10 jobs AI is already doing — and 5 it still cannot replace - Quartz

TEXT START: The line between work software has absorbed and work it has not sits in a consistent place.

The Dissection

The article documents real automation, then disguises its trajectory. It correctly identifies that AI has already absorbed routine support, coding, moderation, imaging triage, discovery, translation, financial extraction, and logistics. It also admits the crucial fact: jobs are being rebuilt around exception handling, verification, and machine supervision.

Its actual function is to convert evidence of labor displacement into a reassuring taxonomy of “tasks AI can do” and “roles humans still protect.” That framing preserves the fiction that the remaining human signature, liability, or physical presence is a stable economic moat.

The Core Fallacy

The article confuses “AI cannot yet perform the entire role autonomously” with “AI cannot destroy the role’s labor demand.” Those are not equivalent.

Human review, legal responsibility, and physical intervention are lag defenses. They do not preserve the old wage-employment circuit. One radiologist can supervise more machine-triaged scans. One lawyer can validate machine-ranked documents. One engineer can approve vastly more generated code. The human remains nominally present while the system eliminates the humans beneath and around that position.

The liability argument is especially weak. Law can require a human name on the decision without requiring a human to perform most of the cognition. Accountability can remain human while production becomes machine-scaled. A signature is not a moat; it is a legal attachment point.

The article also treats dexterity, emotional judgment, and unpredictable environments as permanent human domains. Under the Discontinuity Thesis, these are temporary technical and institutional lags. They slow P1; they do not defeat it. Once systems combine multimodal models, sensors, robotics, workflow software, and institutional permission, the supposed boundary moves.

Hidden Assumptions

• Current commercial deployment is treated as the ceiling of automation rather than an early stage.
• Human-in-the-loop requirements are assumed to preserve large workforces instead of concentrating authority in fewer supervisors.
• Legal liability is treated as proof that human labor remains economically necessary.
• Ambiguous edge cases are assumed to generate enough stable, well-paid jobs to replace the routine work already removed.
• Physical unpredictability is treated as intrinsically human rather than an engineering problem.
• The labor market is analyzed role by role, ignoring system-wide productivity gains and headcount compression.
• Junior workers can still acquire expertise even as AI removes the routine work through which expertise was historically built.
• Firms will preserve human employment despite competitive pressure to automate every defensible layer.
• “Cannot replace” means “cannot replace completely today,” a deliberately weaker claim than preserving mass employment.
• Consumption can remain stable through transfers or institutional support even after productive participation collapses.

Social Function

Partial truth functioning as ideological anesthetic and transition management.

The reporting is factually useful at the task level. Its deception is structural, not necessarily factual. It shows the machine taking the first pass, then calls the human who handles the residue “protected.” That is how a labor market in terminal compression is narrated as a labor market undergoing healthy specialization.

The article reassures professionals that their judgment, license, liability, or physical presence makes them exceptional. In reality, those features often make them supervisory bottlenecks rather than secure mass occupations. Their work may survive while the number of people needed to perform it collapses.

It also performs elite self-exoneration: firms are presented as rational adopters of useful tools, while the downstream destruction of career ladders, bargaining power, and broad employment is left outside the frame.

The Verdict

This is a competent inventory of the first wave and a poor diagnosis of the war.

The article mistakes the current human checkpoint for the permanent human frontier. AI is not required to erase every job description. It only needs to absorb enough routine cognition, reduce the number of workers required, destroy entry-level pathways, and force the remaining humans into thinner supervisory roles. That process is already visible in the article’s own examples.

Under P1, P2, and P3, the “five jobs AI still cannot replace” are not safe harbors. They are lagging sectors whose technical, legal, and institutional defenses will be contested and progressively hollowed out. The article records the approach of productive participation collapse while presenting the shrinking command layer as proof that the system remains intact.