Student-AI Interaction in Computer-Assisted Consecutive Interpreting: Patterns and Performance

Synthesis: Twenty-two Chinese-native interpreting trainees completed bidirectional computer-assisted consecutive interpreting (CACI) tasks in AI-enabled systems built on automatic speech recognition and machine translation, while eye-tracking, pen-recording and voice-recording captured how they actually divided attention between the AI output and their own note-taking. Cluster analysis of 264 stage-level observations produced four interaction profiles — Intensive Engagers, Fast Scanners, Traditionalists and Frequent Switchers — and 58.3% of observations shifted profile between the comprehension and production stages of the same task. Comprehension-stage patterns, but not production-stage patterns, were significantly associated with interpreting quality, and the AI-heaviest profile consistently scored lowest. The study shows that student-AI interaction in a demanding bilingual task is heterogeneous, unstable across task phases, and not automatically beneficial.

Key Findings

1. Four interpretable interaction profiles emerged from the eye-tracking, note-taking and switching measures: Intensive Engagers (heavy AI reliance, AI reading proportion 0.83), Fast Scanners (scanning-based processing, reading proportion 0.27, highest fixation frequency at 218.6), Traditionalists (minimal AI reliance, reading proportion 0.12, deep fixation proportion 0.56) and Frequent Switchers (shifts between AI support and handwritten notes, switch count 36.03 against 3.30 to 7.65 in the other clusters).
2. Interaction patterns were not stable traits: the overall stage-to-stage transition rate was 58.3%, so most students reconfigured how they used the AI between comprehending the source speech and producing the target speech.
3. Prior AI training mattered: targeted CACI training was associated with more stable, AI-oriented patterns, while untrained participants showed greater reconfiguration across stages.
4. Only comprehension-stage patterns predicted product quality. Intensive Engagers scored significantly lower on fluency of delivery (5.46 vs 6.17, 6.08, 6.46) and target language quality (5.70 vs 6.35, 6.28, 6.67) than all three other clusters; information completeness and temporal speech measures did not differ by cluster.
5. The authors caution that the four profiles are context-bound behavioral patterns nested within only 22 participants, not stable learner types, and that the pattern differences for output-stage data were not statistically significant.

Why the heaviest AI users did not perform best

The counter-intuitive finding is that the cluster reading AI output most (AI reading proportion 0.83 at the comprehension stage) produced the weakest delivery and language quality. The authors interpret this as critical engagement mattering more than access: attention spent on the AI subtitle stream is attention not spent on building a personal representation, and the Traditionalists and Fast Scanners retained more of their own processing route. Their recommendation is not to withhold the tool but to teach students to describe and reflect on their own interaction strategy, because the pattern is implicit — hidden in where the eyes go — and therefore invisible to the learner unless it is surfaced.

The study also makes a metacognitive argument for training: since patterns shift as task demands change, students need adaptive expertise — the ability to judge, at each stage, whether the AI support or their own note-taking is the better resource. Instructors are advised to avoid prescribing one correct way of working with AI and instead to guide students toward a strategy they can justify and revise.

Method and measures

Participants performed bidirectional tasks in a CACI environment that integrated ASR and MT features. Behavioral data were triangulated from three channels: eye-tracking (AI reading proportion, regression rate, saccade amplitude, fixation frequency, deep fixation proportion, switch count), pen-recording (note-taking duration and note counts, computed separately for the AI-support area and the notepad area) and voice-recording (target speech duration, pause count, pause duration proportion, fillers and disfluencies). Interpreting quality was rated by two experienced interpreters on an 8-point scale across four bands for information completeness, fluency of delivery and target language quality; measures were computed separately for the input and output stages of each task.

What this means for practice

• Instructors. Coach a stage-specific division of labor — AI reference while comprehending, self-generated notes while producing — because comprehension-stage patterns predicted interpreting quality while output-stage patterns did not.
• Instructors. Put a ceiling on AI reading during the listening stage: the Intensive Engagers cluster read the AI output 0.83 of the time and scored significantly lower on fluency of delivery (5.46 versus 6.17–6.46) and target language quality (5.70 versus 6.28–6.67) than every other cluster, while gaining nothing in information completeness.
• Instructors. Debrief with cued retrospection rather than a lecture on technique: the strategy is hidden in where students look, and 58.3% of observations show the pattern shifting between stages of the same task, so no single prescription will hold.
• Researchers. Analyze at the stage level rather than the task level and pair interaction measures with product ratings, since 41.7% of observations changed profile across stages and only stage-specific models surfaced performance differences.
• Researchers. Treat targeted CACI training as a candidate explanation for behavioral stability — trained participants held their pattern 66.7% of the time against 50% for untrained — and test it against institutional effects, which this cross-sectional comparison cannot separate.

Limitations

• Twenty-seven Master's interpreting students were recruited and 22 were analyzed after five were excluded for incomplete eye-tracking data; the 264 stage-level observations are nested within those 22 participants, which the authors state does not compensate for the limited participant-level sample.
• Trained participants (n = 11) came from a single institution and untrained participants (n = 11) from three others, so the authors cannot separate training effects from institutional influences on academic and socio-cultural context.
• Output-stage interaction patterns showed no statistically significant differences in interpreting quality, and quality depended on ratings from two expert interpreters on an 8-point scale (inter-rater reliability 0.774–0.854).
• Materials were eight ChatGPT-4o-drafted speeches of about one minute (110 English words against roughly 135 Chinese words) with pre-recorded ASR and MT outputs matched for accuracy and BLEURT-20 score, so the AI behavior was fixed rather than live and the design cannot trace change over time.

Citation

Kuang, H., Li, J., & Weng, Y. (2026). Student-AI interaction in computer-assisted consecutive interpreting: patterns and performance. Frontiers in Psychology, 17, 1890808.