Technology & judgment
When AI makes answers cheap, why does judgment become expensive?
Faster generation is not necessarily faster completion. When output outruns verification, the bottleneck moves from producing an answer to earning confidence in it.
A draft is not a delivery
Suppose someone replaces a two-hour writing session with a ten-minute AI draft. The apparent gain is dramatic. But checking references, revisiting assumptions and repairing fluent reasoning gaps may consume much of the saved time. These are hypothetical timings, not experimental results.
Work includes generation, verification, integration and responsibility for consequences. AI’s most visible gain often occurs in the first stage; polished prose does not automatically remove the others.
The useful comparison is completion time at a fixed confidence standard. Otherwise, a speed improvement may simply move verification outside the measured window.
Capability does not form a smooth frontier
A Harvard Business School account of a study involving 758 consultants describes a jagged technological frontier: AI’s usefulness varies across tasks.[1] The important point here is that apparently similar work can fall on different sides of a capability boundary.
A successful email does not establish the factual reliability of a contract summary. Running code does not establish that all business constraints are satisfied. Similar output formats can invite unjustified confidence across tasks.
The study concerns particular models, participants and tasks. It is not a performance table for every current system. It supports a more modest working rule: validate at the task level rather than trusting a product name wholesale.
Verification requires knowledge of what to verify
Proofreading and validating a conclusion are different activities. Validation may require reconstructing the reasoning: whom the data describe, whether comparisons share a baseline, and which omitted conditions could reverse the conclusion.
Without domain knowledge, a complete and confident answer may reduce vigilance. Fluency is not necessarily wrong; it is simply a credibility cue that does not establish truth.
A CHI 2025 survey involving Microsoft Research collected accounts from 319 knowledge workers. It reported an association between confidence in AI and less critical thinking, and described a shift toward verification and integration.[2] These are self-reports and associations, not proof that AI causally reduces intelligence.
Can the result be checked independently?
Two useful questions are whether a result can be tested independently of its generator, and whether a mistake can be reversed.
A conversion script can be checked against known inputs, output files and edge cases. An explanation of market behavior may sound coherent while making an invalid causal claim. Agreement from another model is not automatically independent evidence.
Testable and reversible work permits more automation. Work with weak tests and difficult-to-reverse consequences calls for stronger constraints and judgment. This is an allocation heuristic, not a formal risk-classification standard.
More output can create more work
When a report becomes cheap to generate, an organization may request more reports, variants and updates. The per-document cost falls while the number of items requiring responsible review rises.
If each item still needs substantive checking, the bottleneck moves to reviewers. More documentation can coexist with slower decisions. This is a conditional argument: it applies when output grows faster than verification capacity and meaningful scrutiny is necessary. It is not a claim about every use of generated content.
Possible responses include limiting approval-bound output, attaching evidence to consequential claims, showing changes rather than repeated documents, and permitting uncertainty instead of demanding another confident answer.
The next product opportunity is cheaper proof
A product optimized only for generating more may leave the hardest work to its user. Another route makes results easier to inspect: connect claims to specific evidence, separate facts from inferences, expose changes, support recovery and represent uncertainty honestly.
A coding assistant should report the scope of its changes and the checks actually run. A research assistant should make it possible to see whether a source directly supports a sentence. Long bibliographies and large test counts are not substitutes for relevant coverage.
Expensive judgment is not an argument against automation. It is a reminder that abundant expression increases the importance of knowing why this output deserves confidence. Products that reduce that proof burden can turn generation into deliverable work.
Sources & scope
Research findings are distinguished from the article’s arguments. Illustrations are not DeconstructAlpha experimental results.
京公网安备11010802050788号