What AI Can and Can't Do: An Interview with Brendan Fleig-Goldstein

TEXT START: Brendan Fleig-Goldstein is a philosophy professor and Mellon Postdoctoral Fellow at Brown University’s Cogut Institute for the Humanities.

The Dissection

The interview accurately explains current LLM mechanics and their present weakness at verification. Then it quietly converts those limitations into a social conclusion: AI will redistribute tasks, humans will retain judgment, and humanities skills will remain valuable.

That is the central maneuver. The interview analyzes whether AI can perform a task perfectly, but the economic question is whether it can perform that task cheaply and well enough for firms to need fewer people. It examines workflow substitution at the individual level while avoiding labor-demand collapse at the system level.

The Core Fallacy

It mistakes “not reliably correct” for “not economically substitutable.” A system does not need flawless judgment to displace labor. It only needs to produce usable output cheaply enough that one human can supervise more work, or that institutions can tolerate, test, and distribute its errors.

“Human evaluation” is not a permanent moat. Evaluation can be decomposed, benchmarked, cross-checked, delegated to multiple systems, tested against outcomes, or reduced to a narrow approval bottleneck. If production scales while the number of required evaluators falls, the remaining human judgment may survive as a function while its mass employment value dies.

The claim that AI changes tasks rather than replacing people is also evasive. Task redistribution is often the mechanism of replacement: the machine absorbs enough tasks that fewer workers are required. Under P1–P3, cognitive superiority, institutional adoption, and shrinking labor demand sever the employment-to-wage-to-consumption circuit without requiring magical superintelligence.

The interview’s “either superintelligence or market correction” framing is a false binary. Partial automation can destroy bargaining power long before artificial general intelligence arrives. A market correction and labor displacement can also reinforce each other rather than serve as alternatives.

Hidden Assumptions

• Human oversight will remain economically valuable at mass scale rather than becoming a thin supervisory bottleneck.
• Judgment, interpretation, and critical thinking are inherently human rather than capabilities that can be approximated, automated, or institutionally routinized.
• Productivity gains will create enough new human work to offset labor displacement.
• Workers can individually “adapt” faster than firms can reorganize production around AI ownership.
• The market price of a skill will remain high even as AI makes that skill abundant.
• Current model limitations are durable barriers rather than temporary engineering constraints.
• The important question is how people use AI, not who owns the systems and captures the resulting surplus.

Social Function

This is a partial truth functioning as ideological anesthetic, prestige defense, and transition management. The technical cautions make the argument sound sober, while the assurances about humanities skills protect the professional identity of the speaker’s audience.

Its practical message—use AI for clarification, but retain responsibility for judgment—is valid as operating advice. Its economic message is much weaker. It shifts responsibility for survival onto individuals while leaving ownership, bargaining power, labor demand, and distribution almost entirely unexamined. “Learn to use the tool” is the servitor’s adaptation script; it does not make the worker sovereign.

The Verdict

The interview is competent on what current LLMs do and inadequate on what automation does to the economic system. It promotes human judgment from a present limitation into a permanent employment moat. Under the Discontinuity Thesis, that moat is hospice care: useful for delaying the diagnosis, not preventing the death of mass productive participation.