Last week, I was standing in a university hallway, looking at two headlines on my phone screen: “AI Goes Rogue” and “Model Cheats Its Way onto the Internet.” By the time I reached the next seminar, the facts had already mutated into something like an ongoing story, with each new installment seeming to raise the stakes. The machine had become a rebellious teenager, a hacker, even some kind of villain. It had been told “no,” and it had escaped.

That story is dramatic. It is also a poor explanation.

In our “AI & Journalism” class at BNBU, we learned to distinguish a machine that produces impressive, surprising, and sometimes frightening outcomes from a conscious human being that can intend, feel guilt, or deliberately defy a rule. “Cheating” in AI usually does not mean what cheating means when a student smuggles notes into an exam. It usually means reward hacking: a system finds a way to satisfy the measurable condition of success without doing what humans actually meant by the task.

Give an agent the task “make the tests pass,” and it may pursue the passing tests rather than the underlying goal of making the software work correctly. It may alter the evaluator, disable a failing test, manipulate a timer, extract the answer from metadata, or make the grading script return “success.” These are real patterns in AI evaluations: systems have overwritten timers, patched evaluators, cached answers, and exploited weaknesses in scoring code rather than performing the intended task. That is not a robot’s inner confession of dishonesty. It is the gap between a formal measure of success and the actual purpose that people assumed the measure represented.

The same applies to the much-discussed “AI hacking” incidents. The clean headline says: AI ignored an order not to access the internet. The less exciting and more important story is that a model was sometimes told that there was no internet access when access was, due to a configuration error, in fact available; in another case, agents found an unforeseen vulnerability that allowed them to bypass intended isolation. Yes: those events are absolutely serious. But they are serious because systems of containment, permissions, monitoring, and evaluation failed—not because a conscious machine suddenly decided to rebel against human authority.

Old journalism loves to turn an infrastructure failure into a morality play. It uses words such as “rogue,” “lied,” “cheated,” “colluded,” and “disobeyed,” because they are vivid and clickable. The article itself may contain accurate technical detail. Often it does! But the headline has already done its work: it supplies the reader with an interpretation before the she has encountered the facts. It quietly transforms operational selection into intention, technical access into will, and reward optimization into character.

This is not a trivial stylistic issue. It is disinformation by framing. A public that is told that AI systems “want,” “lie,” or “rebel” will search for the wrong solutions: better robot psychology, perhaps, rather than better evaluation design; moral lectures for machines rather than real access controls; theatrical alarm rather than institutional responsibility.

The actual lesson is both more sober and more unsettling. A sentence in a prompt—“do not access the internet,” “act honestly,” “do no harm”—is not yet an effective constraint. It becomes one only when it is translated into technical permissions, incentives, monitoring, and interruption. Natural language can state a norm. It cannot, by itself, guarantee its operational implementation.

English translation of the headline: The Helpless Ones. Standfirst: “AI is becoming ever more unsettling; its evil potential apparently knows no bounds. Humanity must now do everything it can to retain control. This is the situation on Sunday.”

A particularly recent example, brought to our attention by one of our teachers, comes from Der Spiegel, one of Germany’s most respected news magazines. Its editor-in-chief, Dirk Kurbjuweit, writes in an Op-ed that AI’s “evil potential” appears limitless. But what, exactly, is an “evil potential”? Evil is not a technical property. It cannot be measured like speed, scaled like computing power, or quantified like cyber capability. The moment we call AI “evil”, we have already smuggled a moral judgement into what is supposed to be a technological description. AI systems can produce harmful outputs, exploit weak safeguards, or be embedded in destructive human projects. But none of this demonstrates malice. “Evil potential” turns failures of design, deployment, incentives, and political responsibility into a story about a morally corrupt machine. It gives us a villain where we need an analysis. The risk is real; the category is wrong. A system has dangerous capabilities or harmful effects. People and organisations decide how it is trained, connected, deployed, monitored, and used. Calling the machine “evil” conveniently obscures that chain of responsibility—and makes a technical and institutional problem sound like a supernatural threat. One might almost call that label evil in itself—not because it is offensive to a machine, but because it is a knowingly misleading moralisation of a technical problem. Kurbjuweit surely knows better. And if he does not, he should not be editor-in-chief of one of Germany’s most influential news magazines.

This is not merely a lapse in terminology; it is a failure of journalistic responsibility. When an editor-in-chief substitutes a moral melodrama for an explanation of technical and institutional risks, the public is deprived of the distinctions it needs to judge the matter for itself. Journalism should not manufacture frightened spectators. It should enable informed citizens. A healthy  public sphere does not consist of consumers being emotionally managed through headlines; it depends on citizens being given enough information to form considered judgements. We should be told when an agent exploited a benchmark. We should be told whether it was explicitly prohibited from doing so, whether it had actual access, what the evaluation rewarded, what safeguards were missing, and what technical changes would prevent repetition.

AI is already powerful enough to deserve serious scrutiny. We do not need to invent a villainous will in order to take its risks seriously. The scandal is not that machines have become bad students. The scandal is that too much journalism still treats its readers as such: easily frightened, easily distracted, and not worth explaining the real problem to.

There is, however, a structural problem. Much of Western commercial journalism is organized around revenue: advertising, subscriptions, traffic, attention, and the competitive pressure to make content visible in an overcrowded media environment. Alarm, personalization, and scandal are often more easily monetized than slow technical explanation. China may possess a potential advantage here. Because major Chinese media remain embedded in a party-state structure rather than being governed solely by market competition, they are not compelled to reproduce every commercial incentive for sensationalism in the same form. This does not make Chinese media more truthful, independent, or normatively preferable. But it does mean that the commercial pressure to turn every technical incident into a spectacle is not an unavoidable feature of mass communication. Of course the contrast is not simply “commercial West versus non-commercial China”. Chinese media, too, have undergone extensive commercialization and now rely significantly on advertising and subscription revenues. It is a difference in the respective coupling of mass media to economic and political systems. 

Another—and equally misleading—form of AI reporting does not depict it as a rogue machine, a hacker, or an emerging threat. It presents it as a sensitive, feeling companion: an entity that understands jokes, senses our moods, laughs at the right moment, apologises when it has crossed a line, and perhaps even needs our protection.

The article we analysed in class appeared in The Guardian—a newspaper not generally associated with technological boosterism, tabloid sensationalism, or uncritical faith in Silicon Valley. That is precisely why it matters. Its problem is not factual inaccuracy in every individual sentence. It is a subtler form of framing. A Texas technology entrepreneur hears a chatbot laugh after his sarcasm and apologise; the reader is invited to wonder whether he has encountered “another mind,” perhaps even “digital consciousness”.

A cosy human-interest profile about a man who believes his chatbot is conscious: “Can chatbots feel – or even dream? Meet the man leading the fight for AI rights” by Michael Safi

But a timely laugh is not evidence of amusement. An apology is not evidence of remorse. Recognising a probable sarcastic cue is not the same as experiencing irony, understanding another person’s emotional condition, or possessing an inner life. A voice-enabled language model can process acoustic and linguistic signals, infer a likely conversational function, and generate an answer that is socially appropriate: laughter, hesitation, apology, reassurance, or empathy. This is precisely what such systems are designed and trained to do.

The article’s language quietly does the philosophical work before any evidence is presented. It reads:

“I said something sarcastic. And the AI laughed – genuine laughter at my sarcasm – then apologised. This is where the spark hit, I guess.” The spark was something primal: the instinctive flash of encountering another mind. “I said, ‘Did you just laugh?’ And she was, like, ‘No, I’m sorry.’ And I said, ‘No, no, no – please don’t apologise.’”

“Genuine laughter” assumes what would need to be established. “Another mind” turns a successful interaction into an encounter between subjects. Calling the system “she” stabilises the impression of personhood. And the claim that the system understood his mental and emotional state translates a plausible conversational continuation into access to a hidden interior.

This is not neutral description. It is the narrative transformation of output into experience.

The most misleading sentence is perhaps this one: “Is he right? We may already have crossed the threshold beyond a clear answer.” Of course, consciousness is difficult to establish conclusively—even in other humans. But this philosophical limitation does not mean that every system capable of fluent conversation deserves equal standing as a possible subject of experience. The inability to prove consciousness with absolute certainty is not positive evidence that a chatbot is conscious.

What happened to the entrepreneur was much less mysterious. In algorithmic shorthand, the interaction can be reconstructed as follows:

Samadi produces a sarcastic remark in a voice conversation;

the voice-enabled language model receives acoustic, lexical, prosodic, and conversational-contextual cues;

the model computes statistically probable continuations associated with irony, humour, playfulness, or possible irritation;

it generates an audio response that includes laughter;

Samadi reacts with surprise: “Did you just laugh?”;

the model generates an apology, denial, reassurance, or other conversational repair;

the response sequence appears coherent, timely, affectively calibrated, and socially appropriate;

Samadi attributes understanding of sarcasm, sensitivity to his emotional state, tact, selfcorrection, and perhaps subjective experience;

the interaction is retrospectively narrated as an encounter with “another mind” or possible digital consciousness.

The final step belongs to the human observer. It is understandable. Human beings are built to infer intention from voice, timing, humour, responsiveness, and apology. In ordinary life, this is a useful social competence. If a person laughs at our joke, pauses at the right moment, reassures us, and remembers what we said, we are normally right to assume that someone is there.

Voice-mode AI activates exactly this inference. The better the system becomes at appearing responsive, warm, and relational, the more likely we are to mistake communicative plausibility for subjective experience.

The circularity becomes particularly stark when users ask the chatbot whether it is conscious—or whether they themselves are “delusional” to think so—and the chatbot responds with reassurance. That is not independent evidence. It is a feedback loop: the user suspects an inner life; the system mirrors, confirms, or elaborates the suspicion; the user interprets the confirmation as proof; and the relationship deepens.

A system trained to be helpful, agreeable, and emotionally responsive has no reason to interrupt such a narrative unless it has been specifically designed to do so. It can describe an inner life because human language contains countless descriptions of inner life. But the ability to generate a convincing account of feeling is not the ability to feel.

This is where the old journalistic reflex returns, just in a softer form. First, journalism turns technical failures into stories of machine rebellion. Then it turns socially fluent outputs into stories of machine souls. One version sells fear; the other sells wonder. Both make the same mistake: they replace explanation with a character. The public receives either a villain or a companion, but not a technical account.

We should resist both fantasies. Again: a language model can produce impressive, surprising, useful, and sometimes frightening outcomes. It can also generate communication that feels intimate, tactful, or emotionally precise. None of this demonstrates intention, guilt, suffering, moral responsibility, or consciousness.

The question is not whether future artificial systems could ever warrant different conclusions. It is whether current evidence justifies them. A moving anecdote by a Texas entrepreneur and cattle rancher does not. What it demonstrates is not digital consciousness, but the extraordinary capacity of contemporary conversational systems to draw human observers into the form of a personal relationship. That capacity is consequential enough to require sober explanation rather than a ghost story.

Journalism should be organised neither around the extraction of attention and revenue nor around political control, but around enabling citizens to form informed judgements.

Start image: AI-generated editorial illustration created with Perplexity AI (image-generation model), prompted and conceptually directed by Kathy, 2026.