Dangers of AI: 14 real risks and what you can do about them - Quartz

TEXT START: A cloned voice on the phone and a chatbot that invents court cases sit on the same spectrum as the scenarios that worry the people who built modern AI.

THE DISSECTION

This is a catalog of incidents disguised as a theory of risk. It gathers fraud, election manipulation, sexual abuse, biased decisions, hallucinations, surveillance and job displacement into one administrative stack, then assigns each a countermeasure: verify, audit, appeal, restrict, regulate or reskill.

Its real function is to make systemic disruption legible as household procedure. A cloned voice becomes a secret family phrase. A fraudulent transfer becomes dual authorization. A biased algorithm becomes a request for written reasons. Job loss becomes skill selection. Those measures may reduce individual injuries, but they relocate the burden of adaptation onto users, workers and downstream institutions.

The NIST framework, named agencies and concrete case studies create an evidence-heavy, responsible tone. That credibility is real. The structural blindness is real too. The article places a scam, facial surveillance and the collapse of mass employment on the same continuum of “risks,” flattening the difference between an incident and a regime change.

THE CORE FALLACY

The core fallacy is risk-management substitution: treating AI’s consequences as separate hazards to be managed instead of recognizing AI as a productive system capable of severing the employment → wage → consumption circuit.

Under DT logic, P1 is the decisive mechanism. Once AI achieves durable cost and performance superiority across cognitive work, P2 prevents institutions from preserving stable human-only economic domains at scale. P3 follows: the majority lose access to economically necessary labor.

The article’s job-loss remedy—identify what AI can do and build skills around what it cannot—assumes that the remaining human tasks will be numerous, durable and defensible. Competitive pressure attacks exactly those assumptions. New jobs may be fewer, more concentrated, lower-leverage or eventually automated themselves. Aggregate job-creation forecasts do not prove that displaced workers retain income, bargaining power or productive necessity.

The article correctly identifies local harms. It fails to identify the terminal mechanism connecting them: ownership and control of AI capital. Transfers might preserve consumption, but they do not restore productive participation. Governance can reduce damage; it cannot, by itself, preserve the old economic order.

HIDDEN ASSUMPTIONS

• There will be a sufficiently large remainder of work that AI cannot perform or economically replace.
• Net job creation will translate into stable employment and income for the people displaced, rather than merely improving aggregate statistics.
• Regulators and courts can move faster than deployment and enforce rules against concentrated AI capital.
• Audits, human review and consent requirements will be substantive safeguards rather than procedural rubber stamps.
• Individuals have enough time, knowledge and institutional access to verify claims, appeal decisions and pursue remedies.
• The listed harms are separable problems rather than interacting feedback loops that amplify fraud, surveillance, distrust and power concentration.
• Provenance tools and fact-checking can preserve shared reality even as synthetic content overwhelms verification capacity.
• Uncertainty about catastrophic scenarios justifies treating them as ordinary risk-management questions rather than low-probability, system-level threats.
• The owners of models, compute, energy, logistics and maintenance are background actors instead of the decisive class in the transition.
• “What you can do” is a meaningful substitute for asking who controls deployment and who captures the surplus.

SOCIAL FUNCTION

Primary classification: partial truth and transition management.

Secondary classifications: ideological anesthetic, prestige signaling and limited elite self-exoneration.

The text is not pure copium. Its examples are concrete, and several recommendations are genuinely useful against specific harms. But it trains readers to adapt to AI deployment through personal vigilance, bureaucratic appeals and incremental regulation. It turns a power struggle into a compliance checklist.

Its quiet ideological service is to normalize the transition while distributing responsibility downward. Individuals must verify. Workers must reskill. Employers must audit. Regulators must catch up. The owners of the productive machinery remain largely invisible. The population is taught how to survive the machinery’s malfunctions, not how the machinery is reorganizing ownership and necessity.

THE VERDICT

Useful field notes for surviving individual AI harms; inadequate as a diagnosis of the transition. The article catalogues the bruises while refusing to name the machine: AI capital replacing labor, concentrating control and breaking the social circuit that made mass employment viable.

Under DT logic, this is competent incident response wrapped around structural denial. It may help someone catch a deepfake or challenge an automated decision. It cannot make a displaced majority economically necessary again.