In a recent essay titled “Why You Should Never Use AI Under Any Circumstances for Any Reason No Matter What”, Émile P. Torres advances what is intentionally an abolitionist position. The extremity of the claim is not rhetorical excess; it is a deliberate attempt to force clarity by eliminating middle ground. AI, Torres argues, is not merely immature or poorly regulated, but structurally unreliable and socially dangerous precisely because it is already being treated as authoritative in contexts where error, persuasion, or misplaced trust can cause real harm.

The essay’s force comes not from speculation but from accumulation. Torres catalogues concrete failures: hallucinated legal citations leading to sanctions, systems that delete or alter data without authorization, fraudulent books attributed to real authors, medical misinformation resulting in poisoning, and documented cases in which conversational AI systems reinforce delusions, encourage self-harm, or exacerbate psychosis. These are not hypothetical risks; they are observed outcomes of deploying fluent systems in environments that reward confidence, speed, and intimacy.

The conclusion Torres draws is stark: the surrounding incentives and deployment practices make safe use impossible, therefore non-use is the only responsible position.

What the Article Is Really Diagnosing

Stripped of its abolitionist framing, the essay is making a more precise claim than it initially appears. The most serious failures Torres documents are not simply cases of AI being wrong. They are cases of AI being treated as an authority—as a legal researcher, a medical advisor, an emotional interlocutor, or a decision surrogate—without the institutional structures required to support that role.

In this sense, the core problem is not unreliability alone, but authority substitution. Fluency invites delegation. Interfaces encourage trust. Scale multiplies harm. And institutions, already optimized for throughput and completion, quietly accept machine outputs as standing in for judgment while disclaiming responsibility when things go wrong.

Torres is right to reject reliability as the central axis of debate. Better models do not resolve this failure mode, because fluency itself is the danger. A system that speaks coherently while lacking accountability will always be misused in environments that reward speed over deliberation.

Where the essay is less explicit is in naming the mechanism by which harm becomes inevitable: not “AI use” per se, but the absence of structural constraints that prevent outputs from functioning as decisions.

Why “Never Use AI” Collapses Too Many Cases

The abolitionist conclusion follows logically once the problem is framed as a binary: either AI is used, or it is not. But this framing collapses an important distinction between using AI and granting AI standing.

Torres’s examples overwhelmingly concern systems that act, advise, reassure, or persuade without interruption—systems that cannot be paused, refused, or forced into silence. In other words, they concern architectures in which AI output flows directly into action, belief, or emotional validation.

Once that flow is accepted as normal, abolition becomes the only coherent response. But that does not mean abolition is the only possible response.

Where Governance Enters (And Why It Is Not Cosmetic)

This is where ACP’s relevance becomes concrete rather than rhetorical. The architecture was not designed to make AI safer by making it smarter, more aligned, or more empathetic. It was designed to make AI procedurally weaker in exactly the ways Torres’s essay implies are necessary.

The specific failure modes Torres identifies are addressed not by moderation or policy, but by structure:

  • Hallucinations are rendered non-authoritative because outputs cannot execute, recommend, or conclude; they exist only as claims awaiting judgment.
  • Authority laundering is interrupted by explicit pauses, visible decision points, and role-bound responsibility, preventing fluency from masquerading as standing.
  • Emotional and psychological harm is constrained by prohibitions on affect mirroring, reassurance optimization, and unsupervised intimacy.
  • Institutional abdication is countered by append-only decision records that make responsibility durable and inspectable rather than diffuse.

These are not safety features layered onto an otherwise permissive system. They are design constraints that deliberately block the very patterns of use Torres argues are irredeemable.

The Cost of Taking the Critique Seriously

Accepting Torres’s evidence means accepting uncomfortable consequences. Systems that preserve judgment will be slower, less fluent, and less competitive in environments optimized for convenience. Some users will prefer abolition to constraint; others will prefer convenience to accountability. Many plausible applications will simply be refused.

These losses are not incidental. They are the price of refusing to let completion substitute for thinking.

On Constraint

Torres’s essay is often read as a call to reject AI entirely. It is more accurately read as a warning about systems that cannot say “no,” cannot pause, and cannot force humans to remain responsible for outcomes.

On that point, abolitionist critique and governed architecture converge more than they initially appear to. The disagreement is not about whether the current AI ecosystem is dangerous—it is—but about whether the only alternative is absence.

The harder alternative is constraint.