I have recently written a blog post about various movement metaphors used in AI discourse, especially those of ‘race’ and ‘journey’. A related metaphor, that of ‘pace’ or ‘pacing’, focusing on the rate and regulation of movement, has been made prominent in a September essay by Dario Amodei, CEO of Anthropic, entitled “We must pace the frontier”. In this essay Amodei proposes to slow down the pace of AI development in order to make it safer. This proposal coincided with President Trump’s promotion of the ‘race to win’ framing of AI.
So far, I have not discussed the word ‘frontier’ used in the same context. In this post I want to dig a bit deeper into the pace metaphor but also examine the frontier metaphor with which it is often associated, as in Amodei’s essay.
I’ll first say something general about journey, race, breakthrough and frontier metaphors in science/technology; then home in on Amodei’s use of the relatively novel pace metaphor; and finally link that metaphor to the frontier metaphor. In fact, as, according to the philosopher of metaphor Max Black (1962), many metaphors are models, I’ll show how the old ‘frontier model’ as metaphor is linked to the new ‘frontier model’ as a prominent phrase in AI discourse. In doing this analysis of pace and frontier metaphors, I’ll try to show some tensions in the framing of AI development and what that might imply for AI regulation.
Journeys, races, breakthroughs and frontiers
There are two important conceptual metaphors that structure scientific development, discovery and advancement: SCIENCE/TECHNOLOGY IS RACE and SCIENCE/TECHNOLOGY IS A JOURNEY, with their conceptual components of the traveller, the terrain covered and discovered, the path, steps and speed, indeed pace, milestones, obstacles and barriers, and competition. A traditional example of this sort of metaphor is “Scientists raced to map the human genome”, “Scientists explored uncharted territory” or, indeed “AI development is at a crossroad”.
The prototype of the journey framing of science is perhaps Darwin’s voyage on HMS Beagle in the 1830s, as a slow pursuit aiming to find out new things about how nature works. The prototype of the race framing is perhaps the US/USSR 1950s race to develop the H-bomb, a competition between two nations in fast pursuit of a very specific goal. It was at that time in the history of science and technology that two other terms emerged alongside that of (arms) race, namely that of ‘frontier’ and ‘breakthrough’.
In 1945 Vannevar Bush used the military word ‘frontier’ (related to frontline, border and the American myth of the ‘Wild West’) when talking of ‘science as the endless frontier’. Instead of “frontier” referring to the edge of settled or explored territory, it begins to represent the boundary between the known and the unknown.
In the 1950s the word ‘breakthrough’, until then also used only in the military context, began to be applied to science (OED). Science became a more straightforwardly goal directed race to break through barriers and conquer new frontiers. In 1955 one can find examples of the US arguing that: “Supporting research is the lifeblood of frontier technology, and guided missiles are indeed at the frontier of our technology.” (OED) Frontier became an adjective. In 1966 the television series Star Trek popularised the phrase “space: the final frontier.”
In the 1990s the sequencing of the human genome became a frontier in science, and in a year 2000 announcement of a breakthrough then President Clinton explicitly linked it to Meriwether Lewis’s mapping of the American West. In 2018 there was talk of an “electronic frontier”, and around 2023 the phrase ‘frontier models’ (also ‘frontier AI’, even ‘frontier intelligence‘) started to be increasingly applied to large language models or AI systems that sit at the cutting or leading edge of (possibly dangerous) capabilities, pushing the boundaries, expanding the frontiers of knowledge. Through shortening or ellipsis, ‘frontier’ can now stand for ‘frontier model’ or ‘frontier technology’ as in this rather startling question asked by Sky News: “Could rapidly advancing frontiers kill us?” which means: “Could rapidly advancing AI technologies kill us?” – for an answer see this article in Nature for example.
Critiques of such talk are also emerging. In February 2024 Bruce Schneier and Nathan Sanders wrote a post entitled “How the ‘frontier’ became the slogan of uncontrollable AI” in which they argued that AI companies use ‘frontier’ as a rhetorical device to frame AI as wild territory exempt from traditional oversight and regulation, linking this framing back to Frederick Jackson Turner’s 1893 American frontier thesis and historical expansionist narratives that justify exploitation and evade public accountability. Some critics, like Gina Helfrich, see ‘frontier AI’ as harmful hype.
With words like ‘uncontrollable’, ‘accountability’ and ‘hype’, we have reached the present day, where frontier labs now argue that their creatures have escaped their control. As Emma Beddington said in 2023 in an article for The Guardian: “We’re pushing the frontiers of artificial intelligence, eyeing our creations, Victor Frankenstein-style, with foreboding”.
And with that we come to Amodei’s article “We must pace the frontier”, where a creator eyes its creation and asks it to slow down. As the BBC reported: “Dario Amodei recently called for AI development to slow down and proposed industry-wide deceleration and regulation, as well as independent monitoring of AI model development.” Let’s see how these ambitions were framed.
Pacing the frontier
Like many other tech insiders, Amodei had become alarmed by the speed with which AI systems, models, indeed ‘agents’ were advancing. He feared that ‘frontier models’ were “outrunning industry”, as they were not only getting better at what they were programmed to do but might also be getting better at improving ‘themselves’ – what is called ‘recursive self-improvement’ (where AI systems write their own code). By focusing on the opportunities promised by chasing the ‘endless frontier’, AI developers had forgotten about limits and boundaries, indeed frontiers, that should have ‘contained’ their models and kept them in line and aligned.
Amodei explicitly defined ‘pacing the frontier’ as “building AI at a balanced rate that aims to ensure its safety while achieving its benefits and grappling with important geopolitical dilemmas” – this last part of the definition hints at some problems with the pacing metaphor that I’ll discuss later. China is never far away and racing rather than pacing lurks always in the background (‘pacing the race’). The key point is that: “We must slow the pace at which we improve the capabilities of AI models.” This can be done through “verifiable pacing”, that is, embedding independent evaluators in model development – a rather challenging task. He hopes that by working at a “measured pace” or a “safe pace” one can “achieve much greater operational excellence”.
Pacing while racing
The metaphor of pacing is itself embedded within a wider network of journey metaphors – and here is where some tensions emerge between pacing and racing and where pacing becomes a ‘hybrid metaphor’, combining and exploiting aspects of both journey and race metaphorical framings. I’ll focus on three journey metaphor here: ‘checkpoints’, ‘speed limits’ and finding a ‘middle way’.
‘Checkpoints’ is a journey/travel term. In Amodei’s usage (“I am most enthusiastic about pacing based on what a given frontier AI system can do, and how safe we observe it to be”), this becomes more of a competitive compliance gate: if models have capability X, they need certifications Y and Z. Here the journey metaphor merges with the race one, something I call a metaphorical hybrid, or hybrid metaphor. And the focus is on what frontier AI system ‘do’.
When he talks about a ‘speed limit’ he focuses on pacing as regulating speed of movement, setting limits on the rate of recursive self-improvement. He uses a road-journey metaphor, but it is deployed to manage a geopolitical race with China, as the clause “gives up relatively little strategic advantage” indicates.
And finally, when he evokes a ‘middle way’, he uses a path metaphor “We have sought a middle way: to show that it’s possible to build carefully and succeed commercially, and to make safety something on which AI companies compete. In other words, to create a race to the top.”
In all three cases, we have overt journey metaphors but covert race and competition metaphors. ‘Pacing’ is what I call a hybrid metaphor, gesturing at ‘journey’ and slowing down while staying inside ‘race’ – pacing the race. SCIENCE IS A JOURNEY vocabulary is recruited to do SCIENCE IS A RACE work. This hybridity provides scope for having your cake and eating it, for still sending AI models on a journey even a fast race but putting a few obstacles in their way.
Frontier models, old and new
As Amodei writes, feeding the same hybrid framing: “A coordinated pacing strategy would give frontier AI developers the time to do this vital work without sacrificing commercial advantage or the United States’ lead in AI” and “extra time from a paced frontier would help our researchers improve our understanding of [alignment issues].” In the phrase ‘a paced frontier’, ‘frontier’ performs the same hybrid trick as ‘pace’: journey/territory vocabulary is borrowed to carry race and competition implications.
As with pacing, I want to pull out three uses that show this hybrid framing at work: ‘frontier model’, ‘frontier developers’, and ‘paced frontier’ itself.
In ordinary discourse, the word frontier evoked a place: the outer edge of a territory, the boundary line of a settlement. In AI discourse, a ‘frontier model’ refers to the leading edge of capabilities of an AI system, but it also functions as a ranking: which lab currently has the best model, which one is in the lead, which one is winning the race. ‘Frontier developers’ in turn refers to people who work at or push that edge, but in practice picks out the small set of companies competing for market and geopolitical advantage. And ‘paced frontier’ folds pace’s slowing-down sense into frontier’s spatial one, producing a hybrid of its own: a territory that can be walked rather than raced across, if only the racers agree to it.
Frontier, unlike pace, is loaded with meaning and history. Building on the American frontier mythology of the 19th century, Vannevar Bush gave science and science policy ‘the endless frontier’; Star Trek gave us the ‘final frontier’; the mapping of the human genome was compared to Meriwether Lewis’s map of the West; John Perry Barlow called the early internet an ‘electronic frontier’; and AI developers talked about ‘frontier models’. Each borrowing carried forward the assumptions that Turner attached to the term in 1893: unclaimed territory, opportunity for whoever gets there first, an absence of existing order opening opportunities for unchecked expansion. Schneier and Sanders, writing about this genealogy, therefore argue that AI companies’ use of ‘frontier’ does real rhetorical work: it frames AI as wild ungoverned territory, exempt from the oversight and accountability that a governed terrain would have to answer to.
The trouble with metaphors and models
Max Black’s old observation about metaphors as models is useful here: a model is “an imaginative tool for ordering experience, rather than a description of the world”. This is a point geographers Livingstone and Harrison drew on in their 1980 study of the frontier metaphor and the ‘frontier model’ (p. 130) – a nice pre-AI usage of that term. A modern ‘frontier model’ is a model in Black’s sense: it doesn’t describe an AI system, it organises how we are invited to think about one; as an outpost, bounded and contested, rather than as, say, an infrastructure, a product, or a set of decisions made by named people in named companies.
And that’s the trouble with the model. Livingstone and Harrison’s old frontier-model diagram (which I can’t reproduce here, unfortunately) is quite a good model of the new AI frontier model, as the system itself, its organisation and capabilities are in the foreground. It leaves offstage the external environment with all its uncertainties and the personnel who built the outpost, chose where to place it, and are meant to be patrolling it (what I initially thought ‘pacing the frontier’ meant!). Worrying about the internal organisation of the outpost at the edge of knowledge deflects attention from asking who built the frontier, and for whose benefit.
This brings the two metaphors back together. Pacing borrows journey vocabulary to keep a race intact. Frontier borrows the vocabulary of contested territory to keep attention on the territory, rather than on those who drew its lines. Both are hybrid metaphors. “We must pace the frontier” turns out to be a remarkably economical piece of framing: it tells you to worry about speed, and where to worry about it, while not really settling the more important questions: Whose race or competition is this? And: Who benefits from it staying one.
Related blog posts
Nerlich, B. (2026). AI, metaphors and control. Making Science Public blog, 8 July.
Nerlich, B. (2026). From rogues to collectives: Myths and metaphors after the Hugging Face Incident. Making Science Public blog, 9 September.
Nerlich, B. (2026). From pause to pace: How metaphors frame calls for an AI slowdown. Making Science Public blog, 18 September.
Image: Pexels: Cowboy herding cattle in the Western frontier
(Thanks to Kate Marvel who sent me a pdf of the 1980s article on the frontier metaphor, an article that I loved reading for its early insights into metaphor theory, over and above what it said about ‘fontier’! But that’s for another time, if you are unlucky) (I used Claude for brainstorming and checking my atrocious spelling)
I am not a lawyer. So I never thought about how the race/pace/pause discourse is entangled in anti-trust law for example. If you want to know more about that read this Wired article!
Leave a Reply