The AI Memory Gap: Users Misremember What They Created With AI or Without

Paper Title

The AI Memory Gap: Users Misremember What They Created With AI or Without

Publication Info

• Topic area: Memory and source attribution in human-AI collaboration
• Keywords: AI memory gap, source attribution, human-AI collaboration, large language models, ideation, elaboration, memory accuracy, confidence, source monitoring, mixed workflows

Background and Problem

• Problem / challenge: Users struggle to accurately recall whether content was created by themselves or with AI assistance, particularly in mixed human-AI workflows. Prior studies have not systematically examined source memory in human-AI co-creation.
• Significance: Accurate source attribution is critical for intellectual ownership, accountability, and trust in human-AI collaboration. Misattribution can lead to ethical, legal, and practical challenges in creative and professional contexts.
• Motivation and related work: Previous research shows users often fail to detect AI-generated content and may misattribute AI contributions to themselves. Cognitive psychology highlights that similar outputs from multiple sources impair source memory. This paper addresses the gap in understanding how AI use affects source memory in ideation and elaboration tasks.

Solution

• Proposed approach: A two-phase, within-subjects experiment to investigate how AI involvement in ideation and elaboration affects source memory and confidence.
• Novelty:
  1. Systematic study of source memory in human-AI co-creation workflows.
  2. Quantification of memory accuracy and confidence across consistent and mixed workflows.
  3. Application of the Source Monitoring Framework (SMF) to interactive AI settings.
  4. Introduction of a Multinomial Processing Tree (MPT) model to disentangle memory processes.
• Procedure and key techniques:
  • Phase 1: 184 participants generated ideas and elaborations for problem-solving tasks, alternating between AI-assisted (withAI) and unaided (noAI) conditions.
  • Phase 2: After one week, participants attributed the source of ideas and elaborations, including distractors, and rated their confidence.
  • Analysis: Generalized linear mixed models (GLMMs) and MPT modeling were used to assess memory accuracy, confidence, and cognitive processes.

Results

• Concrete findings:
  • AI involvement reduced item memory accuracy from 97.6% (noAI/noAI) to 87.9% (withAI/withAI).
  • Source memory accuracy for ideas was highest without AI (92.4%) and lowest in mixed workflows (withAI/noAI: 37.7%).
  • Source memory accuracy for elaborations was highest in consistent workflows (noAI/noAI: 91.5%, withAI/withAI: 89.2%).
  • Confidence was highest in human-only workflows (85.2 for ideas, 86.3 for elaborations) and dropped with AI involvement.
  • Participants overestimated their source attribution accuracy, particularly for ideas (by 12%).
• Advantage over baselines:
  • Consistent workflows (noAI/noAI or withAI/withAI) supported better memory accuracy compared to mixed workflows.
  • MPT modeling revealed stronger source memory for elaborations than ideas and identified biases in guessing (toward self for ideas, toward AI for elaborations).
• Experiments / evaluation:
  • Design: 2×2 within-subjects factorial design with ideation and elaboration tasks.
  • Datasets: 8 problem statements per participant, 40 self-generated items, and 20 distractors.
  • Metrics: Item memory, source attribution accuracy, confidence, and MPT parameters (d, s, β, f).
• Limitations and future work:
  • Limited to ideation and elaboration tasks; future research should explore other creative and professional workflows.
  • Focused on one LLM (GPT-4.1-mini) and chatbot interface; results may vary with different models or interaction paradigms.
  • Fixed one-week retention interval; effects of shorter or longer delays remain unexplored.

Summary

This study identifies an "AI memory gap," where AI involvement in ideation and elaboration impairs users' ability to recall the source of their contributions. Memory accuracy was highest in fully human workflows and lowest in mixed human-AI workflows. Confidence often exceeded actual performance, particularly for ideas. The findings highlight the need for system designs that enhance source attribution transparency and support consistent workflows. Future research should explore broader task domains, interaction paradigms, and retention intervals to generalize these insights.