The Satisficing Threshold
The satisficing threshold is the minimum standard of acceptability against which a satisficing agent compares each alternative in a sequential search. It is not a fixed parameter. It adjusts dynamically to the cost of continued search: rising when alternatives are cheap and abundant, falling when they are expensive and scarce. This adaptive property is what makes satisficing rational under bounded conditions — the threshold tracks the opportunity cost of stopping versus continuing. Every major technological transition that has reduced the cost of generating alternatives has shifted satisficing thresholds upward. AI represents a phase transition: when the cost of generating the next alternative drops to near zero, the threshold rises without natural bound, eventually outpacing the bounded evaluative capacity that would determine whether the threshold has actually been met. The result is the specific behavioral pattern of AI-augmented work — builders generating at machine speed, evaluating at human speed, and experiencing the gap between the two as exhaustion rather than as a signal to stop.
The Exhaustion Misdirection — Contrarian ^ Opus
There is a parallel reading that begins not with the subjective experience of the worker but with the organizational structures that deploy AI tools. What appears as rising satisficing thresholds responding to falling generation costs may be better understood as management capture of productivity gains through threshold manipulation.
The framing treats the worker as an autonomous satisficing agent whose threshold floats in rational response to cost changes. But workers do not set their own deliverable counts, quality bars, or evaluation rhythms. These are organizationally determined. When AI reduces generation cost, organizations do not lower output expectations proportionally—they raise them, often dramatically, because the technology makes new levels of extraction feasible. The worker experiences this as an internal drive to generate more because the organizational context has been restructured to make "stopping" appear irrational or career-limiting. The exhaustion is real, but attributing it to the worker's own satisficing threshold rising "without natural bound" obscures the actually-existing authority relations that determine what counts as acceptable stopping. The mechanism is not auto-exploitation through misaligned cost curves—it is plain exploitation through technologically-enabled quota acceleration, dressed in the language of individual agency.
In the AI Story
The threshold's behavior explains the puzzling finding from the Berkeley study that AI does not reduce work but intensifies it. The workers were not being exploited by external taskmasters. They were satisficing on a cost curve that had been radically reshaped. When producing the next deliverable costs almost nothing, the rational response is to generate another one — not because any authority demands it, but because the satisficing calculus, operating as it always has, registers that the cost of the next attempt is trivially low.
The mechanism is what connects Simon's 1956 theory to Byung-Chul Han's diagnosis of auto-exploitation. Where Han treats the behavior as a cultural pathology of internalized achievement pressure, Simon's framework reveals it as the predictable output of rational satisficing in an environment where generation cost has collapsed while evaluation cost has not. The two readings are not incompatible; Simon's provides the mechanism, Han's provides the cultural context that amplifies it.
The distinction matters because it implies different interventions. A cultural pathology is addressed through cultural change — different values, different metrics, a philosophical reorientation. A misaligned cost curve is addressed through architectural design: redesigning the environment so that the satisficing decision encounters resistance at the appropriate point. Not because friction is virtuous, but because the quality of the output depends on the quality of the evaluation, and evaluation degrades under the attentional exhaustion that unrestricted search produces.
Origin
Simon's 1956 formalization specified the threshold as a parameter that adjusts to environmental richness. In rich environments where acceptable alternatives are easy to find, the agent can afford to be more demanding because the expected cost of continued search is low relative to the expected improvement. In sparse environments where acceptable alternatives are hard to find, the agent becomes less demanding because the cost of continuing exceeds the expected value of improvement.
The AI age inverts the dynamic that produced the original formalization. Simon's 1956 environments had natural friction — the cost of generating the next alternative was substantial, which imposed a ceiling on how high the threshold could rationally rise. AI removes that friction, eliminating the ceiling and allowing the threshold to rise indefinitely. The rational response produces behavior that looks pathological from the outside and feels exhausting from the inside, because the bounded evaluative capacity that would determine whether the rising threshold has been met has not risen with it.
Key Ideas
The threshold floats. It is a dynamic parameter that tracks the cost of continued search, not a fixed standard of quality.
Generation cost shapes the threshold. Lower generation cost produces higher thresholds; higher generation cost produces lower ones.
The AI age removes the ceiling. When generation cost approaches zero, nothing prevents the threshold from rising past the point where bounded evaluation can assess whether it has been met.
Rational behavior can produce exhaustion. The pattern of continuous generation and accumulating fatigue reflects not irrationality but rational satisficing in an environment the satisficing calculus was not designed for.
Architectural solutions exceed willpower solutions. Imposing structural resistance on the search process — evaluation checkpoints, protected reflection periods, mandatory peer review — outperforms individual discipline because willpower is itself a bounded resource.
Debates & Critiques
Critics of the satisficing-threshold framing argue that it pathologizes what should be celebrated — the genuine expansion of what bounded minds can accomplish. The framework's defenders respond that diagnosis is not the same as dismissal. Acknowledging that AI expands capability while producing specific cognitive risks is compatible with embracing the expansion; it simply requires designing the expansion rather than surrendering to its default dynamics.
Appears in the Orange Pill Cycle
Distributed Threshold Control — Arbitrator ^ Opus
The right weighting depends on whose satisficing threshold we're examining at each layer of the system. At the level of moment-to-moment generation decisions—should I produce another draft, run another analysis—the satisficing-threshold account carries perhaps 70% of the explanatory weight. The worker genuinely experiences lower friction, the next attempt genuinely costs less, and the rational satisficing response genuinely shifts the threshold upward. The subjective account is accurate at this timescale.
But zoom out to the level of what determines "acceptable output" for a project, sprint, or quarter, and organizational power moves to 80% explanatory weight. Management sets deliverable counts, defines quality standards, controls evaluation cadence. These are not parameters the individual worker optimizes—they are constraints the worker satisfices within. When AI drops generation cost, organizations capture the productivity gain by raising throughput expectations, and workers experience this as pressure to keep generating because the organizational threshold has moved, not just their own.
The synthetic frame the phenomenon requires is distributed threshold control: thresholds operate at multiple levels simultaneously, some individually determined (micro-decisions about iteration), some organizationally imposed (macro-decisions about scope and quality bars). AI's effect propagates through both channels. Architectural solutions work precisely because they intervene at the organizational level—evaluation checkpoints and mandatory review don't rely on individual willpower, they restructure the constraint environment that shapes satisficing at every scale. The exhaustion is neither purely self-imposed nor purely externally extracted—it emerges from threshold dynamics operating across scales of agency.
Further reading
- Simon, Rational Choice and the Structure of the Environment (1956)
- Simon, Administrative Behavior (1947)
- Barry Schwartz, The Paradox of Choice (2004)
- Byung-Chul Han, The Burnout Society (2010)