When the AI Leaves the Tailorshop: Measuring What an LLM Advisor Leaves Behind in Complex Problem Solving

Robin Welsch

cs.HCSep 14, 2026

When the AI Leaves the Tailorshop: Measuring What an LLM Advisor Leaves Behind in Complex Problem Solving

Authors: Robin Welsch

Organizations: Aalto University, Espoo, Finland

Abstract

Complex problem solving depends on acting effectively and understanding how a system works. AI advice may support these outcomes unequally. Two preregistered experiments compared participants managing a simulated clothing factory with and without an LLM advisor. Across studies, AI-supported participants reported greater confidence and understanding with less effort. In the first study (N=200), assistance increased company value but produced no detectable prediction-accuracy difference. After withdrawal, previously supported participants outperformed controls when decisions were scored against repeating previous choices, but not default settings. Within the AI-supported group, more frequent recommendation alterations predicted better unaided performance. In the second study (N=198), AI-supported participants went bankrupt less often and showed a small knowledge advantage in the registered analysis, largely associated with remaining solvent. More frequent recommendation alterations predicted higher knowledge within the AI-supported group. Applied HAI evaluation should assess users' understanding and independent capability alongside the performance achieved with AI support.

Figures & tables

Figure 1. An LLM advisor raised performance while it was available, and left behind little more than a second unaided run would have.

200

participants ran a simulated clothing factory for twelve months with the advisor present for one group (Phase A), then twelve further months with no advisor for anyone (Phase B). Faint lines are individual firms; heavy lines indicate the mean with

\pm 1

standard error of the mean (SEM) ribbons; the dashed line is the mean of

50

autonomous AI-alone runs, and the dotted line is the do-nothing baseline. The

94

k end-of-phase gap in Phase A falls to

13

k once the advisor is gone. A wide two-part chart of company value over twenty-four months. A header strip shows three icons: a shirt for the clothing factory, a coin for company value, and a speech bubble marked AI for the advisor. In Phase A, the \condai{} mean stays high while the \condnoai{} mean declines, ending 94 thousand apart. After the advisor is withdrawn at the phase boundary, both means fall together and end 13 thousand apart.

	Withdrawal of the state-conditional rate (phase $\times$ condition)
Prediction accuracy (mental model)	worse for ai	equivalent	better for ai
worse for ai	substitution	substitution (model)	mixed
equivalent	performance cost only	no measurable cost	augmentation (transfer)
better for ai	mixed	augmentation (model)	augmentation

Table 1. The verdict table diagnoses augmentation and practical equivalence on equal footing with substitution. Pre-registered joint verdict over the two confirmatory dependent variables (DVs). Each DV is judged worse / equivalent / better for AI-supported : its confirmatory mixed model gives the direction (for H1 the between-group prediction-accuracy contrast, for H2 the phase

\times

condition withdrawal interaction, where worse for ai means the state-conditional gap narrows on withdrawal, not that unaided performance falls below controls (the registered plan reports the Phase-B condition effect alongside it)) and the Bayesian counterpart’s

95\%

highest-density interval (HDI) lying within the

\pm 0.40

region of practical equivalence (ROPE) renders the verdict as equivalent ; a DV that is neither, is inconclusive , making the overall verdict inconclusive. Bold cells are the three theory-anchored outcomes. For Study 1 the H1 axis is void rather than equivalent : both groups answered that battery at chance, so the equivalence rule cannot be applied to it and the joint verdict is inconclusive on the model axis which motivated Study 2. A three-by-three verdict table crossing the H1 mental-model outcome (worse, equivalent, better for \condai{}) with the H2 withdrawal outcome, assigning each cell a reading such as substitution, augmentation, no measurable cost, or mixed.

Panel of Figure 2 — Figure 2. Both groups share every stage; the conditions diverge only in the advisor’s presence during Phase A and its withdrawal for the unaided Phase B. Study flow with timing, condition contrast, per-stage measures, and the autonomous-AI baseline. Lanes. Grey boxes span both groups; white = unaided play, coral = LLM advisor available, orange = withdrawal phase (advisor withdrawn). The yellow box is the prediction task (24 engine-scored items), administered once, immediately after Phase A. The dotted box on the AI-supported lane between Phase A and Phase B is the single withdrawal-notice screen shown to participants at the start of Phase B. Time column (left) gives approximate stage duration; the total session runs $\sim$ 90 minutes. Measures column (right) lists the data collected at each stage: descriptive individual-difference scales, the post-Phase-A prediction task, per-round self-report and company value across the 24-month test, the AI-supported -only trust-and-reliance battery (administered after the aided phase, before withdrawal), and the post-task strategy survey and debrief. The AI-alone benchmark (dashed green) uses the same LLM and prompt as the advisor, acting on the simulation via the same structured action interface, run on the same fixed default market seed as every human session. A two-lane study-flow timeline. Both groups share an exploration phase, the post-Phase-A prediction task, and individual-difference scales; the \condai{} lane additionally completes a trust-and-reliance battery after the aided phase, and both complete a post-task strategy survey and debrief. The two lanes diverge in Phase~~A (the \condai{} lane has the LLM advisor available, the \condnoai{} lane plays unaided) and both play Phase~~B unaided after a withdrawal-notice screen on the \condai{} lane. A separate dashed lane shows the AI-alone benchmark run on the same seeds.

Figure 3. Both conditions share one Tailorshop interface; the advisor is the only thing added. The monthly decision dashboard as AI-supported participants saw it in Phase A: the left column reports the firm’s state (company value in the brand accent, plus the coupled economic variables with per-month trend arrows), the right column holds the twelve decision controls submitted each month, and the center a scrolling news feed. Three elements are added over the identical game: an “AI Advisor Active” banner, a queryable advisor panel that rendered the advisor’s full reply (analysis, options, recommendation and risks as text and option cards; the suggested decisions as per-control chips with an Apply action; replies ran to a median of

343

displayed words), and per-control suggested-value chips, while the underlying simulation, seed, and state are unchanged. Control participants saw this same dashboard with those three elements absent. Screens are faithful reconstructions of the deployed frontend with illustrative values, not participant data. A screenshot of the Tailorshop management dashboard in the \condai{} condition, showing a left panel of company-state variables, a right panel of twelve monthly decision controls, and a central news feed, plus a purple ``AI Advisor Active'' banner, an advisor recommendation panel with an Apply button, and suggested-value chips on the decision controls.

Figure 4. Scoring each month against a do-nothing counterfactual credits skilled decisions in bad situations that the raw rising-months rate penalises. How performance is scored (schematic, illustrative data). Primary (state-conditional): at every month the participant’s recorded state is branched once with the pre-filled do-nothing defaults (dashed segments); the month is scored 1 (filled green) iff the achieved company value exceeds that counterfactual, 0 otherwise (filled red). The H2 rate is the proportion of scored months (2–12). Secondary (rising-months rate): the raw rise/fall of company value versus the previous month (triangles). The example shows why the primary is fairer to good decisions in bad states: a steadily declining trajectory scores only 4/11 on the raw rising-months rate but 8/11 against passivity: most of its months are the best of a bad situation. A single declining company-value trajectory over twelve months. From each month's point, a short dashed segment shows where the value would have gone under do-nothing defaults; each achieved point is coloured green when it lies above that counterfactual and red when it does not. A strip below marks each month's raw rise or fall as up or down triangles. The trajectory falls in most months (rising-months rate 4 of 11) yet beats the do-nothing counterfactual in 8 of 11 months.

Figure 5. The advisor’s performance advantage is gone by the first unaided month, not eroded gradually (

n=200

; the both-outcomes analytic sample). State-conditional beat rate (months 2–12) by phase and condition under the registered do-nothing reference (left) and the harder persistence-reference sensitivity (right); points show condition means with 95% confidence intervals across participants (proportion scale), thin lines individual participants. The Phase-A advantage of AI-supported disappears in the unaided Phase B on the registered do-nothing reference; on the ceiling-free persistence reference, a small AI-supported advantage survives (

b=+0.24

p=.022

g=0.32

). Two interaction panels. The AI-supported mean starts above Control in Phase A. In Phase B the means converge under the do-nothing reference, while a smaller AI-supported advantage remains under the persistence reference.

Figure 6. The aided performance advantage vanishes at withdrawal, while no difference in prediction accuracy is detectable on this battery. The confirmatory dependent variables by condition: H1 post-Phase-A prediction accuracy and the H2 Phase-B state-conditional rate. The two panels do not share a denominator : H1 is

24

items scored within each of

200

participants, H2 is

2{,}200

participant-months (

11

scored Phase-B months each), so the panels’ interval widths are not comparable to one another. Dots are participants (jittered), encoded by both colour and shape. The black point and bar are the model-implied marginal mean and its

95\%

interval, taken from the same mixed models that carry the confirmatory tests (

\text{correct}\sim\text{cond}+(1|\text{pid})+(1|\text{item})

;

\text{beat}\sim\text{cond}+(1|\text{pid})

) rather than from a raw across-participant CI, so the interval reflects the item and participant variance the tests actually model. On H2 the model-implied marginal (

.77

.76

) sits slightly below the observed proportion (

.80

in both groups) because the participant-level variance is large (

SD=1.75

on the logit scale); the groups are indistinguishable on either. Dashed lines on the H1 panel mark chance (

.50

) and the ceiling a constant “lower” response would reach against the unbalanced key (

.625

). Two panels, one per confirmatory dependent variable, showing per-participant dots by condition (distinguished by color and shape) with the model-implied mean and its 95 percent interval, plus dashed reference lines at chance and the constant-response ceiling on the first panel. The \condai{} group sits negligibly higher on prediction accuracy, and the two groups are nearly identical on the Phase-B state-conditional rate, with large overlap between conditions.

Figure 7. Study 1: prediction and unaided-level contrasts lie inside the region of practical equivalence. Posteriors of the standardized between-group contrast, with the

\pm 0.40

ROPE shaded and

95\%

intervals as black bars (H1 HDI; Phase-B equal-tail CrI). H1 uses a log-odds-derived standardization, Phase B a probability difference divided by the between-person SD. Positive values favor AI-supported . H1 holds

100\%

of its posterior mass inside the ROPE and H2’s Phase-B level

99.7\%

. That licenses a practically equivalent reading of the Phase-B level , which is not the withdrawal interaction the verdict table crosses on its H2 axis. On that axis the study is worse for AI-supported . It does not on H1: a battery both groups answered at chance cannot register the knowledge the rule presumes, so that axis is void rather than equivalent. Mass inside the band means no effect as large as

d=0.40

. It does not exclude smaller ones. Both panels are beta-binomial models on per-participant counts. Two density plots of standardized effect sizes with a shaded band from minus 0.4 to plus 0.4. Both densities are centered near zero and sit almost entirely inside the band.

Figure 8. Confidence anchored to the system tracks its visible state; confidence anchored to the advisor leaves with it (E4). Mean decision confidence per round across both phases for the

199

complete reflection sequences (ribbons are

\pm

1 SEM across participants); the grey band marks the between-phase reset and, for AI-supported , the advisor’s withdrawal. Control confidence declines as the firms lose value, rebounds sharply at the identical Phase-B reset, and holds through Phase B; AI-supported participants show no such rebound and decline across the unaided months, at performance equal to Control . A line plot of mean decision confidence per round over twenty-four rounds spanning both phases, with a grey band at the phase boundary labeled advisor withdrawn. The \condai{} line sits above the \condnoai{} line throughout Phase A; after the boundary, the \condnoai{} line rises to meet it, and the two converge.

Figure 9. The felt understanding the advisor supplied is not retained: the gap between groups closes at withdrawal, and it closes because Control rises. Reported understanding of the system (1–7), by month and group, for

199

complete reflection sequences. Ribbons are

95\%

confidence intervals across participants, and the grey band marks the reset between phases. While the advisor is available AI-supported participants report understanding the system better than Control participants do (

4.75

vs.

4.27

averaged over Phase A). Across the reset itself, the last aided month to the first unaided one, Control participants rise while AI-supported participants barely move, the firm being new and its early months easy (

+0.40

vs.

+0.06

points). That asymmetry is reliable (

-0.34

t(185)=-2.22

p=.028

g=-0.31

95\%

[-0.59,-0.03]

) and larger on confidence (

-8.14

points,

t(192)=-2.94

p=.004

g=-0.41

95\%

[-0.70,-0.13]

; Figure 8 ). Contrasting whole-phase means rather than the reset alone gives the same asymmetry more strongly (understanding

-0.33

t(188)=-3.18

p=.002

g=-0.45

95\%

[-0.73,-0.17]

; confidence

-8.13

t(196)=-4.84

p<.001

g=-0.68

95\%

[-0.97,-0.40]

), and the group gap narrows from

+0.47

in Phase A to

+0.14

in Phase B. What the figure shows is that the aided group’s advantage in felt understanding does not survive the advisor, on a measure weakly related to prediction accuracy: within this study, reported across-phase understanding correlates with prediction accuracy at

r=.00

( Control ) and

r=+.12

( AI-supported ), both intervals spanning zero ( Table 5 ). A line plot of reported understanding across twenty-four months, twelve aided Phase A months and twelve unaided Phase B months separated by a grey band at the reset. The \condai{} line sits above the \condnoai{} line throughout Phase A. At the phase boundary both lines rise, the \condnoai{} line rising further, and the two run close together for the remainder of Phase B.

Family	Measure	Phase A $g$	Phase B $g$
Outcome (E1)	End-of-phase value	$+0.94$ ∗∗∗	$+0.12$
Process (E3)	Intervention rate	$+0.49$ ∗∗∗	$-0.07$
	Decision latency	$+0.74$ ∗∗∗	$-0.07$
Subjective (E4)	Decision confidence	$+0.44$ ∗∗	$+0.01$
	Mental effort	$-0.45$ ∗∗	$-0.12$
Self-assessment (E5)	Value-estimate error	$-0.66$ ∗∗∗	$-0.03$

Table 2. The measures tabulated here differ during assistance, but show no detected between-group differences after withdrawal. Summary of the exploratory measurement families (E1–E5), all uncorrected. Hedges’

g

is AI-supported

-

Control ; “Phase A” is the aided contrast, “Phase B” the unaided one. Tests. Strategy and subjective measures are Welch two-sample

t

-tests on participant-level means. Decision latency is a Mann–Whitney

U

test on per-participant medians (the distribution is right-skewed). End-of-phase value is a Welch

t

-test on

k. Prediction accuracy and metacognitive sensitivity are elicited once, after Phase A, so they have no phase contrast. Significance is flagged as

{}^{}p<.05 $,$ {}^{}p<.01 $,$ {}^{}p<.001$ ; unflagged cells are non-significant, and every Phase-B cell is. Full statistics per family are in the materials release. A table listing exploratory measurement families with their effect sizes in Phase A and Phase B. Every Phase-A effect shown is significant. No Phase-B effect is.

Figure 10. Study 2’s small knowledge advantage is largely associated with differential solvency. (A) Control firms failed

44\%

of the time against

7\%

under assistance. (B) Structural knowledge

\theta

(registered 2PL; mean

\pm 95\%

CI) by group within each solvency stratum. No within-stratum group difference was detected. Adjusting posterior-mean scores for end-of-phase solvency reduces their group contrast from

+0.28

+0.04

, distinct from the registered model contrast of

+0.42

( Section 6 ). This post-treatment decomposition is not causal. Two panels. The left bar chart shows 44 percent of \condnoai{} firms went insolvent against 7 percent of \condai{} firms. The right panel plots structural knowledge for each group separately among solvent and failed firms. Group means are close within each group, with solvent firms above failed ones.

Figure 11. Structural knowledge is more strongly associated with performance in the unaided group than the aided group; associations with felt understanding are weak. Latent ability

\theta

from the 2PL, registered

n=198

. Points are participants, jittered vertically, and lines are least-squares fits with

95\%

confidence bands. (A) Against the Phase-A rising-months rate the slope is clearly positive in Control and near-flat in AI-supported . That difference is tested at the month level in the text. On the per-person values plotted here it does not reach significance (

t(194)=-1.64

p=.103

), so the panel illustrates the effect rather than carrying it. (B) Final-month understanding is modestly associated with

\theta

(

r=+.21

p=.032

in Control ;

r=+.13

p=.221

in AI-supported ). These final-month ratings differ from the across-month means analyzed in the text. Two scatter plots side by side sharing an x-axis of latent structural knowledge. On the left the y-axis is the Phase-A rising-months rate and the \condnoai{} point cloud trends upward more steeply than the \condai{} cloud. On the right the y-axis is final-month felt understanding, with modest positive slopes and substantial overlap between groups.

Figure 12. The same manipulation, in the same sessions, moved performance and self-assessment more than raw knowledge scores. Every between-group contrast in both studies on one standardized axis: Hedges’

g

for AI-supported minus Control with

95\%

confidence intervals, ordered by magnitude. Filled points exclude zero, open points include it. So that the axis is comparable, each contrast is recomputed here as a between-person

g

on per-participant means, which is not always the estimator the text uses for the same quantity: in particular the unaided-performance point is the Phase-B level contrast between groups, whereas H2’s registered test is the phase

\times

condition interaction, which is large and significant ( Section 4 ). The knowledge points are proportion correct, the one scale the two batteries share. Study 2’s registered analysis was latent

\theta

from a 2PL and is reported in Section 6 . Performance while the advisor is present is the largest effect in either study. The subjective registers follow at between one-quarter and two-fifths of a standard deviation, with confidence and felt understanding close in magnitude across the two samples and perceived control in the same direction but significant only in Study 2. The raw prediction and structural-knowledge contrasts are statistically unresolved, with point estimates of

+0.12

and

+0.16

. Effort runs the other way, aided participants reporting less of it. Because every contrast is standardized within the same participants and the same sessions, the figure compares magnitudes. It does not establish knowledge equivalence. Study 1

n=200

, Study 2

n=198

; per-contrast

n

varies slightly with instrument completion. A forest plot of fifteen standardized effect sizes, Hedges' g for AI-supported minus Control, ordered from largest positive at the top to largest negative at the bottom. Firm solvency in Study 2 (about plus 0.91) and aided-phase performance in Study 1 (about plus 0.76) sit near plus one as filled points, followed by performance on rising months in Study 2 (about plus 0.51). The confidence, felt understanding and perceived-control contrasts from both studies cluster between plus 0.26 and plus 0.44, mostly filled, and the unaided-phase persistence-reference contrast from Study 1 (about plus 0.32) sits filled among them. Near zero, as open points with intervals crossing it, are the prediction and structural-knowledge contrasts (about plus 0.16 and plus 0.12) and the Phase-B performance level contrast (about zero). The two mental-effort contrasts sit as filled points between minus 0.45 and minus 0.59.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 13 — Figure 13. Study 2 keeps Study 1’s aided phase and drops everything downstream of the mental-model probe. Study 2 flow: shared onboarding and unaided exploration, a single 36-month Phase A (advisor in AI-supported only), the 41-item structural-knowledge battery, and the closing checks. There is no Phase B and no withdrawal; the AI-alone benchmark and engine seeds match Study 1. A two-lane flow diagram for Study 2. Both lanes share consent, walkthrough, and a six-month exploration phase, then split into a 36-month Phase A with no advisor versus advisor available, and rejoin at the 41-item structural battery followed by the debrief; Study 2 carries no manipulation check.

Panel of Figure 14 — Figure 14. The advisor could inform every decision but execute none: it saw only what the participant saw, and its advice reached the simulation only through the participant’s own submission. The AI-supported consultation loop. (1) The participant may consult the advisor with a free-text query; the server automatically adds the observable dashboard state, a four-month trend summary (company value, bank-account cash, shirt sales and worker satisfaction), and the running Phase-A transcript (the model retains the full transcript; the panel shows only the latest exchange). (2) The advisor’s free-text reply is parsed server-side into a structured recommendation: one value per decision control, plus a one-line rationale, against the output schema fixed in the system prompt; the parsed values render as the panel recommendation and per-control chips, and Apply pre-fills the controls with them. (3) Only the participant submits the month, applied, edited, or ignored advice alike, so the advisor has no actuation rights on the engine (dashed). (4) The engine returns the next month’s state to the dashboard. In Phase B the panel is hidden and the loop reduces to steps (3)–(4), identical to Control . A diagram with three boxes: Participant (dashboard), LLM advisor, and Tailorshop engine. Arrows show an optional query from participant to advisor with server-added context, a recommendation returned to the participant, the participant submitting monthly decisions to the engine, and the engine returning the next state. A dashed crossed-out line between advisor and engine marks that the advisor has no actuation rights. An orange tag notes that in Phase B the advisor panel is hidden. By scrolling, participants in the \condai{} group could see the advisor's full analysis, the options it weighed, and the risks it identified.

	Control	AI-supported	Test
Baseline gradient on Phase-A outcome
Slope per baseline SD	$+72.4$ k	flattened	$t(196)=7.41$ , $p<.001$
Within-group $r$	$+.50$	$+.23$	interaction $-62.4$ k/ SD , $t(196)=-4.92$ , $p<.001$
A $\to$ B change in end value, by baseline half
Weak half	$+52$ k	$-61$ k	interaction $+61.0$ k/ SD ,
Strong half	$+30$ k	$-19$ k	$t(196)=4.36$ , $p<.001$

Appendix

Table 3. The advisor’s help flowed to the participants least able to survive its withdrawal. Stratification on the pre-manipulation exploration-phase baseline, which precedes any divergence between groups. Values are end-of-phase company value in $k unless noted; “weak” and “strong” are the within-group halves of the baseline distribution. All exploratory and uncorrected. A table stratifying both conditions by pre-manipulation baseline skill, showing that the baseline gradient predicting performance is flattened under assistance, that the weak-baseline aided participants lose the most at withdrawal while their unaided counterparts gain, and that weak-baseline aided participants nonetheless do not end below unaided ones.

Measure	$n$	$p$	$\mathrm{BF}_{01}$	Effect [95 % CrI]	in ROPE	Evidence
per-item prediction accuracy (24 items)	200	.379	—	$+0.04$ [ $-0.05$ , $+0.12$ ]	100 %	in ROPE
per-participant mean accuracy	200	.384	4.6	$+0.12$ [ $-0.16$ , $+0.39$ ]	98 %	moderate
month-1 beat indicator vs do-nothing reference	200	.642	5.6	$+0.08$ [ $-0.26$ , $+0.41$ ]	97 %	moderate
month-2 beat indicator vs do-nothing reference	200	.875	6.2	$+0.03$ [ $-0.31$ , $+0.37$ ]	98 %	moderate
state-conditional beat rate, months 2-12 Phase B	200	.749	6.5	$-0.05$ [ $-0.30$ , $+0.20$ ]	100 %	moderate
per-participant Phase-B beat rate (do-nothing reference)	200	—	6.5	$-0.01$ [ $-0.28$ , $+0.26$ ]	100 %	moderate

Appendix

Table 4. Confirmatory nulls, both studies.

\mathrm{BF}_{01}

is the Bayes factor for the null over the alternative, so larger values favour absence of the effect; moderate marks

\mathrm{BF}_{01}\geq 3

, strong

\geq 10

, and inconclusive

1/3<\mathrm{BF}_{01}<3

—which is not support for the null. † marks a value between

3

and

3.5

, i.e. only just past the moderate threshold. The effect column is the posterior median with its

95\,\%

credible interval, on the standardised

d

scale (or

\rho

for correlations and latent

\theta

for the 2PL); “in ROPE” is the share of the posterior inside the

\pm 0.40

region of practical equivalence used throughout the paper. Non-zero, but inside the ROPE marks the one case the other labels cannot express: a posterior whose credible interval excludes zero while lying almost entirely within

\pm 0.40

, so the effect is directional and within the registered, relatively broad ROPE. That is a different finding from evidence of no effect, and is not counted as one. BF labels concern the point null; ROPE labels concern magnitude. A positive interval need not settle either question. Frequentist

p

values refer to the specified tests. Bayesian counterparts for the registered confirmatory contrasts. These carry the H1 equivalence verdict and the Phase-B level check.

Measure	$n$	$p$	$\mathrm{BF}_{01}$	Effect [95 % CrI]	in ROPE	Evidence
mean self-rated understanding (1-7) vs prediction-battery accuracy	100	.996	6.5	$-0.00$ [ $-0.19$ , $+0.19$ ]	—	moderate
mean self-rated understanding (1-7) vs prediction-battery accuracy	100	.253	3.4	$+0.11$ [ $-0.08$ , $+0.30$ ]	—	moderate †
mean per-round decision confidence (0-100 VAS) vs prediction-battery accuracy	100	.743	6.2	$-0.03$ [ $-0.23$ , $+0.16$ ]	—	moderate
mean per-round decision confidence (0-100 VAS) vs prediction-battery accuracy	100	.733	6.2	$+0.03$ [ $-0.16$ , $+0.23$ ]	—	moderate
mean self-rated understanding (1-7) vs prediction-battery accuracy	200	.355	6.0	$+0.06$ [ $-0.07$ , $+0.20$ ]	—	moderate
mean per-round decision confidence (0-100 VAS) vs prediction-battery accuracy	200	.924	9.2	$+0.01$ [ $-0.13$ , $+0.15$ ]	—	moderate

Appendix

Table 5. Calibration nulls: self-rating against measured knowledge.

\mathrm{BF}_{01}

is the Bayes factor for the null over the alternative, so larger values favour absence of the effect; moderate marks

\mathrm{BF}_{01}\geq 3

, strong

\geq 10

, and inconclusive

1/3<\mathrm{BF}_{01}<3

—which is not support for the null. † marks a value between

3

and

3.5

, i.e. only just past the moderate threshold. The effect column is the posterior median with its

95\,\%

credible interval, on the standardised

d

scale (or

\rho

for correlations and latent

\theta

for the 2PL); “in ROPE” is the share of the posterior inside the

\pm 0.40

\pm 0.40

p

values refer to the specified tests. Bayesian counterparts for the self-rating-versus-accuracy correlations in both studies. These quantify associations between self-ratings and measured knowledge.

Measure	$n$	$p$	$\mathrm{BF}_{01}$	Effect [95 % CrI]	in ROPE	Evidence
metacog_sens (conf correct - conf incorrect)	200	.428	4.8	$+0.11$ [ $-0.17$ , $+0.37$ ]	—	moderate
auroc2 (type-2 AUROC, confidence resolution)	200	.286	3.8	$+0.14$ [ $-0.13$ , $+0.41$ ]	—	moderate
declarative importance rating spread (within-participant SD of 10 ratings)	200	.903	6.5	$+0.02$ [ $-0.25$ , $+0.28$ ]	—	moderate
Phase-B signed company-value forecast error ($k)	196	.797	6.4	$-0.02$ [ $-0.29$ , $+0.25$ ]	—	moderate
beat (B2-B4 first unaided months)	600	.499	2.5	$-0.11$ [ $-0.47$ , $+0.25$ ]	94 %	inconclusive
beat \| prior unaided month missed	430	.589	2.8	$+0.08$ [ $-0.27$ , $+0.43$ ]	96 %	inconclusive

Appendix

Table 6. Mental-model nulls.

\mathrm{BF}_{01}

is the Bayes factor for the null over the alternative, so larger values favour absence of the effect; moderate marks

\mathrm{BF}_{01}\geq 3

, strong

\geq 10

, and inconclusive

1/3<\mathrm{BF}_{01}<3

—which is not support for the null. † marks a value between

3

and

3.5

, i.e. only just past the moderate threshold. The effect column is the posterior median with its

95\,\%

credible interval, on the standardised

d

scale (or

\rho

for correlations and latent

\theta

for the 2PL); “in ROPE” is the share of the posterior inside the

\pm 0.40

\pm 0.40

p

values refer to the specified tests. Bayesian counterparts for the Study 1 mental-model contrasts that stand behind the H1 verdict.

AI-Assisted Decision Making

AI-assisted decision making studies how computational systems can inform human choices while accounting for context, uncertainty, and consequences.

See all 144 papers

Sep 15, 2026

Available but Unclaimed: An Empirical Study of Human-AI Synergy

People increasingly reason with large language models (LLMs), yet complementary capabilities do not guarantee outperforming both components. In a between-subjects study, participants (N=535) solved a 40-item battery of matrix reasoning, mental rotation, syllogisms, and letter-string analogies, unaided or with GPT-5.6-Luna, Claude Opus 4.8, Gemini 3.6 Flash, or Kimi K3. Each assisted trial required consultation with the model. Each model answered every item alone 100 times under matched elicitation. The assisted-unaided accuracy difference increased with item-level LLM competence. Deference varied across tasks and increased with competence within tasks. Post-advice confidence distinguished correct from incorrect answers less strongly than unaided confidence. In a reference comparison, about half the increase in LLM accuracy carried through to assisted accuracy. How much of that accuracy gain reached participants differed across the models. These findings motivate evaluating LLMs in interaction with humans and designing support for selective deference that preserves independent reasoning.

Robin Welsch, Michelle Rausch, Pascal Knierim +4

Aalto University, Espoo, Finland · University of Innsbruck, Innsbruck, Austria · Humboldt-Universität zu Berlin, Berlin, Germany +1

Human-AI Complementarity AI-Assisted Decision Making Human-AI Collaboration LLM Reasoning

Ask PDF

Jul 15, 2026

AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized

Knowing when to say "I don't know" is fundamental to human judgment, yet AI assistants offer a fluent answer to almost any question. In five experiments (N = 3,132; four preregistered, one direct replication), participants answered difficult questions and could always decline to respond. We engineered the questions so that AI advice was wrong, separating AI use from its accuracy. Merely having access to AI nearly eliminated participants' willingness to suspend judgment, and this held whether the advice was actively requested or simply displayed. Consequently, participants answered more questions but were correct about a third as often as when AI was unavailable-yet their confidence nearly doubled. Incentivizing accuracy and penalizing inaccuracy led participants to seek and follow AI advice less, answer more accurately, and suspend judgment more often, though still far less than when AI was unavailable. As AI suggestions grow ubiquitous and unsolicited, they may not simply affect answer accuracy; they may even alter the metacognitive threshold at which people decide whether they know enough to answer.

Chiara Marcoccia, Walter Quattrociocchi, Valerio Capraro

Ecole Normale Supérieure, Paris, France · University of Rome “La Sapienza”, Rome, Italy · University of Milan-Bicocca, Milan, Italy

AI-Assisted Decision Making Human-AI Decision Making

Ask PDF

Aug 24, 2026

How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles

While AI assistance can improve human task performance in the short term, it may also undermine the development of skills in the longer term. We examine this tension in a controlled logic-puzzle experiment involving on-demand AI assistance, where participants complete tasks before, during, and after AI is available. By experimentally varying AI request costs, we find that lower-cost assistance induces more frequent AI use. We also find that participants who request AI assistance during the AI-access phase perform worse at the task after assistance is removed, and their subsequent unassisted performance is overestimated when predicted from earlier AI-assisted performance. We use a Bayesian latent ability model to separate initial ability, post-AI ability, and participant-specific skill change, while estimating how independent reasoning during the AI-access phase relates to skill development. The results show that greater independent problem-solving effort is associated with larger gains in latent ability, consistent with the interpretation that skill development is weaker when AI assistance substitutes for independent reasoning.

Shang Wu, Catarina G Belem, Shuyuan Fu +2

University of California, Irvine USA

Human-AI Collaboration

Ask PDF

Abstract

AI-Assisted Decision Making

Explore similar work

Available but Unclaimed: An Empirical Study of Human-AI Synergy

AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized

How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles