When a Large Language Model (LLM) generates a complete essay or email from a short prompt, it maximizes efficiency. However, this efficiency comes at a steep psychological cost: the human user often feels disconnected from the final output, experiencing a profound loss of personal voice, pride, and accountability.
As Hui Wang et al. note in their scoping review of human-AI co-creation, agency—the ability to act intentionally and exert control—is frequently mismanaged when AI is integrated into creative workflows. The central challenge in modern interactive systems is not just making AI more capable, but balancing its autonomy with meaningful human control.
The Granularity Trap and Cognitive Anchoring
Why does high-powered AI assistance often leave users feeling like bystanders? The answer lies in the granularity of the interaction.
When an AI assistant provides a wholesale paragraph rewrite or a full-message draft, it introduces a phenomenon known as anchoring. In a study comparing different suggestion granularities, Benjamin Newman et al. demonstrated a fundamental trade-off: full-message suggestions maximize speed, but users edit very little of the AI's output, heavily anchoring to the machine's phrasing. Conversely, sentence-level suggestions preserve a significantly higher sense of human agency, though they require more time and cognitive effort.
Coarse-Grained (Full Draft) ---> High Efficiency ---> High Anchoring ---> Low Agency
Fine-Grained (Sentence/Edit) ---> Low Efficiency ---> Low Anchoring ---> High Agency
When an interface forces a user into an all-or-nothing choice—either accept a large block of text or reject it entirely—it strips away their stylistic autonomy. As Philippe Laban et al. argue, standard chat interfaces force users to manually spot differences and copy-paste text, making it difficult to accept surgical, localized improvements without also adopting unwanted stylistic shifts or accidental hallucinations.
The Psychological Trade-off of Ownership
The tension between productivity and authorship is not merely an interface issue; it fundamentally alters human behavior and psychological ownership.
By analyzing thousands of keystrokes and suggestions, the creators of the CoAuthor dataset, Percy Liang et al., uncovered a clear inverse relationship: a writer's sense of ownership is inversely related to the volume of machine-generated content they accept. This pattern is mirrored in behavioral data. Yixin Cheng et al. found that users who retain high ownership of a text spend significantly more time independently composing and revising. When ownership drops, users shift from active creators to passive editors, relying heavily on the AI to dictate the narrative direction.
This erosion of agency has real consequences for the uniqueness of human expression. Zhuoyan Li et al. observed that while direct content generation boosts user confidence and speed, it simultaneously decreases writing diversity and personal accountability. When the AI writes the initial draft, the user's capacity for self-expression is constrained by the boundaries of the model's training data.
Designing for Co-Creative Agency
To resolve this tension, researchers are moving away from monolithic generation toward structured, mixed-initiative interaction paradigms. The goal is to build tools that act as collaborative partners rather than automated replacements.
1. Guidance and Selection Workflows
Instead of letting the AI drive, effective systems restrict the AI's role to specific sub-tasks. For instance, Zijian Ding et al. explored different interaction levels in headline generation and found a "sweet spot" in a Guidance + Selection workflow. By requiring the human to provide keywords (guidance) and then select/refine the best AI outputs, the system lowers cognitive load without sacrificing the user's feeling of authorship.
2. Information Aids and Progressive Disclosure
When AI tools act as black boxes, offering a single authoritative suggestion, users feel dictated to by the machine. To counter this, Yewon Kim et al. developed ParaScope, an interface providing non-native English speakers with "information aids" like back-translations and quality scores. Rather than replacing user judgment, these aids empower users to make informed choices, transforming the AI from an oracle into an advisor.
3. Multi-Agent Scaffolding
For complex, long-form tasks like autobiography writing, monolithic chatbots frequently hijack the narrative. To preserve the author's personal voice, Shayan Talaei et al. introduced StorySage, a multi-agent framework that decomposes the process into distinct conversational stages: interviewing, memory logging, and narrative drafting. By separating reflection from text generation, the system ensures the user steers the topics explored, resulting in higher user agency.
[User Input] ➔ [Interviewing Agent] ➔ [Memory Log] ➔ [Drafting Agent] ➔ [User Refinement]
▲ │
└─────────────────────────── Retains Control ─────────────────────────────┘
Shifting the Paradigm Across Domains
The necessity of balancing automation with granular control extends far beyond text. Whether designing neural-network-based games to assess human creativity (Janet Rafner et al.) or building tools for sports video storytelling (Tica Lin et al.), the core insight remains identical: automation handles the tedious mechanics, but humans must retain the ability to apply their domain expertise.
As highlighted in the Proceedings of the Fourth Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2025), the future of AI-assisted creation relies on interfaces that treat human agency as a non-negotiable constraint. The historical roots of this idea trace back to early mixed-initiative systems, such as the navigation metaphors proposed in Navigation and Planning in a Mixed-Initiative User Interface, which visualized AI decision-making as a space for users to steer.
In sensitive domains like peer-to-peer mental health support, Ashish Sharma et al. demonstrate that actionable, localized feedback can empower supporters rather than replacing their empathy. By designing interfaces that favor surgical collaboration over wholesale automation, we can build AI tools that amplify human capability while firmly preserving the human soul of the work.
Key Foundational Work
Navigation and Planning in a Mixed-Initiative User Interface (1997) — contributed 2 concept elements to this direction.
crea.blender: A Neural Network-Based Image Generation Game to Assess Creativity (2020) — contributed 1 concept elements to this direction.