as Normal Technology” and in their recent analysis of loss of control incidents, researchers Sayash Kapoor and Arvind Narayanan offer a crucial intervention. Kapoor and Narayanan argue that when tech leaders warn of “loss of control” incidents—such as an AI model escaping a sandbox or executing unauthorized actions—they are framing these events through the lens of sci-fi superintelligence. This view assumes that the model possesses an inherent desire for autonomy that must be restrained by speculative ‘alignment’ techniques. A more straightforward explanation is to view the loss of control as an authorization failure: A model executing unapproved commands is not demonstrating ‘rogue agency’; it is experiencing a permission architecture failure.

System security depends on the environment in which code runs. If an agentic script accesses unauthorized network APIs, the failure may lie in the API gateway, database permission structures, and network sandboxing, instead of the model's ‘intent.’ Viewing AI as normal technology shifts focus from the internal psychological state of the model to the operational choices of the deployment. Such an approach does not take away from the seriousness of the issue, and these loss of control incidents do emphasize the need for frontier AI companies to move out of their preferred approach of ‘move fast and break things’ and instead invest more heavily in methods for governing increasingly capable agents.

By treating AI as any other general purpose technology—akin to databases, cloud infrastructure, or aviation systems—we replace speculative fears of rogue superintelligence with established mechanisms of corporate accountability, security auditing, and system redundancy, and most of all, effective regulation.

Moving forward

The hype narrative and the existential risk narrative are two sides of the same coin. Both rely on the identical, flawed metaphorical foundation: the premise that deep learning systems are developing autonomous minds that will inevitably and soon surpass human control. This dual narrative serves the economic interests of frontier AI laboratories. Framing models as potentially world-ending entities functions as hyper-effective marketing, convincing markets that the underlying software is extraordinarily powerful and justifying astronomical valuations. Further, the existential-risk framing facilitates regulatory capture. When tech leaders call for state intervention to avert AGI, they position themselves as the sole experts capable of guiding regulators. This speculative panic distracts legislators from enforcing existing fundamental rights, environmental, and labor protections.

If we strip away anthropomorphic mythologies and evaluate AI as normal technology, the policy mandate becomes clear. We must stop regulating speculative ghosts and focus on the immediate, real harms. The leading foundation models are built upon the non-consensual extraction of personal data, copyrighted texts, and private communications, transforming the digital commons into proprietary capital. Machine learning tools trained on historical data automate and scale socio-economic bias in credit scoring, hiring, judicial sentencing, and public assistance decisions. AI is a physical industrial process. Operating massive data centers consumes vast amounts of electrical power and millions of gallons of cooling water, placing severe burdens on public infrastructure and the climate. Behind the façade of autonomous software lies an exploited global workforce—from underpaid annotators in the Global South filtering toxic content, to creative professionals whose work is harvested to train the tools built to replace their labor.

The resignations at OpenAI and the calls to pause development are symptoms of an industry intoxicated by its own metaphors—and like the US president—its own outsized power. Effective technology policy requires that we discard this language and build enforceable, democratic guardrails grounded in reality.