Two games, two rounds, one assistant removed
The paper reports two problem-solving user studies with 139 and 111 participants, built around Connect Four and Othello. In each, a participant works on a puzzle with an AI assistant available, then faces related puzzles with no assistant. The outcome of interest is that later unassisted performance, which the authors call transfer, not the score achieved while help was on screen.
To describe the conversations, the authors borrowed a dialogue act framework from tutor-student dialogue research. A dialogue act is a label for what an utterance is doing in the exchange, such as articulating reasoning, asking for the answer, or reporting the current board state. The same scheme is applied to both sides of the conversation, so participant turns and assistant turns are coded with parallel categories.
The assistant was deliberately given the optimal next move, which reduces variation in whether its advice was correct but makes the interaction less like a real AI model. Participants were also screened for the ability to produce high-quality prompts. Both choices sharpen the comparison inside the study while narrowing who and what it describes.
Thinking aloud tracks better unassisted play
The participant-side act the paper calls Think Aloud covers turns where the user verbalizes their own thought process rather than asking the assistant to decide. In Othello, Think Aloud prevalence correlated with averaged unassisted disc count at ρ = +0.358. In Connect Four the same association appeared at a weaker size, ρ = +0.182 (p = .032).
The authors checked whether this was simply a proxy for talking more. Turn count, total words, and mean words per turn each correlated only weakly with transfer, and each dropped to statistical noise when Think Aloud was held constant, while Think Aloud stayed at ρ = +0.352, +0.302, and +0.288 depending on which volume measure was partialled out. Adding mean words per turn to the regression left Think Aloud at +0.329 and the volume term at +0.011.
A dissociation makes the point sharper. In Othello, turn count related to the assisted round (ρ = +0.261) but not to transfer (+0.075), whereas Think Aloud was unrelated to the assisted round (−0.052) and strongest on transfer. One qualification the authors state: the count of Think Aloud turns also tracked transfer (+0.327), so the data separate reasoning talk from talk in general, not a rate from an amount.
Asking for the next move goes with weaker transfer
Solution Request labels turns where the participant asks the assistant directly for the next best move. In Othello this act was coupled with worse unassisted performance. The authors read it as executive help-seeking: the user hands over both the interpretation of the game state and the choice of action, so a good position can be reached without building the representation that produced it.
That reading is the authors' interpretation, not a measured mechanism. The study is observational, and the paper states that the relationships between participant behavior and transfer remain observational. What the data show is an association between a conversational pattern and a later score, not that changing the pattern would change the score.
The two participant acts are also entangled. Think Aloud and Solution Request were negatively correlated (r = −0.26), and once Think Aloud entered the Othello regression, Solution Request showed no significant relationship with transfer. The paper therefore treats the two as partially overlapping interaction patterns rather than as independent levers.
Assistant replies mirror the participant pattern
The assistant side shows a parallel split. When the assistant explained the current problem state, coded as Board Report, participants subsequently performed better without help. When it directly provided the next action, coded as Provide Correct Answer, subsequent unassisted performance was worse.
These assistant acts are not independent of what users do. In a simulation of 1,347 turns across 270 conversations, Solution Request strongly elicited Provide Correct Answer, and Board Report was more likely after certain participant turns. The authors describe the participant behaviors as driving assistant responses accordingly: elaborating thought processes elicits more tutor-like behavior, while soliciting solutions produces direct answers.
That makes the assistant pattern a mirror of the user pattern rather than a separate finding about assistant design. The paper's own framing is that performance without assistance tracks how reasoning work is divided between user and assistant, and the assistant's reply style is part of how that division is enacted.
The assisted score is a poor guide to transfer
One of the more consequential results is a mismatch between the two outcomes. In Othello, the correlation between assisted and unassisted performance was negative, so the participants who looked strongest while help was available were not the ones who did best after it was removed.
Survey and workload data point the same way. Participants who went on to transfer better rated the assistant as less easy to use (4.03 against 4.39, U = 1129, p = .013). During the assisted round they reported higher effort (4.90 against 4.34, p = .045) and rated their own performance markedly lower (2.72 against 3.66, p = .012); in the final unassisted round that self-rating gap reversed (3.23 against 2.64, p = .054).
The practical implication is specific to this setup: a smooth assisted session, or a user's own satisfaction with it, is not a reliable signal that the user will retain the skill. The paper argues that curricula targeting human-AI collaboration performance may overlook retention of performance without AI assistance.
Similar-looking reasoning turns are not equivalent
The aggregate Think Aloud label hides variation. Qualitative and subtype analyses showed that ostensibly similar reasoning turns can elicit different assistance, so two utterances both coded as thinking aloud may not represent the same kind of cognitive work.
Annotation reliability reflects this fuzziness. Across a three-model panel, Think Aloud had the lowest agreement of the participant acts in both games (Fleiss' κ = 0.81 in Connect Four and 0.75 in Othello, with majority-label prevalence of .36 and .31). A single blind human coder against the pipeline reached κ = 0.62 on 150 turns, and that coder's own test-retest agreement on Think Aloud was 0.40, against 0.84 to 0.96 for every other act.
The authors connect this to the limits of coarse dialogue taxonomies: the category is useful for finding the aggregate pattern, but it is the act whose boundary is hardest to apply consistently, for a human as well as for the panel.
What the evidence cannot settle
The authors tested whether prior skill explained the Think Aloud association by adding three pre-assistance measures: whether the unaided first move was optimal, confidence in that move, and self-rated game skill. None correlated with Think Aloud (Spearman ρ = +0.02 to +0.14, all p ≥ .13) or with transfer (ρ = +0.05 to +0.15, all p ≥ .12), and the Think Aloud coefficient stayed stable when they were added (β = +0.303 with all three, against +0.375 unadjusted).
The authors are explicit that this does not close the question. Those measures are coarse, range-restricted by screening, and barely predicted unassisted performance even in a no-assistant baseline. The paper says the data cannot rule out that Think Aloud is a symptom of participants who already attend closely to the board, and that a validated pre-test or a randomized manipulation of Think Aloud would be needed.
Scope is also bounded. The studies measure immediate near-transfer to related puzzles of the same game type, not delayed retention or far transfer, so the authors do not claim durable learning. Connect Four and Othello have objectively optimal solutions, which may not generalize to open-ended domains such as writing or programming.
Where the priority lies for AI-assisted work
The paper's distinctive claim is a comparison, not a verdict on AI use. Immediate assisted performance and later unassisted performance can move in opposite directions, and the conversational behaviors that track them differ: Solution Request tracked better in-task performance in prior work the authors cite, while in these studies it accompanied worse transfer.
That makes the user's own reasoning talk the more informative signal in this setting. The authors' suggestion follows from the same variable: preserving users' cognitive participation in problem-solving may support performance beyond the assisted interaction, and they propose helping users steer assistants toward pedagogical roles rather than treating executive help-seeking as something productive collaboration must avoid.
The threshold a reader should carry away is about which outcome is being judged. If the goal is the assisted session's output, the assisted score and the user's comfort are the relevant measures; if the goal is what the user can do afterward, this paper's evidence points to how much of the reasoning the user kept, not how well the session went.
Sources used in this answer
Human-AI Conversational Behaviors Predict Unassisted Task Performance
Across two game-based studies (n=139 and n=111), participants who verbalized their own reasoning showed better unassisted performance, while directly requesting solutions was associated with worse transfer; assistant explanations of the current state tracked better transfer than directly providing the next action.