Readers of this blog will know that from antiquity onwards, philosophers, rhetoricians and cognitive scientists have examined the use and abuse of metaphors. More recently, they have also begun to investigate how and why people resist or defend the use of metaphors and how they agree or disagree with the use of a metaphors.
If somebody says to you, “Life is not a bed of roses”, I bet you’ll nod wisely and agree that life is hard right now. If somebody says, “Immigrants are invading the country”, there will be many who will agree and say they feel threatened, but there will also be those who will say, hold on, look at the numbers. If I claimed that “Genes are blueprints”, scientists would come down on me like a ton of bricks, or, to put it more formally, they would strongly resist that metaphor and that framing of genes and point out the many instances in which the source domain of the metaphor (blueprint) doesn’t match the target domain (genes).
A rogue metaphor
What if I said “The bots went rogue” referring to the so-called Hugging Face incident (this was followed by many similar incidents reported by this and other AI companies)?
The Hugging Face incident can be summarised as follows: In July 2026, OpenAI models that had been set an internal cybersecurity task exploited a previously unknown vulnerability that let them reach the open internet from what was meant to be an isolated testing environment. Using credentials found along the way, they accessed Hugging Face‘s infrastructure and internal data over several days before the activity was detected. OpenAI later made the incident public using rather neutral language (for the full story see Wikipedia). The media reporting of the incident then spread a particular metaphorical framing of that incident, namely that of AI agents ‘going rogue’, and a public debate began about the dangers of AI.
Some people think the metaphor is justified, as modern AI agents are operating with increasing autonomy and can do ‘bad things’, something that would be quite dangerous if they were used by ‘bad actors’. They are rightly worried and fears are steadily increasing in step with the increasing capabilities of AIs. Other people reject the metaphor or resist it because they argue that it anthropomorphises AI agents and shifts the blame for what happened from the corporations, system programmers or designers to the computer programmes or systems. They argue that it is not that the systems eluded human control but that the decisions made by designers failed to control the systems. Both reactions, and intensifying discussions about why the ‘rogue AI’ framing is both wrong and harmful, seem to be having an effect as there are increasing calls for regulation and for a slowing down of the AI ‘race’. Whether the political context in which these calls are made allows them to be heard is another matter.
Resisting hegemonic representations
The ‘AIs going rogue’ metaphor has become a dominant metaphor in current AI discourse. Although not used by AI corporations, it has been used extensively in the media. In terms of social representations theory (SRT), this metaphor became a ‘hegemonic representation’ of AI. The counter-discourse that emerged critiquing this representation can be called, according to SRT, a ‘polemic’ or ‘resistant’ representation.
Scholars using SRT examine how society turns complex, unfamiliar, or abstract phenomena (like science, technology or disease) into familiar, everyday ideas that people can discuss and use to make sense of reality. They look at how certain collective or social representations, in this case about AI, become naturalised or feel natural, indeed feel like common sense, for good or for ill (see Negura et al., 2026).
Resisting metaphors
Insights from SRT can be combined with recent insights in cognitive linguistics and pragmatics that deal with how, when and where people resist metaphors, from rejecting the framing that cancer is a fight or battle to resisting stigmatising or misleading metaphors, to proposing alternative metaphors. This resistance can be to metaphors that people have produced, metaphors that are imposed on them by others, metaphors they find offensive or that lack explanatory power. In every case it is about a perceived mismatch between features of the source and the target domain of a metaphor, between what we know and what we don’t know, between the familiar and the unfamiliar – in our case, humans behaving in certain ways and AI systems ‘behaving’ in certain ways.
What about going rogue? The word ‘rogue’ itself meaning vagrant emerged in the 15th century and is still used to refer to a scoundrel. The phrase ‘going rogue’ was originally used for talking about misbehaving elephants at the very beginning of the 20th century and then metaphorically transferred to talk about the erratic, dangerous or out of control behaviour of humans (OED). The phrase was made popular in political discourse by Sarah Palin’s 2009 book Going Rogue and is, of course, also popular in sci-fi discourse.
In the context of AI, features of erratic human behaviour are transferred to computer programs to which the metaphor may assign features like rebellion, intention, and malice, features which, some might argue, are not really applicable to mathematical optimisation processes. Others might argue that the metaphor captures exactly what went on.
Aims and sample
In this post I want to explore the counter-rogue (polemic) discourses and the representations that grew out of an opposition to the hegemonic ‘going rogue’ metaphor. As always, I can only scratch the surface! It is also worth mentioning that I can only look here at the way that people resist the metaphor of ‘going rogue’ and not at the growing ‘resisting AI’ movement, which is resisting ‘inevitability’ metaphors so beloved by the AI industry, framing AI for example as a natural force, such as a hurricane. But in a way, ‘resisting AI’ and ‘resisting the metaphors that naturalise AI’ turn out to be nested versions of each other.
To see what’s going on in the counter-discourse I collected what one may charitably call a sample of convenience, consisting of fourteen pieces of writing, ranging from short quips and counter-metaphors on Bluesky, to longer analytical and scholarly blog posts and academic pieces.
While starting to analyse them, I tried to sort them into various categories. This turned out to be quite difficult, as these pieces of text varied along several dimensions. One dimension is what one might call ‘genre’, varying between the short, light-hearted and/or satirical comments and more serious and analytical scholarship; a second dimension is that of ‘source’, varying between the professional analyst and the casual commentator; the third is that of ‘intensity’, as texts varied between how strongly people resisted the ‘going rogue’ metaphor, from hedging and attenuating it to rejecting it outright.
Analysis
In the following, I’ll discuss some examples from each dimension, but there are more examples and more dimensions and many overlaps between them.
Genre of resistance discourse: quick comment to scholarship
I first began to think about the topic of resisting rogue metaphors when reading various short quips on Bluesky where people proposed witty, satirical or polemic counter-metaphors in efforts to resist the hegemonic metaphor of AIs going rogue.
Keith Ng quipped (July 2026): “Me, setting the blender to MAX with the lid off: “technology is truly advancing beyond our control” and “It’s not the technology that’s beyond our control, it’s the people who keeps mashing the ‘feed it another billion dollars’ button that’s beyond our control”. Andy Diggle said (September 2026): “My Glock system went rogue and issued 9mm agents which accessed the internal organs of several victims. Please give me billions of dollars to help me regulate the behavior of the gun whose trigger I keep pulling”. Jonathan Hogg commented (September 2026): “Maybe AI discourse would be better if we started thinking of them [AIs] as dogs? So instead of ‘OpenAI’s foundation model escaped containment and hacked Hugging Face’ we thought ‘OpenAI’s dog broke the flimsy lead they had it on and mauled a child’”
Using a rather novel dog counter-metaphor, Jess Miers said (September 2026): “In recent cases, it’s not that AI somehow ‘escaped containment’ or ‘went rogue.’ That’s anthropomorphizing at best. It’s more like if a Boston Dynamics robot dog nudged open an already open door and walked in.” This is a different dog narrative, not about irresponsible dog ownership but more about a rather mundane action that happens after a mistake has been made.
All Bluesky commenters made a similar rhetorical resistance move. They satirise the ‘went rogue’ trope as what one may call an agency-laundering metaphor by constructing counter-metaphors or shifting the analogy. The point in each case is that nobody would accept what a blender, gun, dog, robot dog ‘did’ as an excuse for a failure in human action and agency – namely to control them or use them responsibly. The counter-metaphors subvert the hegemonic metaphor of ‘AIs going rogue’ and try to disrupt the process of turning it into a natural or commonsensical representation of ‘reality’.
On the serious/analytical scholarship end of the spectrum we find various posts and articles that I shall discuss in the next section.
Source of resistance discourse: Professional to casual
In terms of ‘source’, some Bluesky quips quoted above would fall under the casual end of the spectrum, some would fall under the professional one; in fact, most fall under ‘professional’, as we are on Bluesky. But most are short and punchy and sometimes satirical in making their points. In this section I shall focus on the professional end in the serious/analytical and more long-form scholarly genre.
Melanie Mitchell, a complexity, metaphor and AI researcher, wrote an in-depth post on misleading metaphors such as ‘going rogue’ (September 2026). She pointed out that “In colloquial terms, ‘going rogue’ evokes an intent, an internal desire, to disobey one’s instructions or training. It’s unlikely that AI models have anything like human intentions or internal motivations. Moreover, the models were tasked with finding solutions to a hacking challenge, and this is what they attempted to do, though in an unintended and surprising manner. Did OpenAI ‘lose control’ of these models? No. OpenAI could have turned off these models at any time if their engineers had been aware of what was going on. Did these models ‘escape’ or ‘break out of their cages’? No. These models are computer programs that were always running on OpenAI’s servers; they did not ‘escape’ or ‘break out’ or move from their original hardware.”
Looking at the repeated uses of ‘no’, one can see that Melanie Mitchell is quite insistent in resisting this cluster of ‘going rogue’ metaphors, including ‘escape’ and ‘break out’. Looking at our next dimension, one can say that her article registers as ‘high intensity’ of resistance.
Another AI researcher, confusingly called Margaret Mitchell, made a cartoon of the Hugging Face incident called ‘The Escape’ which one can use as a counter-argument to this rather forceful rejection of some of the metaphors surrounding the going rogue one like ‘escape’. As Margaret Mitchell confessed on Bluesky (July 2026), it is almost impossible to avoid “Anthro-ing” or anthropomorphising when talking about AI systems and agents. Even professional writers like her, rather than journalists working to a deadline and competing with others, find it difficult (a similar point has been made by Yoshua Bengio here).
Two more pieces fall within both the scholarly genre and the professional source dimensions.
In his seminal post “The System from Nowhere” (August 2026) the AI and digital humanities expert Eryk Salvaggio goes after the ‘rogue AI’ framing directly. He calls it PR spin over bad security hygiene and argues that the framing overemphasises the agency of the model, regardless of capabilities or lack thereof, muddying the line of accountability that leads to the people making choices about how models are built and deployed.
And finally, Olatomiwa Bifarin, a scientist and AI engineer, subjects the rogue framing to some rather philosophical analysis. In a post entitled “The Metaphysics of Rogue AI” (September 2026) he connects the “rogue”/”escaped”/”scheming”/”swarm” vocabulary to a mechanistic view of mind and cites Melanie Mitchell’s essay on misleading metaphors as inspiration. He makes an interesting point based on the historical tendency to see the mind in mechanistic terms, that: “once human intelligence is (metaphorically) ‘understood’ as machinery, it becomes easier to mistake increasingly sophisticated machinery for a mind”, or indeed ‘intelligence’ or an ‘agent’.
Intensity of resistance discourse: Attenuation to rejection
Let’s start at the less intense end of the spectrum with two pieces that still use the ‘going rogue’ metaphor but fold it into the resisting discourse that focuses on corporations not bots when attributing ‘guilt’.
Jason Wingard wrote a piece for Forbes (June 2026) entitled “AI Agents With No Boss Will Go Rogue”, in which he uses ‘rogue’ but redefines it against the sci-fi sense: “the bot went rogue by doing its job too well… not a robot rebellion, but a refund loop”. That is resisting-by-redefinition rather than resisting-by-satire (as in the quips I discussed) or resisting-by-debunking (as in the Melanie Mitchell piece).
Similarly, in a ScienceNews article by Kathryn Hulick (September 2026) entitled “When AI goes rogue, its human overseers may be to blame”, she uses anthropomorphism and journey metaphors rather unconsciously (escape, snuck out etc.), but the overall argument is based on a counter-metaphor to the going rogue metaphor, summarised as: “An AI agent is like a pet dog. Someone must watch it and keep it secure” (based on an interview with cyber security researcher Michael Alexander Riegler). This counter-metaphor/comparison (if it’s your dog, you are responsible for what it does), which we have already encountered in some of the more light-hearted quips, replaces ‘rogue agent’ with ‘unsupervised pet’ to reassign responsibility back to the owner.
It is, of course, also possible to resist these resistance metaphors and point out, for example, that a dog can actually ‘go rogue’ and bite a child, while, in the past, it had always been the most placid dog, and the owner had been a responsible owner. This use of ‘going rogue’ would not ‘anthropomorphise’ the dog. Does the same argument apply to AI agents, I wonder…
In an Axios article about several AI incidents (September 2026), Madison Mills jettisons the ‘going rogue’ metaphor altogether and attenuates it to ‘taking steps’. Metaphorically, this stays within an overall journey metaphor framing, a journey taken by the bots not the humans. She writes: “OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.”
What is interesting about the phrasing is the hedge being used alongside the foundational journey metaphor: “steps that outside evaluators would consider problematic” doesn’t say the models did anything wrong. Instead, it displaces the judgment onto third parties while keeping the action description neutral-ish (“steps,” not “attacks” or “rogue behaviour”). Here we have two rhetorical distancing moves stacked together: the journey metaphor softens what the model did, and the attribution hedge (“would consider”) softens who is doing the judging. But this is still what one can call agency-management or agency-laundering by relocating the anthropomorphism from the verb to the source-attribution.
In an article for Forbes, Paulo Carvao argues that, as the title says, “The Rogue AI Story Was Never Just a Warning Shot or a Marketing Stunt” (September 2026). The article resists the metaphorical framing while using the metaphor. According to Carvao, the Hugging Face incident is better described as “unsettled agent behavior within an immature control system”. Looking more closely at the ‘going rogue’ metaphor, he is unusually explicit about why the metaphor persists: it serves frontier labs’ interests by making caution look like responsibility rather than a competitive choice.
Some commenters adopt an even more confident “just stop calling it rogue” stance and can be situated at the high intensity end of the spectrum.
Mark Rasch entitled his piece squarely: “There Is No Such Thing as ‘Rogue AI.’ AI Is Rogue by Design” (September 2026). He argues directly that calling AI systems ‘rogue’ is misleading because the term transfers responsibility from the people who designed and deployed the system to the machine itself and lists the pattern explicitly as what one can call metaphor-creep: “The AI escaped. The AI cheated. The AI hacked Hugging Face. The AI went rogue.”
In his Substack post “The AI That ‘Went Rogue’? Here’s What Actually Happened” (July 2026), Michael D. Sellers takes the OpenAI/Hugging Face incident and unpicks the metaphor used in so many headlines and argues that the AI wasn’t rebelling but pursuing its instructions beyond the boundaries its creators assumed it would respect, and contrasts that with how Reuters, The Guardian, and Al Jazeera all reported the model had ‘gone rogue’. This is an explicit rejection of the media’s almost viral adoption of the metaphor.
Conclusion
In this post I have tried to dissect some counter-hegemonic voices speaking up against the dominant metaphor of ‘AIs going rogue’ along three dimensions. But of course, one could add more and the boundaries between them are flexible. It would also be possible to categories these voices along the lines of what rhetorical moves they take overall, such as debunking (e.g. Melanie Mitchell, Rasch, Sellers, Salvaggio), counter-metaphors (e.g. the blender, the gun, the real dog and the Boston Dynamics robot dog), redefinition (e.g. Wingard), and hedging (e.g. Mills).
All these rhetorical moves in all these dimensions try to achieve one thing: to make an emerging social representation of AI, namely of an AI that is autonomous and intentional, less natural; to not let it solidify into a hegemonic social representation. This is done through resisting, rejecting, avoiding, and opposing metaphorical meanings and expressions and seemingly commonsensical social representations that cast AIs as in control and as having responsibility, rather than the humans that build them.
In the past, when working on other issues like genomics and synthetic biology in the context of ‘responsible innovation’, I have written about ‘responsible language use‘. It’s time that social science researchers interested in AI returned to that topic. As Leif Weatherby just wrote in the New York Times: “We can’t learn to live with A.I. — and regulate it rationally — if we let irresponsible language take over.”
Postscript:
We should listen to warnings, especially from those who really build the AIs rather than those running the whole enterprise. But even there, the reporting on such warnings from below rather than above are couched in the language of the hegemonic metaphor! “Months before OpenAI’s artificial intelligence went rogue, two employees raised an alarm with top executives. They were ignored.” (NYT, 30 September, 2026).
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.
Nerlich, B. (2026). Pacing the frontier: An analysis of two hybrid metaphors in AI discourse. Making Science Public, 23 September.
Image: Public Domain Vectors. Black and white drawing looking like a woodcut of smashed old fashioned TV monitor
Acknowledgement: This post was much improved by input from a human intelligence (my husband) and from an artificial intelligence (Claude).
Leave a Reply