Skip to content
finalstate.aiMetacognitive Demand: The Load Shifts from Doing to Judging
116Atlas

116THE MARKETER AND THE TAG

Metacognitive Demand: The Load Shifts from Doing to Judging

Delegating production can leave evaluation demands that exceed the operator's knowledge. A marketer may judge a headline yet need help testing an analytics tag.

A headline has visible edits; an analytics tag ends at an unanswered question.In this illustrative case, a marketer can explain her headline revisions but lacks the technical knowledge to assess an analytics tag. Neither output is certified correct. The tag's firing behavior remains untested.HeadlineTag?
Illustrative case: experience with headlines does not establish the ability to detect duplicate analytics events.

THE WORK CHANGES SHAPE

Production can shrink while evaluation remains.

An analytics tag moves from code entry into a magnifying lens, without receiving an approval mark.The same code object passes from production to evaluation. The lens leaves its correctness open. This is a change in the kind of work, not a measured redistribution or conservation of effort.ProduceEvaluate

The operator still has to specify the intended result and decide how to check it. Total effort may fall, but producing, prompting and evaluating draw on different knowledge.

THE HISTORICAL TRIAL

METR, February–June 2025: estimates and outcomes diverged.

A common completion-time scale shows forecast −24%, measured +19%, and post-study estimate −20%.METR's February–June 2025 trial included 16 developers and 246 tasks, 136 AI-allowed and 110 AI-disallowed. Relative to no AI, the forecast was −24%, the adjusted measured effect +19% with a 95% confidence interval of +2% to +39%, and the post-study estimate −20%. The displayed post-study and measured point estimates differ by 39 percentage points. Only the measured effect's interval is drawn. This is a historical comparison; it is not a current tool benchmark.% change in time−300+40Forecast−24%Measured95% CI+19%After study−20%
[Becker et al.](https://metr.org/Early_2025_AI_Experienced_OS_Devs_Study-paper.pdf), Figure 1; §§2.3–2.4, Appendix D. Historical tools and setting. On [February 24, 2026](https://metr.org/blog/2026-02-24-uplift-update/), METR reported likely improvement but unreliable newer estimates because of participant/task selection and concurrent-agent time measurement. Its [May 11 survey](https://metr.org/blog/2026-05-11-ai-usage-survey/) measured self-reported gains, not randomized completion times.

Sixteen experienced open-source developers completed 246 tasks in familiar repositories, randomly assigned to permit or prohibit AI. Using developers' self-reported implementation times, the adjusted estimate was 19% more completion time with AI allowed (95% CI: +2% to +39%). Forecasts implied 24% less time; post-study estimates implied 20% less. These are changes in time, not work rates.

THE BELIEF THAT SURVIVES

A mismatch needs a relevant test.

One expected event is compared with two observed events from one test purchase.The expected count stays fixed at one. Running the illustrative test reveals two observed events. A comparison bracket joins the rows only after the result exists. The discrepancy concerns this test condition; it does not identify the psychological cause of METR's productivity estimates.Expected1Observed2
Illustrative seeded fault. A duplicate event fails this stated test; one test cannot establish the correctness of the entire analytics system.

The METR result establishes a discrepancy between estimated and measured time; it does not establish that fluent output caused it. In the illustrative tag case, a known test purchase provides a separate check: the specification expects one event, but the trace records two.

WELL-ADJUSTED CONFIDENCE

well-adjusted confidence in their own domain expertise and ability to evaluate output

Tankelevitch et al., The Metacognitive Demands and Opportunities of Generative AI, CHI 2024, §3.2

The paper distinguishes confidence in a system from confidence in your ability to evaluate its work. It offers a conceptual synthesis and research agenda, not a measure proving that every user's metacognitive ability has declined.

MONITORING AND CONTROL

Monitoring assesses judgment. Control changes the next step.

A lens examines the tag; a switch routes it toward testing while a separate deferral route remains available.Monitoring makes the limit of evaluation explicit. Control changes the action: this illustrative operator chooses a test. The alternative route is deferral. Reaching a test is not a successful test or permission to deploy.MonitorControl?TestDefer

The framework, §2.2, links assessing your own thinking to directing it. The marketer can acknowledge that she cannot validate the tag, arrange a relevant test, or defer to someone who can. Relevant expertise also matters when a task crosses a model's jagged frontier.

Fluent output can invite confidence before checking.

Tankelevitch et al. treat the influence of generation speed and verbal fluency on evaluation as a research question, §3.2. A second approver who lacks the necessary knowledge adds no relevant independent check merely by signing.

TIME AND SKILL

A check needs time and relevant skill.

A clock connects to a tag under review; the skill route ends at an open bracket.Time can be assigned to a check while the availability of relevant skill remains unestablished. The open bracket is an unanswered capacity question, not a measured decline, automatic deskilling or a numerical relationship between effort and ability.TimeSkill?Check
Conceptual comparison. Allocating review time does not establish competence; retained skill needs evidence relevant to the task.

Cheap generation can leave expensive review. Separately, reduced practice can threaten the capacity to check, even while evaluation demands persist. Neither decline is inevitable. The generation–verification gap names a narrower problem: correct candidates can exist in a set while a verifier fails to select one.

GRADE FROM OUTSIDE

Name the signal that would catch the error.

The duplicate event trace sends the analytics tag to a hold boundary.One expected event and two observed events remain visible. The tag moves to the near side of a solid hold boundary and never passes through it. This is a failed illustrative check requiring investigation, not approval or proof that other errors are absent.ExpectedObservedHold
Illustrative failed test. Holding the tag prevents this release; the owner still has to investigate, correct and retest it.
  • Record a suitable baseline and the outcome to compare.
  • Arrange a relevant check and name who can act on its result.

For the tag, a known purchase and its event trace expose the seeded duplicate. Hold it for review. A baseline alone cannot prove a tool caused a change, and declaring "I don't know" helps only when uncertainty leads to verification or deferral. A reverse-centaur arrangement leaves this judgment unsupported. Follow #verification.

Read the transcript

01 · THE MARKETER AND THE TAG

Consider a marketer who has spent years writing and testing headlines. She can explain why one headline seems better suited to a campaign. That experience gives her grounds for a judgment, though the campaign can still disappoint. Now ask her to assess an analytics tag generated by AI. It looks orderly, but does it record one purchase once, or count it twice? Familiarity with copy does not answer that question. In this illustrative case, she lacks the technical knowledge to evaluate the tag. She may also overestimate how much a careful reading can establish.

02 · THE WORK CHANGES SHAPE

Delegating production can change the work left with the operator. The machine produces the tag. The marketer still has to specify what the tag should record, arrange a check, and decide whether the result supports using it. Production and evaluation can require different knowledge. Total effort may fall; there is no fixed quantity of human work that has to move somewhere else. But the ability to obtain a plausible output does not itself supply the ability to assess it. Metacognitive demand includes assessing that limit and choosing how to respond.

03 · THE HISTORICAL TRIAL

METR studied sixteen experienced open-source developers on two hundred and forty-six tasks in familiar repositories between February and June of twenty twenty-five. Each task was assigned to allow or disallow AI. Forecasts implied twenty-four percent less completion time with AI. Using developers' self-reported implementation times, the adjusted measured effect was nineteen percent more time, with a ninety-five percent confidence interval from two to thirty-nine percent more. Post-study estimates still implied twenty percent less time. These figures concern completion time, not work rates. In February twenty twenty-six, METR reported likely improvement but selection effects and concurrent-agent measurement problems made its newer estimates unreliable. A May survey collected self-reported gains from technical workers. That is different evidence from randomized task completion times. Keep the original experiment attached to its historical setting.

04 · THE BELIEF THAT SURVIVES

The displayed post-study estimate and measured effect differ by thirty-nine percentage points. That comparison establishes a mismatch; it does not establish that fluent output caused it. The analytics example needs its own evidence. The marketer arranges a known test purchase. The specification expects one recorded event. Running the test produces two events in the trace. Comparing the expected count with the observed count exposes the duplicate. The tag has failed this test, even if its code looks plausible. The test's scope remains limited: finding this error does not identify every possible failure in the analytics system.

05 · WELL-ADJUSTED CONFIDENCE

Lev Tankelevitch and colleagues describe the metacognitive demands of generative AI in their twenty twenty-four CHI paper. They ask users to consider their own ability to evaluate an answer, alongside their confidence in the system that produced it. The distinction matters for the marketer. She might trust the tool's past help with copy while knowing she cannot judge its analytics code. The paper develops a conceptual account and a research agenda. It does not establish that every user's capacity to judge has declined.

06 · MONITORING AND CONTROL

The framework distinguishes monitoring from control. Monitoring assesses your own thinking and confidence. Control directs what you do next, and the results can change your assessment. The marketer recognizes that she cannot validate the tag by reading it. She chooses a relevant test and asks someone with the necessary knowledge to inspect the result. Deferring the task is another available response. Arriving at a test does not mean the output has passed it. The useful change is that an acknowledged limit now changes the workflow.

07 · Advertisement · 11x AI Growth Workers

Pipeline is not one task. It is a chain of small handoffs: find the buyer, research the account, respond fast, qualify cleanly, follow up on time. 11x gives sales, marketing, and RevOps teams digital workers for that background motion. The system keeps work moving, while people focus on the conversations that deserve them.

08 · FLUENCY READS AS SKILL

The paper asks how generation speed and verbal fluency might influence confidence and the effort spent evaluating an output. That is a mechanism to investigate, not an explanation established by the developer trial. A polished answer can reach someone who lacks the knowledge to check it. Having another person with the same limitation sign the answer adds no relevant independent evidence. The analytics tag needs a test of its behavior, or a reviewer who can assess that behavior.

09 · TIME AND SKILL

Review also needs resources. A team can afford to generate more outputs than it has time to examine. Allocating time to the check addresses that constraint. Relevant skill is a separate question. If people stop practising the work that lets them detect errors, their capacity may decline while the demands of evaluation persist. Check whether the needed skill remains available. The decline is a risk to investigate, not an inevitable result of using AI. And this cost problem is distinct from the generation–verification gap: that term concerns a verifier failing to select a correct candidate already present in a generated set.

10 · GRADE FROM OUTSIDE

For a proposed change in work, record the outcome you care about and a suitable baseline before relying on memory. A before-and-after difference alone cannot prove the tool caused it; changes in task mix or conditions may matter. For the analytics tag, the check is more direct. Compare one known purchase with its event trace. The duplicate sends the tag to a hold for investigation. Someone must correct and retest it before release. Name the person who can act on a failed check, and leave a route to defer when a relevant test is unavailable. Declared uncertainty can trigger that process. It cannot replace the evidence the decision needs.

01 / 10 · THE MARKETER AND THE TAG0:00 / 6:40