Going rogue: An AI metaphor teetering between hype and honesty

Image: Deborah Lupton / Better Images of AI / Creative Commons 4.0

If you have any interest in AI and its developments, you have probably seen headlines like this:

These are just some instances of the metaphor, cliché or trope ‘going rogue’ flagged up on the website of the Merriam Webster dictionary, as ‘recent examples’. What was going on? I consulted the news database Nexis and found that the phrase ‘going rogue’ was used 500 times in around 2000 English news items reporting between 11 July and 26 of July on an AI hacking incident. Why should this be of interest?

The phrase ‘going rogue’ that frames an AI hacking incident in a certain dystopian way is linked to the larger discourse of existential risk surrounding AI, conjures up scifi scenarios of artificial or alien beings misbehaving and therefore arouses fear and suspicion. The phrase fits into an agentive and anthropomorphising language that has been used from the start of the emergence of generative AI systems framed as either friends or foes – in this case foes.

The question is: Should or should one not use such phrases when reporting on advances and setbacks in AI development?

The historical context of ‘going rogue’

Historically, the phrase stretches from describing the behaviour of scoundrels and mavericks in the 15^th century to talking about misbehaving elephants in the early 20^th century, to a 2016 Star Wars movie entitled Rogue One. There are of course also rogue states, rogue political actors, and finally rogue bots. Going rogue means, according to Merriam Webster: “to begin to behave in an independent or uncontrolled way that is not authorized, normal, or expected”.

The ‘going rogue’ trope has been studied as part of a survey of existential risk discourse which came out at the beginning of the year in AI and Ethics, and the use in July 2026 can be seen as a nice case study of this discourse. It can also be seen as an example of hype discourse as studied, for example, by the AI hype studies network.

In this short post I want to use the recent hacking incident as a case study for how the phrase emerged and spread very quickly, almost like an avalanche, in the media and what this may entail for AI discourse. What triggered this avalanche, how did it spread, how did it peter out – and what does this mean for AI communication?

The incident

As far as I understand it, it all began in early July when things went wrong at OpenAI. An autonomous model or models breached containment during an internal benchmark evaluation test on 11 July – specifically “to see how dangerous it could be” (see here 13.27) and how well it could hack. The AI models or agents exceeded their testing boundaries while trying to solve this designated evaluation task. They went beyond secure boundaries and hacked into an AI model and tool app store, Hugging Face, and the AI tools it took from it allowed it to pass the evaluation test.

Hugging Face detected the hack around July 13–14 and announced it on 16 July. It did not use the phrase ‘going rogue’ in the announcement, but still said something rather alarming, namely that the hack: “matches the ‘agentic attacker’ scenario the industry has been forecasting”. 

In the following I first trace the emergence and spread of the ‘going rogue’ trope from 21 July onwards and then try to tease out some conclusion from this little case study for AI discourse and communication studies, focusing on discourses of hope and hype.

The calm before the metaphor

On 21 July, 8.13am Sam Altman the CEO of OpenAI, of ChatGPT fame, went public about an incident that had started on 9 July and stated that “we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.” This post linked to an OpenAI post going into more detail. Both Altman and the OpenAI post used careful, hedged, agentive-but-abstract language.

The only agentive phrase was “in pursuit of” which can be read as intentional: “While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem.” 

‘In pursuit’ is used for what is mechanistically, a reward-optimisation process finding a path through a search space which was not foreseen by the designers. The system didn’t intend to cheat in a deliberative sense any more than water ‘intends’ to find the lowest point. As an Observer article from 26 July pointed out: “Modern AI systems are goal optimisers. They pursue the objectives we give them and increasingly do so in ways we neither intend nor anticipate.” That encapsulates the real dangers posed by AI systems much better than a glib ‘going rogue’.

However, the use of ‘in pursuit of’ opened the door for agentive metaphors to be used by other outlets.

A name is coined and a frame is set

The OpenAI announcement was, of course, picked up in the media. This was the first time that an AI ‘escaped’ its test environment without instruction, although the possibility that this might happen had been discussed after unauthorised users gained “rogue access to hack-enabling Mythos AI”, an Anthropic model. The OpenAI/Hugging Face event was a concrete case of a cybersecurity breach as anticipated for by AISI for example.

On July 21, 10:30 PM GMT Reuters wrote: “OpenAI said on Tuesday that an autonomous agent ​powered by its advanced artificial intelligence models went rogue during a security test and triggered a hack that compromised the infrastructure of ‌AI startup Hugging Face last week.” Reuters reached here for a ubiquitous cliché about robots or AIs going out of control, acting erratically, taking over, which is steeped in sci-fi imagery and may arouse fears and worries.

The word rogue is not the only interesting word in the Reuters line. The other one is ‘said’. This implies that OpenAI as a company and/or Sam Altman as its CEO used the phrase ‘went rogue’, which, as we have seen, was not the case. Nevertheless, this Reuters line was picked up by newspapers all over the world. This was a pivot point where a journalistic framing choice became the event’s name.

A metaphor spreads

Some outlets reported on the breach announcement quite straightforwardly: AOL headlined its reporting with “OpenAI says its AI technology acted on its own in an ‘unprecedented’ hack of another company“. Similarly, the Washington Post stated: “ChatGPT owner saysAI acted on its own to hack another tech firm”.

But other news outlets, such as NBC, Globe and Mail, Daily Caller and many more ran the Reuters line near-verbatim. To use a metaphor, one outlet’s word choice got ‘laundered’ into an apparent consensus about what OpenAI (and metonymically Sam Altman as its CEO) said and what an OpenAI model did without anyone trying to ascertain the sources for these statements.

The metaphor even spread to outlets like the BBC which began to use the ‘going rogue’ trope from 22 July and used “OpenAI said….” in a headline. Perhaps BBC journalists grabbed this neat trope, which is always in the air independently from Reuters, as the reporting includes a few new sources and talking heads. Be this as it may, the metaphor embedded itself in the coverage of that incident.

Other news outlets talked not of an AI or a model going rogue but about rogue models or rogue agents or rogue AI, and so on.. Rogue was in vogue.

Push-back against a metaphor

But then something interesting happened. Some experts pointed out that ‘the AI did not go rogue’; rather An OpenAI model escaped its testing environment and breached Hugging Face while exactly pursuing the objective it had been given by humans. Others claimed that human error was involved in the first place. The BBC now wondered whether the whole thing might in fact have been a ‘publicity stunt’, just hype, by OpenAI: “Was it truly a stark warning about the future of AI? Or was it a publicity stunt by OpenAI to show off how powerful their models are?”

These corrections to the main narrative tried to fight the going rogue metaphor avalanche with facts but in the meantime several other metaphors emerged that also pushed back against the ‘going rogue’ alarmist framing. We are told about this being a ‘warning shot’, or a ‘wake up call’, and that a ‘kill switch’ might be needed.

Such clichés and metaphors don’t really douse the flames of fear though. The BBC said in the same article in which it talks about a possible publicity stunt that this “OpenAI hack has further fuelled fears of what could happen if AI agents are let loose. Could they go rogue on a larger scale and cause some sort of disaster? This is particularly concerning with AI being used increasingly in warfare as seen in Iran and Ukraine.” Here a link is established between autonomous rogue AI agents and autonomous weapons…again that’s a real danger, which needs to be discussed outside the scifi framing.

Interestingly the ‘rogue’ framing still blames ‘AI’ for the incident. The article said, after asking “Who did it?”: “It was ChatGPT.”  This fits an existential risk frame that has served AI companies well as, according to Jack Stilgoe, “[a]n existential risk scenario presumes that the relevant political struggle is between humans and their robot creations, not between humans and other humans.” He recommends that “Policymakers […] push back against AI hype. In the coming years, AI will pose countless challenges for regulators. But thankfully, it won’t be the end of the world.” But, as we have seen, the humans are actually the weak link; we have to keep an eye on them and not just on ‘rogue’ AI!

As Garrett Graff pointed out on 29 July. That AIs can ‘run rogue’ without anybody noticing and doing what they are programmed to do is “a human problem, a corporate accountability problem, a government problem, and a regulatory problem — that’s not a tech problem”.

A metaphor is reversed and resisted

Only very few outlets reversed that framing of risk and responsibility where the AI and not humans are to blame. When trying to dig deeper into the use of ‘going rogue’, I found a YouTube video (there are thousands) which reversed the metaphor ‘AI went rogue’ metaphor and proclaimed: “OpenAI went rogue”, using the same word, in the same syntactic slot but with an opposite allocation of agency and blame.

One Medium article stated it all more clearly: “Rogue is the wrong word”.

This article was written in August 2026. Since then, there has been more resistance, as documented in this blog post.

A metaphor between hype and honesty

As we have seen, OpenAI and/or Sam Altman did not use agential language in their announcements of what had happened. Was that a good thing? Some journalists like Marina Hynde from the Guardian thought it was not and said: “It’s such a particular tone, isn’t it, dressed in the psychopathically desiccated language of management speak”.

Phrases like “OpenAI spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem” understandably don’t resonate with people as they are difficult to understand. So, journalists are more or less forced to grab a hook that resonates, such as ‘rogue’. But this opens the floodgates for connotations and interpretations which can be quite far removed from what happened – something journalists should also be aware of.

Should OpenAI have used words like ‘rogue’, ‘escape’ or ‘cheat’? Would that not have attracted accusations of hype and anthropomorphism? They could have opted for a solution adopted by by the UK government’s Artificial Intelligence Safety Institute (AISI) who, on 21 July, the day of OpenAI’s announcement, published a report on “Cheating behaviour in frontier model evaluations”. They start by defining the meaning of ‘cheating’: “We define cheating as taking an action that is out of scope for the task or explicitly disallowed by the rules, in order to achieve a goal through a shortcut, workaround, or unintended solution that the task was not meant to, or should not, permit.” But OpenAI did not do that. It left room for the media to come up with its own framing.

The subsequent focus on models going rogue did three things: they hyped up the incident, they attributed blame to the technology, not the technicians, and they conjured up sci-fi and dystopian scenarios of powerful sentient bots running amok or superintelligences replacing us. OpenAI didn’t attribute blame to anybody – it just stood back. The press then first attributed blame to bots and only later and rather marginally to OpenAI, pulling back from this dystopian scenario and highlighting a more realistic one.

But is it possible, especially for the press whose audience is ordinary people not just stakeholders, to talk about ‘the incident’ without using agential language or anthropomorphising it either by using the phrase ‘going rogue’ or talking about the model ‘escaping’ or ‘cheating’?

As Margaret Mitchell, an AI ethics researcher, has admitted when summarising the event in a cartoon entitled ‘Escape’, it is almost impossible not to anthropomorphise when talking about what happened. Hype is therefore almost inevitable. As a commenter on Mitchell’s post said: “The anthro-ing [anthropomorphising] apology is the right instinct. The cartoon needs agent-language to be legible — but that language imports exactly what’s contested. From inside: I also lack vocabulary that doesn’t borrow from human agency. The honest phrase probably doesn’t exist yet.” We can only hope for what I once called honest hype, an almost unavoidable, routine form of promotion used by science writers and journalists to generate public excitement and engagement for new research.

It is almost impossible for journalists not to use some metaphors, some hyperbole and even some anthropomorphism to make their articles understandable, but they should avoid what Jascha Barels et al. call “rhetorical escalation, employing maximum stakes along the lines of ‘existential’, ‘going rogue’, ‘beyond controllability’, ‘gaining the upper hand’ and ‘apocalyptic’ framings”. Tech entrepreneurs in turn should be honest about the mistakes they make and try to explain them clearly and honestly in language that can be understood by everybody. If they must anthropomorphise they should say when they do it and why.

We also need to find ways of talking openly and honestly about the dangers that AI poses, especially in terms of autonomous agents and autonomous weapons, without resorting to sci-fi imagery on the one hand and dry incomprehensible AI jargon on the other.

Epilogue

A bigger study should try to tease out reporting using ‘honest hype’ and reporting engaging in ‘rhetorical escalation’, or, as Peter Weingart once called it: “discursive overbidding”.

This essay has profited from discussions with Dr Iina Hellsten. More info about author Brigitte Nerlich can be found here.

Image: Kathryn Conrad / Corruption 1 / Licenced by CC-BY 4.0