AnchoredAI: Precise Localized Feedback
- AnchoredAI is a writing-assistance system that attaches AI-generated feedback directly to precise text spans, enhancing context and clarity.
- The system uses an Anchoring Context Window and update-aware retrieval to maintain feedback relevance even after document edits.
- Empirical findings indicate that AnchoredAI promotes targeted revisions and boosts writer agency while mitigating automation bias.
AnchoredAI is a writing-support system that anchors AI-generated feedback directly to relevant text spans rather than presenting suggestions in a detached chat window. The system was introduced to address a recurrent problem in AI-assisted writing: chat-based interfaces can obscure referential context and amplify automation bias and overreliance. Its core design combines an Anchoring Context Window (ACW), which ensures that a comment refers unambiguously to a unique span, with update-aware context retrieval, which preserves the intent and location of prior comments after document edits. In a controlled comparison with a chat-based LLM interface, AnchoredAI produced more targeted revisions and stronger reported agency and ownership among writers (Lou et al., 19 Sep 2025).
1. Definition and intellectual setting
AnchoredAI belongs to a line of systems that adapt anchored commenting practices from CHI and CSCW to LLM-mediated writing assistance. Its central premise is that feedback should remain local, inspectable, and referentially precise at the point of revision. In this formulation, “anchoring” does not denote philosophical value anchoring or trajectory anchoring; it denotes the attachment of AI comments to exact spans in a document, together with mechanisms that maintain those attachments as the document changes (Lou et al., 19 Sep 2025).
The problem setting is specifically editorial rather than generative. Chat-based writing assistants often produce globally phrased suggestions in a separate interaction pane, which can detach recommendations from the textual locus to which they apply. AnchoredAI was designed as a countermeasure to that detachment. The system addresses three intertwined issues named in the study: cognitive load, ambiguity, and automation bias. In that sense, it is not merely a prompting layer over an LLM, but an interface architecture for constraining how model outputs are situated within a document (Lou et al., 19 Sep 2025).
This motivation is consistent with broader findings on anchoring and overreliance in AI-mediated decision support. In purchase-decision experiments, targeted AI support reduced the effect of unfair price anchors, whereas XAI alone did not significantly reduce anchoring bias (Haag et al., 2024). In LLM prompting studies, simple strategies such as Chain-of-Thought, Thoughts of Principles, Ignoring Anchor Hints, and Reflection were reported as insufficient to mitigate anchoring bias, while broader, multi-angle context was more promising (Lou et al., 2024). These results situate AnchoredAI within a wider research program that treats interface structure and context presentation as central to bias mitigation.
2. System architecture and document workflow
AnchoredAI integrates an LLM-powered feedback generator with an anchoring layer, document state tracking, anchor management, and version history. The paper describes a four-step pipeline. First, Document State Tracking detects insertions, deletions, and replacements by comparing document versions. Second, Structured Anchor Generation requires the LLM to generate feedback for selected spans in a strict format, such as JSON, including a reference to the text span and, when necessary, an ACW. Third, Anchor Validation/Mapping matches anchors to text using string matching, with uniqueness enforced through the ACW. Fourth, Feedback Rendering displays valid anchors as interactive, persistent comments tied to the text spans (Lou et al., 19 Sep 2025).
Two interaction modes organize system use. In meta-commenting, the system analyzes the whole document in response to a user query such as “find all verb tense errors” and emits multiple anchored comments across relevant locations. In text-anchored commenting, the writer selects a specific span and initiates a contextual AI chat about that span. The first mode supports distributed, document-level review; the second supports localized deliberation around a selected passage (Lou et al., 19 Sep 2025).
The implementation described in the paper uses a Microsoft Word add-in, Word’s API for comment and anchor management, document state tracking on the client side, and OpenAI GPT-4o for LLM output. Prompt design requires verbatim text span extraction in JSON, mandates ACW expansion for non-unique spans, and discourages hallucinated anchors. Sessions persist version history, comments, and mappings so that the referential state of the document survives across editing episodes (Lou et al., 19 Sep 2025).
A concise comparison with the chat baseline captures the system’s intended differentiation.
| Aspect | Chat-based LLM UI | AnchoredAI |
|---|---|---|
| Feedback locus | Disconnected chat window | Exact text span, persistent through edits |
| Revision style | Large, wholesale replacement | Small, targeted, context-specific edits |
| Agency | Lower | Higher on control and ownership items |
| Overreliance risk | High | Mitigated by localized review |
3. Anchoring Context Window
The Anchoring Context Window (ACW) is the system’s primary mechanism for referential disambiguation. The motivating problem is straightforward: a short string such as “more” may appear multiple times in a document, so a comment like “fix ‘more’” is not uniquely resolvable. AnchoredAI therefore expands the anchor’s referential context recursively until the referenced span becomes unique within the document (Lou et al., 19 Sep 2025).
The expansion policy is hierarchical. If the anchor text is unique, it is used as-is. If it is not unique, the system expands from word to sentence, then from sentence to paragraph, and, if ambiguity persists, to section. When the expansion changes the effective referential span, the system can re-prompt the LLM with the expanded ACW so that the generated feedback remains appropriate to the larger local context. The pseudo-formal description given in the paper is: if , set and repeat until (Lou et al., 19 Sep 2025).
This mechanism has two functions. First, it solves a technical anchoring problem: mapping a model-generated comment onto a unique textual object in a mutable document. Second, it changes the epistemic status of the comment from generic advice to context-bounded editorial feedback. That distinction matters because ambiguous feedback increases the interpretive burden on the writer and can encourage default acceptance or dismissal rather than localized judgment. The paper frames the ACW as a response to exactly this ambiguity problem (Lou et al., 19 Sep 2025).
A plausible implication is that ACW acts as a form of interface-level regularization. Instead of allowing the LLM to emit loosely grounded suggestions, the system compels the output to carry sufficient context for deterministic attachment. That design choice parallels, at the interface level, the broader observation that AI assistance is most effective when it is tightly aligned to the decision task rather than presented as abstract explanation (Haag et al., 2024).
4. Update-aware context retrieval and persistent comment continuity
AnchoredAI’s second key mechanism is update-aware context retrieval, introduced to handle the fact that anchored text often changes after feedback is generated. A comment may shift position, its referent may be edited, or the original text may be deleted altogether. The system therefore tracks versions and deltas between edits and reconstitutes the context needed for coherent follow-up interaction (Lou et al., 19 Sep 2025).
When a follow-up occurs inside a comment thread, the workflow described in the paper is four-stage. First, Change Localization determines what changed near the anchor using diff algorithms. Second, Dynamic ACW Recomputation expands the context window so that it encompasses both the original anchor and the modified region. Third, Dual-Context Prompting supplies the LLM with the updated ACW, the full updated document, and the prior thread history. Fourth, Response Generation yields feedback that is localized, current, and reference-preserving. The resulting response is then rendered persistently on the updated text (Lou et al., 19 Sep 2025).
This mechanism is important because it treats anchored comments not as one-shot annotations but as evolving conversational objects embedded in a mutable artifact. The system therefore supports a continuity of editorial intent: the discussion attached to a span can survive textual revision without collapsing into detached chat history. In practical terms, this is what allows a writer to revise the text, return to a prior comment, and still receive contextually coherent follow-up assistance (Lou et al., 19 Sep 2025).
The technical implementation combines client-side document snapshots, normalized string matching for revalidation, and persistent storage of comments and mappings. These elements collectively make the comment layer version-aware rather than statically positional. That property is central to the system’s claim that anchoring should remain robust “even as the document evolves” (Lou et al., 19 Sep 2025).
5. Empirical findings on revision behavior, agency, and cognitive load
AnchoredAI was evaluated in a within-subjects lab study with university students. Each participant revised academic essays under two conditions: a chat-based feedback interface and the AnchoredAI interface. The study used copy/paste logs, percentage of document changed, surveys on agency including authorship, ownership, and control, and NASA-TLX measures of cognitive load (Lou et al., 19 Sep 2025).
The reported revision behavior differed systematically across interfaces. Under the chat condition, participants often performed large, wholesale text replacement, sometimes pasting hundreds of words per action. Under AnchoredAI, edits were more, smaller, and targeted, with an average of approximately 14 concise comments per task and about 29 words per comment. Survey results indicated significantly higher agreement that writers were the main contributors, made substantial contributions, felt in control, and should be recognized as the reviewer; the paper reports these differences as significant at across most agency and ownership items (Lou et al., 19 Sep 2025).
The qualitative evidence matches the behavioral traces. One participant, identified as P20, contrasted the two interfaces as follows: in the chat condition, the participant “kind of just copy-pasted the whole revised essay,” whereas in the anchored condition, the suggestions arrived “a little bit at a time,” which made it feel necessary “to actually think about changes” before applying them (Lou et al., 19 Sep 2025). That statement is consistent with the authors’ interpretation that anchored feedback operates as a cognitive forcing function.
AnchoredAI also increased perceived mental demand. The study interprets this as desirable difficulty rather than detrimental overload: the interface introduces friction, but that friction is taken to promote deeper engagement and critical thought rather than passive acceptance. The paper explicitly connects this effect to authorship and accountability, arguing that moderate effort can help writers remain active participants in revision rather than outsourcing authorial responsibility to the model (Lou et al., 19 Sep 2025).
6. Interpretation, misconceptions, and broader significance
AnchoredAI advances a specific thesis about AI-assisted writing: interface design shapes not only the efficiency of revision but also the distribution of agency between writer and model. Its empirical results suggest that precision of attachment, persistence through edits, and comment-level granularity can reduce the tendency toward wholesale acceptance of AI output. The system therefore proposes an alternative to frictionless chat-based assistance: not maximal ease of substitution, but structured collaboration at the level of local textual decisions (Lou et al., 19 Sep 2025).
A common misconception is that any additional friction in AI interaction is necessarily a usability defect. The study does not support that interpretation. It reports higher mental demand under AnchoredAI, but treats that increase as compatible with stronger ownership and control. In this framing, friction is not simply an obstacle; it is a mechanism for preserving deliberation in the presence of highly fluent model outputs (Lou et al., 19 Sep 2025).
Another misconception is that prompt-level reflection alone is enough to counter overreliance. Related work on LLM anchoring bias found that Chain-of-Thought, Thoughts of Principles, Ignoring Anchor Hints, and Reflection were not sufficient mitigation strategies (Lou et al., 2024). Likewise, in AI-assisted purchase decisions, direct task-fitting AI support reduced anchoring more effectively than explanation alone (Haag et al., 2024). This suggests that AnchoredAI’s contribution is not reducible to better prompt wording; it lies in a representational and interactional reconfiguration of where AI feedback appears, how it is grounded, and how it persists during revision.
The broader implication is that anchored interfaces may be especially appropriate where precise referential context matters and where overreliance on synthetic revisions is a concern. The paper itself recommends using anchoring for precision and agency and chat for brainstorming and breadth, and it proposes that hybrid or adaptive interfaces may switch between the two modes depending on task and context (Lou et al., 19 Sep 2025). That recommendation positions AnchoredAI less as a universal replacement for conversational assistants than as a design pattern for cases in which writer agency, ownership, and traceable revision matter more than frictionless generation.