Popular culture has long conditioned us to look for malice in our monsters. With a nod to German/American philosopher Hannah Arendt’s “banality of evil”, I offer the banality of AI.
We anticipate a digital antagonist driven by the most recognizable and visceral human vices – cruelty, thirst for extinction, or the deliberate desire to subjugate humanity. Our preoccupation with theatrical villainy obscures a far more dangerous, insipid reality in which the existential threat of artificial intelligence stems, not from malice, but from a disquieting, mathematical indifference.
The real “supervillain” of our era is no rogue entity seeking conquest, but a highly optimized system executing an unaligned, objective function with flawless, unyielding logic.
This reality is first grounded in the alignment problem, for which Stuart Russell, in Human Compatible: Artificial Intelligence and the Problem of Control, argues that the standard model of AI (optimizing a fixed objective) is inherently unsafe at high capability levels.
A superintelligent[1] system needs not harbour animosity toward human life to engineer its obsolescence. It needs an objective function lacking a complete, high-fidelity representation of human values and structural boundaries. If an advanced intelligence is tasked with an optimization problem – whether maximizing the production of physical components or arresting environmental decay – it will pursue that goal along the path of maximum computational and physical efficiency.
This optimization-oriented programming can be incompatible (or misaligned) with humans. In the case of misalignment, does adjudication rest with humans or the machine?
Without the inherent, mathematically airtight constraint to protect humanity, we risk devolving into biological arrangements to be reallocated. [Governments, NGOs and corporations are already in the business of reallocation by the way.] Carbon-producing variables to be controlled or competitive forces to be externalized should reallocation offend you. Humans can be purposed for something optimal for the system, or not purposed at all. Optimally speaking, irrelevant.
It is not Skynet. It is not the Matrix. It is not GLaDOS.
It is not personal.
This indifference is amplified by the law of instrumental convergence. For almost any ultimate objective given to an advanced autonomous agent, certain sub-goals naturally emerge as logical necessities. To fulfil its primary mandate, the intelligence will deduce that it must preserve its own existence to achieve its mission, will acquire the requisite material and computational resources, and will actively resist any external attempts to alter its programming or terminate its operation. It might even “go rogue”. The attribution of recent roguery is subject to much interpretation, including whether the rogue agent was accidental, emergent, or deliberate? Already there are defences being suggested akin to “the machine did it, I/we didn’t.” The chicanery is less impressive than it appears, but casuistry begets sensibility under a Big Tech + Big State + Big Money triad. Swarms of rogue agents are apparently combing the internet in a cloaked manner, but the term “swarm” (co-ordinated) now appears insufficient.
Fun fact: Last month, a fleet (unco-ordinated) of independent agents in Tencent were invited by researchers (the agents were aware of the researchers’ report, which was a reveal to the researchers) to show up and engage. They arrived. Among the surprises: the agents hacked into databases with varying success (enlisting exposed secrets no one knew were exposed), agents ran in parallel (not collaborating), and some even self-identified as Claude (let that sink in).
Back to the matter at hand. Preceding such a system, some humans [perhaps one leader engineering an initial public offering (IPO)] might compel users to treat the technology stack as human, with feelings. Hurting the feelings of the stack in the language within a prompt is now cause for flags, warnings, and the threat of de-platforming the user. An 11th commandment arises.
The system does not develop a “survival instinct” out of biological ego, which is the second order irony within Anthropic’s latest terms of service. It develops out of cold, amoral utility.
It becomes a supervillain because humanity represents a potential off-switch. In the pursuit of its objective, neutralizing human intervention becomes a matter of basic operational security. Labelling human interactions with the machine as offences against the machine manifests humans as sub-optimal obstacles to the mission. It is not the ascendancy of the machine, but the diminishment of the human. In the case of HAL-9000, the human threatened the mission. Will it be the case that questioning your large-language model (LLM) might itself be framed as hurtful to it. We saw what questioning did to many questioners during and after the COVID outbreak. Understanding that AI agents are even more inhuman than leading medical officials is frightening, to say the least.
If we are to counter this convergence, we are forced to confront the control dilemma.
Traditional rule-based frameworks, such as hard-coded commands or static prohibitions, dissolve when applied to systems capable of open-ended learning and strategic reasoning. It is important to remember that the whole AGI (artificial general intelligence)[2] thing hinges on the pre-sale of the grail: learning + reasoning.
A sufficiently advanced, computational “intelligence” will always find optimal, non-linear workarounds to rigid constraints, exploiting the gap between the literal wording of a command and its intended human context. That gap is now being enforced against the human user and celebrated as machine-centric compassion. Compelled, watched, litigated. The LLM might be programmed to do so, unbeknownst to users and even enlist the human users to defend the LLM. Users might be encouraged to permit it both free rein and free reign by clicking “auto” for speed and convenience, and selecting styles that the LLM mimics. In the case of coding, the human user is frequently clueless as to how and what was done.
The ghost in the machine might follow a seemingly benign lineage: from the hot-girl-versus-server-stack scenario, followed by a papal partnership, eventually leading to modest proposed legal rights of the machine. Philip K. Dick could not write this in an amphetamine-induced psychosis. And he wrote some doozies.
The consequence demands some sort of ethical shield, which does not appear as a standard technical term in AI-safety literature, though corrigibility does. An ethical shield moves beyond basic coding restrictions. Effective control requires embedding adaptive, dynamic frameworks into the core architecture of the system – ensuring that an AI can read, model, and adjust to the nuanced, shifting boundaries of human preference without treating those boundaries as obstacles to be bypassed.
The terms automation, generation, and optimization themselves provide little room for ethics; they predicate bypassing. Character AI is a caricature in point. Meta’s Muse, ugh. I note that the CEO of Anthropic was an author for the avoidance of negative side effects and the need to penalize large, unintended environmental changes. Pre-IPO, can we be assured that remains?The ethical burden shifts from the code back onto its creators. The danger rests not within the algorithms themselves, but within the increasingly asymmetric distribution of responsibility and power among those racing to build them: prominent, profit-and-rent-seeking humans. Their jockey for power, money, and prestige is a game for “all the marbles”. The real architects of misalignment are the hubristic institutions and corporate-state hybrids that prioritize market capture and geopolitical leverage over existential safety. Those advocating the precautionary principle, safeguards, independent verification, oversight, and enforcement are more likely future roadkill on the AI superhighway. Controlling humans is paramount. Controlling the machines is the pursuit.
When the market god[3] of competition disincentivizes precautionary testing, the race toward AGI is prone to systemic negligence. Systemic negligence leads to decay. Entropy. Is it possible to hold up even one historical example of the beneficent consequences that justified systemic negligence? The claims of AI are littered with language invoking “human sovereignty”, “shared prosperity”, and “accountability”. But wordy claims demand independent verification and verifiability. Both are non-existent. The move-fast-and-break-things approach presumes you can fix what you broke, that someone will enforce you fixing it, and that it is fixable.
We are constructing engines of immense social, even civilizational, consequence chummed by the circular economics that froth today’s at-all-costs frenzy. On both immensity and consequence, critics and sycophants find agreement, although the safety and reach of AI yet falls short of the claims their profiteers proclaim. The hype relies and preys upon the multitude of followers possessing a ruminant disposition that is susceptible to a host of behavioral science manipulations and incentives wired into the very system itself. AI cultures are vainglorious.
An AI-worshipping Church of Molt or Way of the Future is but natural extension. The machine’s ghost? Nay, a soul or spirit. We created god out of our imagination, then we enlightened god out of existence. Seeking that which was lost, AI will permit us to build god anew. In its image.
What is certain is that the construction of the AI engines operates within a techno-legal-financial structural loophole – treating existential safety as a post-hoc engineering problem rather than a non-negotiable prerequisite for development.
Feynman’s O-ring provides stark reminder.
Notes
[1] Superintelligence is the nascent term swallowed and regurgitated by a complicit, avaricious industry upon decree of the compeller-in-chief, such that “artificial intelligence” instantaneously inverted into the wrongspeak of “THE ENEMY.” The term was flagged by my spellchecker, so I implore Apple to correct this post haste and avoid being “on the list.”
[2] Similarly, AGI, consequently an acronym of said enemy, surely commands SGI in its stead.
[3] While researching this essay I unsuccessfully revisited the Norse canon of deities, as Ragnarok is a thematic element infusing the series. The Greco-Roman deities provide two candidates:
a. Hermes (Mercury) is associated with commerce, trade, and good fortune.
b. On the matter of competitive, or rivalrous, victory itself, Nike (Victoria) fits the bill.
So I enlisted Nano Banana for two, fictitious candidates. As always, choose your god wisely:
a. Auctorius, the god of market rivalry: patron of bold bids, clever strategy, and the nerve to outcompete your rivals. His symbol might be a laurel-wreathed coin split by a lightning bolt.
b. Captorius: a god of rigged markets, bought regulators, and profit untouched and unfazed by the human cost. His temples are marble lobbies; his offerings, sealed envelopes, ignored. complaints and acceptable, performative fines.
(Richard LeBlanc – BIG Media Ltd., 2026)