I. We Average Constantly

We average constantly.

We average profit, productivity, risk, growth, performance, efficiency. We compress complicated systems and vast data sets into single numbers because, most of the time, it works. The underlying mechanisms are similar enough that the distortion introduced by aggregation is small, invisible, and safely ignored.

That is why averages are powerful. They let us see patterns without tracing every mechanism. They turn noise into signal and make complex systems legible.

And despite their ubiquity, paradoxes are rare.

A paradox does not appear every time we use an average. It appears only when the shortcut fails—when the things being averaged no longer respond in the same way. When one part of a system improves under an intervention while another degrades.

This is the moment when an average stops summarizing reality and starts misrepresenting it.

In those moments, the world remains consistent. It is the average that has become contradictory.

This failure mode links many famous contradictions that appear unrelated on the surface.

II. The Paradox That Shouldn’t Exist

In 1865, William Stanley Jevons noticed something unsettling.

Jevons was not hunting for contradiction. He was asking whether efficiency could preserve Britain’s industrial future in a coal-dependent economy.

As steam engines became more efficient, Britain’s consumption of coal increased. This result ran against prevailing intuition. Efficiency was supposed to mean fewer inputs for the same output. Better engines should have reduced coal use, not expanded it. Instead, improvements in efficiency coincided with rising total coal consumption.

This observation came to be known as Jevons’ Paradox. It has often been read as a warning that progress carries a hidden cost, that technological improvement can undermine its own aims.

This is not unique to coal. It reappeared across domains and decades. Lighting became cheaper and total illumination expanded. Transportation became more efficient and miles traveled increased. Automation removed human effort from routine tasks and intensified the human burden when systems failed. In each case, improvement at one level appeared to produce deterioration at another.

Jevons was not observing irrational behavior. He was observing multiple responses being combined and treated as one. Efficiency reduced coal consumption in some uses and expanded it in others. Both effects followed directly from incentives and demand. The contradiction appeared only after these responses were collapsed into a single aggregate and labeled “the effect of efficiency.”

Nothing contradictory occurred within the system itself. The contradiction emerged in the summary.

Once the components are separated, the outcome is no longer surprising. Efficiency lowers consumption where demand is already saturated and, at the same time and for the same reasons, expands it where demand remains elastic. Total coal use reflects the weighted contribution of both regimes.

Jevons’ observation demonstrates a boundary. On one side, averaging clarifies. On the other, it conceals structure. Crossing that boundary does not produce paradox in the world. It produces paradox in our description of it.

Not all contradictions vanish with better sorting. While paradoxes explored here are of a certain type, other types of paradoxes exist.

· Physical Limits: Constraints where improvement in one area necessitates the decline in another due to the laws of physics or biology.

· Zero-Sum Tradeoffs: Where different systems are in direct competition for the same resources.

In these cases, the paradox is truly a feature of the world. These examples are outside the scope of this essay.

III. What These Paradoxes Share

Not all paradoxes share the same cause. Some reflect real constraints, genuine tradeoffs, or limits that cannot be engineered away. In those cases, contradiction is not an artifact of analysis but a feature of the world itself.

The paradoxes considered here belong to a narrower class.

They are paradoxes of improvement—cases in which an intervention that works as intended at one level appears to fail when viewed in aggregate. Efficiency rises, capability improves, prevention expands, transparency increases. Yet the system-wide outcome seems to move in the opposite direction.

These cases share a common structure.

First, the population under study contains distinct regimes.
Second, each regime responds predictably to the same change, but not in the same way.
Third, the analysis treats the population as uniform.
Finally, the aggregate outcome appears contradictory.

When the effect on the whole conflicts with the effect on the parts, we call it a paradox.
It isn’t. It’s a missing category.

This pattern is a close cousin of a familiar statistical result. Simpson’s Paradox shows how aggregated data can reverse relationships that hold within every subgroup. The paradox does not arise from strange behavior in the data, but from collapsing distinct populations into a single average.

People often encounter Simpson’s Paradox as a statistical curiosity. It is better understood as a warning: when aggregation crosses a regime boundary, averages can reverse the direction of every underlying relationship.

The classic example comes from university admissions in the 1970s. Aggregate data suggested that women were admitted at lower rates than men. Yet when admissions were examined by department, women were admitted at equal or higher rates in nearly every program. The apparent bias emerged only after applications to programs with very different acceptance rates were averaged together.

Nothing contradictory occurred in the system. Each department behaved consistently. The contradiction was introduced by aggregation that erased the relevant categories.

The paradoxes examined here extend that same logic beyond statistics and into system design. Heterogeneous responses are combined and treated as though they belong to a single mechanism. Once that assumption is made, the aggregate outcome can point in a direction that no individual regime actually follows.

At that point, the average no longer summarizes the system. It replaces structure with contradiction.

IV. How the Paradox Forms

The paradox forms at a specific step in the analysis.

Consider a system composed of two regimes, A and B. Each responds to an intervention in a stable and predictable way, but not in the same way.

The problem arises when the system is treated as uniform. Instead of observing each regime separately, the analysis collapses their responses into a single weighted average and treats it as though it reflected one mechanism.

A contradiction appears only when the regimes respond in opposite directions—when one improves under the intervention and the other degrades. The aggregate outcome can then move in a direction that matches neither regime on its own.

Nothing contradictory occurs within the system itself. Each regime behaves exactly as its incentives and constraints predict. The contradiction is introduced by aggregation.

Once the regimes are separated, the paradox dissolves. The error lies not in the intervention or its effects, but in averaging across mechanisms that no longer belong in the same category.

V. What’s Actually Inside the “Paradoxes”

Across domains, the same structure appears.

Jevons’ Paradox

What is improved is energy efficiency.

Two regimes are being averaged together. Some uses are already saturated, where efficiency reduces total consumption. Others remain elastic, where efficiency lowers effective cost and expands demand. Efficiency pushes consumption in opposite directions depending on the use.

Total energy consumption reflects the weighted contribution of both regimes. Calling that aggregate movement “the effect of efficiency” collapses distinct responses into a single category.

Automation Paradox (Lisanne Bainbridge)

What is improved is machine capability.

Two regimes are being averaged together. In routine operation, automation displaces human labor and attention. In rare or novel failures, automation makes human intervention indispensable. The human role shrinks most of the time and becomes critical when it matters most.

Averaging across system states produces the appearance that automation both removes and intensifies the need for human operators.

Prevention Paradox (Geoffrey Rose)

What is improved is preventive intervention.

Two regimes are being averaged together. High-risk individuals receive large individual benefit. Low-risk individuals receive small individual benefit but dominate population outcomes by sheer number.

The intervention works in both regimes. The apparent contradiction arises only when individual and population effects are collapsed into a single assessment.

Productivity Paradox (Robert Solow)

What is improved is information technology.

Two regimes are being averaged together. Some tasks are transformed by technology and see large productivity gains. Others absorb technology as coordination overhead, increasing complexity without proportional output.

Averaging across task types and calling the result “the economy” produces the illusion that productivity stalled.

Transparency Paradox

What is improved is visibility.

Two regimes are being averaged together. In routine execution, transparency improves accountability and performance. In learning and experimentation, transparency suppresses risk-taking and degrades outcomes.

Visibility helps one regime and harms the other. Aggregating across both produces the appearance that transparency undermines performance itself.

--

In every case, nothing contradictory happens inside the system. Each regime responds predictably to the intervention applied to it.

The contradiction appears only after distinct regimes are collapsed into a single average and treated as though they belong together.

VI. Sort First, Optimize Second

The corrective move is direct but non-obvious.

Before optimizing a system, sort it.

When paradoxes of improvement appear, the failure is not in the intervention itself but in how responses are grouped. The system is treated as uniform when it is not. Optimization is applied before classification, and the result is contradiction.

The remedy reverses the order.

First, identify the dimension along which behavior differs. Determine where responses diverge rather than assuming they align.

Second, name the regimes explicitly. Make clear which cases belong together and which do not.

Third, specify what works for each regime on its own terms. An intervention that improves performance in one context may degrade it in another without being flawed in either.

Finally, route cases accordingly. Apply the appropriate mechanism to each regime rather than forcing a single solution across heterogeneous responses.

The categories will vary by domain. The mistake does not.

Whenever improvement appears to undermine itself, the first question is not whether the intervention failed. It is whether the system was sorted before it was optimized.

VII. Why This Pattern Matters Now

In earlier cases, this failure mode mostly produced confusion. Economists argued about efficiency. Policymakers debated tradeoffs. The cost of misclassification was conceptual rather than operational.

The stakes have changed.

Modern systems increasingly depend on continuous optimization at scale. When classification errors persist in those environments, they do not merely mislead analysis. They break incentives, collapse coordination, and force institutions into unstable responses.

Human–AI collaboration is a clear example that is presently being researched and debated.

AI systems are widely expected to improve performance by automating routine work while keeping humans in supervisory roles. As models become more capable, the expectation is that oversight becomes easier rather than harder.

Recent formal models suggest the opposite. As AI reliability improves, the economic conditions required to sustain human oversight deteriorate. The probability of error declines, but the cost of maintaining attention, verification, and responsibility does not decline in parallel. Oversight becomes increasingly expensive relative to the failures it prevents.

Organizations respond predictably. Some abandon human–AI collaboration altogether. Others deploy less capable systems to justify continued oversight. Still others remove humans from the loop entirely and accept unmanaged risk.

This pattern is often described as a “human–AI contracting paradox.” Improvement appears to undermine supervision. Greater reliability seems to make the safety provided by human review harder to sustain.

What matters here is not the novelty of the result, but its structure. The same classification failure that once produced interpretive paradoxes now produces institutional failure. When aggregation mistakes migrate from analysis into system design, the cost is no longer confusion. It is collapse.

VIII. The Same Error, Resolved

The human–AI paradox forms in exactly the same manner as the earlier cases.

It arises because two different kinds of outputs are treated as though they belong to the same category.

Some AI outputs are execution-stable. Once validated, they can be used without continuous human inspection. Others are inspection-sensitive. They require active oversight because errors remain costly, ambiguous, or difficult to detect automatically.

When these outputs are averaged together, oversight incentives are tied to the wrong quantity.

Imagine two separate AI systems with a combined overall error rate of 1%. System A, accounts for 90% of the combined volume and has a negligible error rate. In system B, 90% of outputs are flawless and 10% are deeply defective but it accounts for only 10% of the combined volume. The overall average error rate is 1%. If an inspector is required to review 100% of the combined volume of Systems A and B with an error rate of 1%, wages soar. However, if the inspector only reviews the outputs of System B, the wages required to maintain vigilance scales with that concentrated risk, not the total overall risk of Systems A and B.

This formalization captures the effect.

Let p denote the unconditional probability that an AI system produces an error. Let ρ denote the share of outputs that are classified as execution-stable. The relevant error rate for inspection is not p itself, but the conditional error rate over the remaining inspection-sensitive outputs:

p′ = p / (1 − ρ)

The intuition is straightforward. As AI improves, p declines—fewer overall errors. But as classification improves, ρ rises—more outputs become execution-stable and exit the inspection pool. The remaining inspection-sensitive outputs now represent a higher share of the work that still requires checking.

An inspector reviewing all outputs at 1% average error rate faces wages that scale with total volume. An inspector reviewing only the 10% flagged as inspection-sensitive, those where the conditional error rate is 10%, instead faces wages that scale with concentrated risk, not volume. The inspector faces fewer items, but each carries proportionally more risk. The wage required to sustain vigilance tracks p′, not p.

Here’s the mechanism in concrete terms: Imagine two AI systems with a combined error rate of 1%.

System A produces 90% of volume with negligible errors, these outputs are execution-stable. System B produces 10% of volume where 10% of outputs are deeply defective and are considered inspection-sensitive. If an inspector must review 100% of combined volume at 1% error rate, wages soar as volume scales. If the inspector reviews only System B’s outputs, wages scale with that concentrated 10% conditional risk, not the 1% average. The oversight remains viable because it’s no longer spread across outputs that don’t need it.

The paradox disappears once we stop treating all AI outputs as inspection-eligible.

In practical terms, removing work that no longer needs checking from the inspection pool keeps oversight viable instead of making it prohibitively expensive.

A Limitation

This separation is easier to describe than it is to implement. Execution-stable and inspection-sensitive outputs can be challenging to distinguish. The boundaries blur: an output that is stable in context becomes sensitive in another or a reliable pattern fails under distribution shifts.

However, even this does not regress entirely to inspecting every input as the blurriness of the boundaries is also specialized work. There is still no one inspecting every input, there are now multiple specialists (or the same with additional training) to review the currently flagged inspection-sensitive inputs and determine which inspection-sensitive inputs are being treated as execution-stable.

IX. When the Average Stops Working

Aggregation is not the enemy.

We average because it is efficient, legible, and usually accurate. In many systems, the underlying mechanisms are similar enough that the distortion introduced by aggregation is small and safely ignored. The shortcut holds. The average summarizes reality instead of obscuring it.

Paradoxes appear only when that condition breaks.

They emerge when systems cross a threshold of heterogeneity—when the mechanisms being averaged diverge far enough that a single number no longer describes a single response. At that point, the average stops functioning as a summary and begins to generate contradiction.

Nothing about the system itself has changed. The world remains consistent. What fails is the category used to describe it.

This pattern recurs across automation, prevention, productivity, transparency, and human–AI collaboration. These are not warnings about improvement itself. They are signals that the categories we rely on have outlived the mechanisms they were meant to represent.

Paradoxes do not appear when systems behave strangely.
They appear when we continue to average long after the things being averaged have stopped behaving alike.

For much of the nineteenth and twentieth centuries, the cost of this mistake was largely absorbed. Misreadings produced confusion, debate, and delayed adjustment, but the systems themselves remained stable enough to tolerate the error.

That is no longer true.

Modern systems operate through continuous optimization at scale. When categorization lags behind improvement, the resulting error does not remain confined to analysis. It propagates into incentives, coordination, and institutional design. Optimization is applied to the wrong object, oversight is routed to the wrong places, and systems are forced into brittle equilibria that grow more complex and costly over time.

The paradoxes of improvement have not changed. What has changed is the price of misclassification.

When improvement appears to undermine itself, the remedy is not to slow progress or distrust optimization. It is to recognize when aggregation has crossed a boundary it can no longer safely span—and to stop treating different mechanisms as though they belong to the same average.