There is something oddly persuasive about a chart that rises sharply to the right.
Elon Musk recently shared one alongside the declaration that artificial intelligence is now a “supersonic tsunami.” The horizontal axis runs from July 2023 to July 2026. For most of that period, the blue bars remain close to the bottom. An arrow points to November 2024 with the words “it’s so over.” Then, as the chart reaches the spring and summer of 2026, the bars climb dramatically.
It resembles the familiar shape of exponential progress. There is only one small problem: the vertical axis does not tell us what is being measured.
There is no unit for intelligence, economic impact, scientific discovery, autonomy or social change. The dates are precise; the meaning of the bars is not. The chart gives an impression of measurement without defining the object being measured. It does not demonstrate the tsunami. It depicts the feeling of one.
This could be dismissed as ordinary social-media theatre. Musk is hardly the first executive to dramatise a product category, and a post on X is not a scientific paper. Yet the image belongs to a larger and increasingly familiar story about artificial intelligence.
In July 2026, OpenAI chief executive Sam Altman declared that we are now “in the singularity.” He described the transition as gradual rather than as a single dramatic rupture, but the phrase still invoked a much stronger idea: a hypothetical point at which machine intelligence surpasses human intelligence, begins improving itself recursively and produces changes that humans can no longer predict or control.
Two months earlier, Demis Hassabis, the Nobel Prize-winning scientist and CEO of Google DeepMind, had offered a more cautious geographical metaphor. Closing the Google I/O keynote, he said that when people look back at the present period, they may conclude that humanity was standing in the “foothills of the singularity.”
Musk gives us a wave. Altman gives us a threshold. Hassabis gives us a mountain.
The scenery changes, but the plot remains remarkably stable. AI is advancing. Its advance is accelerating. A decisive boundary is approaching—or has already been crossed. The future is no longer something societies will negotiate. It is something arriving from elsewhere.
When a Metaphor Makes an Argument
Metaphors are not decorative additions to thought. They influence what kind of thing we believe we are looking at, which questions appear reasonable and which responses seem possible.
A tsunami is powerful, fast and indifferent to human preference. It cannot be debated with, voted against or redesigned through public institutions. Once it becomes visible, the meaningful choices have already narrowed to adaptation, escape or survival.
Calling AI a tsunami therefore smuggles a political conclusion into what appears to be a technical observation. Resistance is futile. Delay is dangerous. Those who detected the wave early deserve authority.
The singularity performs a similar function. It divides history into a before and an after, while suggesting that normal forms of prediction and governance may no longer apply. If we have already entered it, demands for evidence can begin to sound quaint. How could yesterday’s institutions evaluate a phenomenon that has supposedly escaped yesterday’s categories?
The foothills are less alarming, but they still place us on a predetermined route. Once we accept the mountain as our destination, the debate shifts from whether we should make the journey to how quickly we will reach the summit.
These metaphors capture something real. AI systems have improved rapidly in coding, language production, scientific problem-solving and the completion of multi-step tasks. They are changing workflows, organisational structures and expectations of what software can do. It would be a mistake to answer inflated rhetoric with complacency.
But it is equally mistaken to treat rapid improvement in selected capabilities as proof of an autonomous historical force.
An AI model does not arrive like a weather system. It is developed by particular organisations, using particular data, chips, energy, labour and capital. It is released under particular commercial incentives. Its uses are shaped by employers, governments, schools and consumers. Even the pace of development depends on decisions about investment, infrastructure, regulation and access.
The wave has owners.
The Convenient Fiction of a Single Ladder
Both the rising chart and the singularity story depend on another simplification: the idea that intelligence is one quantity moving in one direction.
If intelligence can be placed on a single ladder, the central question becomes easy to formulate. Are machines below human beings, level with us or already above us? Every new benchmark result appears to move them another rung upwards.
Human and machine capabilities do not form such a tidy hierarchy.
Kai Riemer and Sandra Peter make this point in their response to Altman’s claim. Current systems can outperform most people at some highly valued tasks while lacking forms of embodied experience, continuous learning and participation in social life that are central to human intelligence. Today’s large language models do not autonomously rewrite their own parameters while they operate. Their improvement still depends on human-initiated training and engineering processes.
This does not make the systems unimpressive. It makes the comparison more difficult.
A language model can process more text than any individual could read, produce usable code in seconds and identify patterns across enormous collections of information. It can also generate a confident answer without understanding why a claim requires stronger evidence, when uncertainty should stop it from answering or what the consequences of an error might be.
That distinction matters because intelligence is not merely the ability to produce a correct response. In social life, we also care about how someone knows, whether the evidence warrants the claim, whether the speaker recognises uncertainty and whether they can be held responsible when they are wrong.
Harry Collins and his co-authors describe part of this as “classroom sensibility”: the learned expectation that public claims should be supported by appropriate warrants. Their comparison of human knowledge and large language models argues that people acquire this sensibility through socialisation. Current models, by contrast, are designed to generate responses without possessing an equivalent understanding of what knowing, not knowing or being responsible for a claim entails.
The phrase is especially useful because the problem is not confined to machines. It can also describe the way powerful humans speak about them.
What is the warrant for declaring a singularity? What exactly does a rising bar measure? Which capabilities improved, over what period, according to whose benchmark and under which conditions? What would count as disconfirming evidence?
A claim about unprecedented intelligence should not require less scrutiny merely because it is delivered by someone building the technology.
What the Benchmarks Leave Out
Numerical benchmarks are useful. They allow researchers and developers to compare systems under defined conditions. The problem begins when performance on selected tests is translated into a general claim about intelligence.
Collins and his colleagues argue that the financially backed competition between AI companies encourages precisely this interpretation. Models are compared through numerical scores, creating the impression that the main problem of intelligence has already been solved and that machines are now simply competing to see which one is most intelligent.
A model that scores 87 per cent rather than 83 per cent has improved according to a particular test. The number alone does not tell us whether the system understands the task as a person would, whether it can recognise when a problem falls outside its competence or whether it should be trusted when an incorrect answer carries serious consequences.
Benchmarks can measure performance. They do not automatically measure judgement.
This is why the missing vertical axis in Musk’s chart is more than a humorous omission. It is an exaggerated version of a much wider problem. “Progress” is presented as a single, cumulative quantity even though AI development consists of many uneven and sometimes contradictory changes.
A model may become better at writing code while becoming more expensive to operate. It may improve on a scientific benchmark while remaining vulnerable to fabricated sources. It may complete longer tasks while introducing new security risks. It may produce more persuasive language without becoming more reliable.
Which of these developments belongs in the rising bar? How should each one be weighted? Does greater persuasiveness count as progress when confidence can make an incorrect answer more dangerous?
Without answering such questions, the curve tells us less about intelligence than it does about our desire to see intelligence as a curve.
The Prophet Is Also the Vendor
Musk, Altman and Hassabis may sincerely believe what they say. Their claims do not need to be cynical in order to serve strategic purposes.
The leaders of AI companies occupy an unusual double role. They are presented as expert witnesses to a technological transformation while also competing to finance, build and distribute the systems driving it. Their proximity gives them valuable knowledge. It also gives them interests.
Grand forecasts can attract capital, talent, customers and political attention. They can make extraordinary infrastructure spending appear prudent and turn commercial competition into a civilisational mission. Even warnings about existential risk can elevate the companies issuing them. If a technology might transform or destroy the world, its creators become unavoidable participants in every discussion about the future.
The rhetoric can move between utopia and catastrophe without losing its basic function. AI will create abundance, or it may escape our control. It will liberate humanity, or leave societies unable to adapt. The emotional register changes, but the imperative remains the same: this technology stands at the centre of history, and the people building it are closest to the centre.
That is why the recurrent theme is larger than hype. It concerns who gets to narrate technological change.
When corporate forecasts are reported as neutral descriptions of the future, commercial actors gain the power to define both the problem and the acceptable range of responses. The public is invited to choose between enthusiasm and fear, while more ordinary questions receive less attention.
Who benefits from a particular deployment? Who bears the risk of its mistakes? Which work should be automated, and which decisions should remain open to human challenge? What evidence should be required before an AI system enters a school, hospital, workplace or public institution?
Those questions are less cinematic than a singularity. They are also where the future is actually being made.
Taking AI Seriously Without Surrendering to It
There are two easy positions in the AI debate. One is to accept every dramatic forecast as evidence that history has accelerated beyond politics. The other is to treat every forecast as marketing and conclude that little of substance is changing.
Neither position is adequate.
AI does not need to become a superintelligence to alter the distribution of power. A system can remain fallible and still influence hiring, education, welfare, policing, media and war. It can fail to understand in a human sense and still be treated as an institutional authority. It can possess no independent goals and still cause harm when organisations give it autonomy, scale its errors or hide their own decisions behind its outputs.
The most immediate danger may therefore be neither an awakened machine nor an entirely fictional bubble. It may be a society that attributes too much agency to machines and too little to the people and institutions deploying them.
Natural-disaster language makes that mistake easier.
If AI is a tsunami, job losses become consequences of the wave rather than of managerial decisions about automation. If the singularity is already here, weak oversight becomes evidence that governments cannot keep pace rather than the result of political choices to leave development largely in corporate hands. If progress is a line on a chart, questions about what is being optimised—and for whom—disappear behind the direction of the curve.
We need language that restores those choices to view.
Artificial intelligence is not the weather. It is an extraordinary and rapidly evolving collection of human-made systems. Its development may produce changes on the scale its advocates anticipate, although the timing, form and distribution of those changes remain uncertain. Precisely because the stakes are high, its public meaning should not be set by metaphors alone.
The blue bars may continue to rise. Before reorganising society around them, however, we should be told what is on the vertical axis.
Phil Haunhorst, “Elon Musk Says AI Is Now a ‘Supersonic Tsunami’—His Chart Shows Where It Broke Out”, BeInCrypto/Yahoo Tech, 3 August 2026.
Sarah Shamim, “Sam Altman says AI has entered ‘singularity’: Should we be worried?”, Al Jazeera, 27 July 2026.
Kenrick Cai, “Google’s Demis Hassabis goes on the offensive”, Reuters, 20 May 2026. See also “AI x Society: I/O 2026 Keynote”
Kai Riemer and Sandra Peter, “Sam Altman says we’re ‘in the singularity’ with AI. Here’s why he’s wrong”, The Conversation, 31 July 2026.
Harry Collins, Simon Thorne, Paul Newbury and Robert Evans, “More problems of large language models in comparison with human knowledge”