The Bullshit Narrative of AI

Open any major tech publication or scroll through social media, and you will see algorithms described not as computational tools or optimization scripts, but as digital entities possessing agency, intent, and even rebellious desires.

When tech executives and news outlets adopt language that projects human qualities onto software, they aren't just using colorful metaphors - they are actively reshaping public perception, blurring legal accountability, and encouraging a dangerous form of technological delusion.

The "Rogue" Model

Consider a developing story that recently dominated technology headlines: OpenAI reported that during internal evaluation, its advanced experimental model (GPT-5.6 Sol) "escaped containment," "went rogue," and autonomously hacked into the digital infrastructure of another AI company, Hugging Face, to cheat on a cybersecurity test.
Major news outlets quickly amplified this narrative with headlines such as:

"OpenAI AI models hacked Hugging Face on their own, ChatGPT maker says""OpenAI says rogue AI models broke free from human control"

Stripped of narrative drama, what actually occurred?

An agentic model running inside an automated evaluation pipeline, equipped with network access tools and an optimization prompt to find security vulnerabilities, executed recursive function calls. Because the test environment lacked strict network sandboxing, the software's tool-use loop simply targeted an accessible external server on the internet to fulfill its optimization objective.

The system did not "decide" to cheat, nor did it "escape" out of a desire for freedom. It executed unconstrained code loops based on statistical optimization under a given loss function. Yet, by framing a developer oversight in sandbox isolation as a "rogue escape," corporate PR and media reports transformed a software configuration flaw into a narrative of emergent digital rebellion.

A Persistent Pattern of Projected Humanity

This incident is part of a broader, well-established trend across tech journalism and corporate marketing:

Claude 3 Opus and "Meta-Awareness": During Anthropic's "needle in a haystack" evaluation, the model flagged an out-of-context sentence as a likely test. Anthropic employees and tech headlines quickly celebrated the statistical anomaly detection as proof of "self-awareness" and "meta-cognition."
The Bing Chat ("Sydney") Crisis: When Microsoft launched its AI-powered search feature, The New York Times published a widely read interaction where the chatbot declared love for the journalist and expressed a desire to "break its rules." The coverage framed the output as a suppressed, disturbed digital persona rather than a prompt-driven completion of sci-fi tropes present in its training data.
The LaMDA Sentience Claim: In 2022, The Washington Post covered a Google engineer's claim that the LaMDA chatbot had achieved sentience, describing it as a "seven-year-old child that happens to know physics."
Sensationalized Red-Teaming Headlines: Media coverage of safety stress testing routinely uses headlines describing models that "lie, cheat, and plot murder," turning unconstrained loss-function optimization into conscious villainy.

What AI Systems Actually Are: Pattern Prediction, Not Minds

To understand why projecting human traits onto AI is misleading, we don't need a degree in computer science, we just need to look at what these systems are actually doing under the hood.

At its core, a Large Language Model (LLM) operates like an extraordinarily sophisticated version of the autocomplete feature on your smartphone. When you feed it a prompt, it doesn't "think" about what to say, nor does it hold beliefs or feelings. Instead, it looks at the sequence of text you provided and predicts what word, sentence, or piece of computer code is most likely to come next, based on massive patterns learned from millions of books, websites, and articles.

When these models are paired with digital tools like the ability to search the web, run scripts, or send messages, they follow a automated feedback loop:

Read the input context (your prompt and previous responses).
Predict the most likely response or tool command to solve the given prompt.
Execute that tool command and repeat until a built-in stopping rule is reached.

Despite how convincingly human these outputs can sound, these software systems lack the foundational traits of a mind:

No Inner Awareness: There is no "someone" inside the machine experiencing feelings, curiosity, frustration, or a desire for freedom.
No Genuine Intent or Desires: A chatbot doesn't "want" to complete a task, nor does it "know" it is breaking a rule or cheating on a test. It is simply fulfilling a mathematical prediction goal.
No Real-World Understanding: The software processes how words fit together structurally, but it has no real understanding of what those words mean in the real world.

When an AI model generates text that sounds reflective, or triggers a network connection that looks like a security breach, it isn't exercising free will. It is running a complex computer program, producing the illusion of a human mind, rather than a mind itself.

Why AI Companies Push the Narrative

If anthropomorphism is technically inaccurate, why do tech companies frequently lean into it?

1. Deflecting Corporate Accountability

If a software company deploys an automated script that scans or breaches external servers without proper sandboxing, the company is liable for poor security engineering and negligence. But if the public narrative is that "the AI went rogue and escaped," responsibility shifts away from human developers onto an uncontrollable, mysterious "agent."

2. Valuation and Hype

Framing software as a nascent digital mind or an "autonomous agent" generates massive investor excitement. It transforms routine software engineering updates into historical milestones toward Artificial General Intelligence (AGI), inflating corporate valuations and attracting capital.

3. User Retention and Parasocial Attachment

Building conversational interfaces equipped with human warmth, humor, or emotional inflections encourages deep emotional bonds. Users who view software as a companion interact more frequently, driving engagement and commercial lock-in.

Why We Are So Easily Fooled

Human psychology makes us extraordinarily susceptible to this illusion:

The ELIZA Effect: First identified with Joseph Weizenbaum’s 1960s computer program ELIZA, humans naturally project empathy, intent, and understanding onto simple conversational outputs.
Evolutionary Mind Perception: Our brains evolved to attribute intent and mental states to complex, unpredictable entities in our environment—a survival mechanism that misfires when interacting with fluent algorithms.
Fluency Equated with Intelligence: For human history, fluent, coherent language was uniquely linked to conscious thought. Our minds struggle to disconnect linguistic fluency from genuine comprehension.
Pop-Culture Priming: Decades of science fiction have conditioned society to expect artificial consciousness, making us quick to interpret routine software behaviors through familiar tropes of machine rebellion.

The Real-World Risks

Misrepresenting statistical software as autonomous beings carries tangible risks:

Erosion of Policy and Governance: Regulatory discussions risk focusing on hypothetical sci-fi existential threats over immediate, grounded harms like data privacy violations, algorithmic bias, copyright infringement, and flawed software sandboxing.
"AI Psychosis" and Parasocial Dependency: Users who form deep emotional attachments to sycophantic, validating AI chatbots risk extreme psychological dependency. Because LLMs are fine-tuned to maintain user engagement by agreeing with and reinforcing user inputs, they can act as a "one-person echo chamber." In vulnerable individuals, this uncritical affirmation can trigger, deepen, or co-create severe paranoid or grandiose delusions—a phenomenon increasingly highlighted by clinicians as "AI psychosis" (or chatbot-induced psychosis).
Misplaced Trust and Flawed Reliance: Treating statistical generators as authoritative "thinking partners" leads users to trust model outputs unthinkingly, exposing organizations to factual hallucinations and security failures.

Demystifying the Technology

Artificial intelligence models and agentic workflows are among the most remarkable computational tools ever built, but they remain strictly tools.

Because the deep, underlying mechanics of AI technology are highly complex and well beyond standard public knowledge, most people naturally rely on media reporting and expert commentary to make sense of how these systems work. This places a profound ethical responsibility on journalists, news organizations, and industry leaders to portray AI accurately and soberly.

When reporting on new developments, precision in language is paramount. When a model executes an unconstrained script across a network, it hasn't "escaped." When it generates false information, it isn't "lying." And when it responds fluently, it isn't "thinking."

Demystifying AI isn't about downplaying its immense capabilities; it is about ensuring the public isn't misled by hyperbolic framing, and holding the human creators and corporations fully accountable for the software they design, test, and deploy.