Generative AI Steering and Human Authoring: Complementary Search-as-Learning Scaffolds in Cross-Session Aggregated Academic Search

Synthesis: Academic search rarely fits in one sitting, so learners must reconstruct context when they return. Momeni and Hoeber compare two consolidation scaffolds for that moment: STAIS, which drafts GenAI summaries the learner can correct, reorganize and fact-check, and NT, a rich-text note editor where the learner authors the consolidation. In a controlled between-subjects study (n = 36) run across two sessions one week apart, the scaffolds did not make learners uniformly better or worse — they supported different facets. STAIS cut time to reacquaintance (0:55 vs 1:35) and raised precision of saved resources, while NT produced richer connections between concepts. The authors frame the division of consolidation labor between system and learner as a design variable rather than a settled good.

Key Findings

1. STAIS shortened reacquaintance and sharpened resource precision. Time to reacquaintance was 0:55 with STAIS against 1:35 with NT (t(34) = 3.13, p < 0.01, d = 1.04), and precision of saved resources was 0.83 against 0.74 (p < 0.01, d = 0.95).
2. STAIS participants reorganized more. They moved 5.8 items between clusters in the resumed session against 3.5 for NT (p < 0.01, rrb = 0.52), evidence of active restructuring of the workspace rather than passive reading.
3. Steering became more frequent on return. Inline bullet corrections rose to a mean of 5.4 in the resumed session from 2.9 initially (Wilcoxon Z = 3.52, p < 0.001), so learners kept editing the AI summaries rather than accepting them.
4. The note-takers shed lower-order work. NT participants cut verbatim extraction from 1.2 to 0.3 (p < 0.01) and paraphrased summaries from 1.4 to 0.6 (p < 0.01), raising the share of higher-order activity in their notes while its absolute count held steady.
5. The AI summaries still needed factual repair. STAIS factual-error corrections rose from 1.2 to 2.1 (p < 0.01), so the scaffold shifted work toward correcting the model rather than toward reproducing sources by hand.
6. The two conditions produced contrasting profiles, not a winner. STAIS was associated with broader conceptual coverage and higher self-reported domain knowledge; NT was associated with richer inter-concept connections and higher self-reported certainty.

Two consolidation scaffolds, two profiles

Aggregated academic search combines heterogeneous sources — a university library and a reference encyclopedia, in this study — into one workspace, and multi-session tasks require learners to reconstruct what they found and how they organized it. The authors treat end-of-session consolidation as the under-explored step in search-as-learning scaffolds. STAIS automates part of it by generating GenAI summaries, then leaves steering affordances (regeneration, inline corrections, fact-checking) with the learner. NT supplies rich-text editors and leaves the consolidation wholly to the learner. Both were enhancements of a shared baseline interface built in earlier work, so the contrast isolates the division of consolidation labor rather than a general search-interface effect.

What the outcomes did and did not share

The initial 15-minute session produced near-identical time on task (14:55 vs 14:50), as expected from a fixed allocation, and the two conditions diverged only through the condition-specific actions each added. In the resumed session three utility measures separated them: reacquaintance time, precision, and items moved. The content analysis showed both conditions moved toward more cognitively demanding activity, but by different routes: NT participants shed reproductive work, while STAIS participants added higher-order corrections — precision refinements and cross-source work — on top of a continuing need to correct factual errors the summaries introduced. The authors read this as complementary rather than rivalrous: each scaffold strengthens a facet of post-resumption learning the other leaves weaker.

Why the division of labor is the design variable

The finding that matters for learning design is that neither scaffold dominates. When the system performs consolidation, resumption is fast and coverage broad, but the model's errors become the learner's repair task and inter-concept linking is thinner. When the learner performs it, connections between ideas are richer but reacquaintance is slower and the learner carries more reproductive labor. The authors position this trade-off as something a designer chooses deliberately — how much of the consolidation to hand to the model, and which steering affordances to leave in the user's hands — rather than as a capability the model either has or lacks. The worry that Generative AI shifts cognitive work away from the learner becomes a design question the study can specify — which consolidation tasks does the model take, and which stay with the learner — rather than a verdict it settles.

What this means for practice

• Instructional designers. Decide explicitly which part of a multi-session task the AI summaries own and which part the learner owns, then expose controls for the part the learner keeps — regeneration, inline editing and fact-checking were the affordances learners used most on return.
• Software developers. Treat reacquaintance time as a measurable design target: on a workspace that persists across sessions, a summary-and-steer scaffold cut it by roughly 40 seconds against note-taking in this study.
• Software developers. Build fact-checking into any generated summary, because the summaries introduced claims needing correction and learners used the correction tools heavily in the resumed session.
• Instructors. Warn students that fast resumption and broad coverage are not the same as connected understanding; the note-taking condition produced more links between concepts even though it resumed more slowly.

Limitations

• The study enrolled 36 participants, all from one university, and both conditions used a single unfamiliar topic, so the profile differences may not hold for domains where learners already have prior knowledge.
• The initial session was a fixed 15 minutes, which limited how much material either scaffold had to consolidate and may compress the differences between them.
• The domain-knowledge and certainty outcomes are self-reported; the researcher's coding of notes and corrections supplies the behavioral evidence, and the authors did not report delayed learning measures beyond the one-week resumption.
• The scaffolds were evaluated with a single pair of source types (a library catalogue and Wikipedia), so generalization to other aggregated collections is untested.

Citation

Momeni, M., & Hoeber, O. (2026). Generative AI Steering and Human Authoring: Complementary Search-as-Learning Scaffolds in Cross-Session Aggregated Academic Search. Information Processing & Management Conference (IP&MC 2026); arXiv:2610.12755.