Skip to content

From AI Reviewers to Evidence Assistants: Quantifying the Human-AI Responsibility Boundary in Peer Review

Conference: EMNLP 2026
Anthology: 2026.emnlp-main.1296
Paper: ACL Anthology
Area: Reasoning
Keywords: peer review, responsibility boundary, evidence assistant, rating calibration, grounded synthesis

TL;DR

Leveraging 600 ICLR 2026 submissions, 2,231 official human reviews, and 3,600 AI reviews, this paper quantifies the human-AI responsibility boundary across review usability, score calibration, panel diversity, and grounded synthesis, showing that AI should serve as an inspectable evidence assistant while scoring and final evaluative synthesis must remain human-led.

Background & Motivation

The fundamental bottleneck in academic publishing is rapidly transitioning from the production of papers to their rigorous evaluation. With the ubiquitous integration of large language models across literature retrieval, experimental coding, and automated scientific discovery pipelines, the barrier to drafting technically plausible submissions has diminished drastically, driving exponential increases in submission volumes across premier computer science conferences. In contrast, the community pool of vetted, competent expert reviewers cannot expand at a matching pace. Under acute deadlines and heavy reviewing burdens, peer review infrastructures risk severe systemic failure, creating urgent pressure to explore automated machine assistance to conserve scarce expert attention.

In response, recent initiatives have explored autonomous LLM-based reviewer agents intended to generate end-to-end evaluation reports. Nevertheless, critical vulnerabilities identified across peer-review evaluations sound severe alarms: unconstrained AI reviews remain fragile, easily manipulated by hidden prompt-injection attacks or cosmetic stylistic modifications, prone to hallucinated criticisms, and heavily miscalibrated against human score distributions. Assigning algorithmic outputs the authority of independent reviewers introduces profound governance and epistemic hazards, potentially corrupting the collective standard of scientific merit with ungrounded, uninspected generative assertions.

Because automated agents cannot and should not unilaterally substitute for expert human judgment, the central question for the community shifts from whether AI should assist in peer review to which exact responsibilities AI can reliably assume and which critical judgments must remain strictly human-led. Prior literature primarily relies on qualitative position pieces or isolated single-metric evaluations, lacking a unified empirical framework across foundation models, agentic workflows, and reviewing eras. Core idea: formalize peer review as a multidimensional responsibility-boundary problem across review usability, rating calibration, panel-level breadth, and grounded synthesis, demonstrating through large-scale empirical analysis that AI functions effectively as an evidence assistant while evaluative synthesis, score calibration, and final accountability must remain firmly human-led.

Method

Overall Architecture

The investigation establishes a comprehensive benchmark comprising 600 ICLR 2026 papers balanced across acceptance outcomes (Oral, Poster, Reject) with 2,231 official human reviews and 3,600 AI-generated reviews, spanning two frontier base models (GPT-5.5, DeepSeek-V4-Pro) and three representative review agent workflows (DeepReviewer2, ReviewGrounder, AgentReview). The framework is further benchmarked on an out-of-distribution validation set of 100 NeurIPS 2022 papers with 358 official reviews written prior to the release of ChatGPT using GLM-5.2. Across this multi-dimensional corpus, the methodology measures usability, score calibration, panel breadth, and grounded synthesis, translating the resulting boundary into Review Copilot—an interactive human-led workspace.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Target Paper Full Text & Retrieved Evidence Chunks"] --> B["Multidimensional Responsibility Boundary Quantification"]
    B --> C["Review Usability & Decision Signal Diagnostics<br/>Actionability, unsupported claims, and tone civility"]
    D["Score Calibration & Distributional Scale Analysis<br/>Assessing absolute score sensitivity & positive shift"]
    C --> D
    D --> E["Panel Diversity & Semantic Convergence Decomposition<br/>Separating framework-driven vs model-driven convergence"]
    E --> F["Grounded Synthesis & Residual Dispersion Modeling<br/>Decoupling local text grounding from evaluative synthesis"]
    F --> G["Review Copilot Human-in-the-Loop Workspace<br/>Grounded QA, point editing, locking, and human-only scoring"]

Key Designs

1. Review Usability Diagnostics: verifying artifact-level competence and outcome signal recovery

Peer-review reports must provide actionable, grounded, and civil critiques rather than merely plausible text. The methodology formulates three granular diagnostics: Actionable is audited at the question-unit level, requiring both an explicit reviewer uncertainty/rationale and a concrete, rebuttal-addressable path for the authors; Unsupported is audited across weakness units paired with the top-3 retrieved manuscript evidence chunks to catch ungrounded or contradicted claims; and Professionalism is checked across the full review text to ensure civility and filter out personal or hostile remarks. Empirical results demonstrate that AI reviews satisfy basic professional criteria and achieve lower unsupported-claim rates than human initial drafts (e.g., GPT-5.5 + ReviewGrounder achieves 100% professionalism and only 11.0% unsupported claims). However, significant trade-offs between decision accuracy and critique actionability reveal that artifact usability alone does not confer reviewer authority.

2. Rating Calibration Analysis: revealing severe scale compression and conservative central tendency

Reviewer numerical ratings serve as critical process signals for Area Chairs (ACs) to rank borderline submissions and determine discussion priorities. To evaluate whether AI ratings can function as calibrated process signals, the paper evaluates Mean Absolute Error (MAE) and Spearman rank correlation \(\rho\) against paper-level human mean ratings, modeling full score density distributions. Even the strongest configuration (GPT-5.5 + DeepReviewer2) achieves only a moderate rank correlation (\(\rho = 0.696\)), with fitted regression slopes substantially flatter than the identity diagonal. Across all configurations, AI scores exhibit severe scale compression, heavily clustering within the narrow 5–7 range (mean shift of +0.85) while avoiding the lower end of the rating spectrum. Because LLMs lack sensitivity to the absolute meaning of the conference rating scale, numerical scoring and borderline prioritization cannot be safely delegated to automated agents.

3. Panel Breadth and Convergence Decomposition: exposing the illusion of independent multi-agent panels

A peer review panel derives its epistemic validity from diverse individual perspectives—different reviewers focus on distinct methodological assumptions, omitted baselines, and data leakage risks. Using Qwen3-8B-Embedding representations of review sections \(r_{p,k}\), the framework compares pairwise semantic similarity \(s_{ij}^{(p,k)}\) within human review panels versus AI review panels, decomposing convergence via \(\Delta_{F-M} = S_F - S_M\) (where \(S_F\) denotes same-framework cross-model pairs and \(S_M\) denotes same-model cross-framework pairs). In evaluative sections (Strengths and Weaknesses), AI reviews exhibit dramatically higher similarity than human reviews (\(\Delta = 0.187\) and \(\Delta = 0.203\)), with \(\Delta_{F-M} > 0\) proving that convergence is predominantly driven by shared agentic prompting structures rather than foundation models. Consequently, deploying multiple AI reviewer instances merely creates an echo chamber of redundant observations, confirming that panel-level breadth remains an irreplaceable human responsibility.

4. Decoupling Local Grounding from Meta-Level Evaluative Synthesis: geometric modeling of the judgment trajectory

A comprehensive review must bridge concrete evidence retrieval and high-level academic merit assessment. The paper formulates an orthogonal diagnostic coordinate space: the first axis measures the Balanced Top-\(k\) Evidence Support Margin \(M_i = T_i - N_i\) against 20 balanced hard-negative candidate pools, quantifying manuscript-specific factual anchoring; the second axis isolates the evidence-conditioned residual vector \(u_i = f(r_i) - f(c_i^+)\) by subtracting the nearest paper passage embedding, quantifying post-local-evidence residual information volume \(V_{\text{unit}}^{\text{res}} = \frac{1}{2n} \log \det (I + \alpha \tilde{U}\tilde{U}^\top)\). The empirical regime map demonstrates a stark functional divergence: ReviewGrounder achieves superior factual grounding (support margin 0.1125) but exhibits constrained residual movement; conversely, human reviewers maintain the highest residual dispersion (\(V_{\text{unit}}^{\text{res}} = 2.4048\)), reflecting extensive meta-level synthesis (e.g., implicit comparisons to external literature, field-wide methodological standards). This geometric evidence underscores that AI should assist with evidence localization while humans synthesize evaluative meaning.

5. Review Copilot Architecture: structuring human-mediated oversight with mandatory evidence inspection

Operationalizing these quantified boundaries, the authors develop the Review Copilot interactive environment. The workflow enforces a structured four-stage pipeline: (1) generation of a provisional starting draft based exclusively on the paper; (2) mandatory evidence-grounded question answering requiring at least seven interactive Q&A rounds with PDF source highlighting before export unlocks; (3) granular point-level inspection, revision, deletion, and locking of individual critique units; and (4) final review synthesis strictly reflecting human-vetted and locked points. By architectural design, official scores and recommendation fields are excluded from AI generation, ensuring that scoring and final responsibility remain human-authored.

Key Experimental Results

Main Results: Comprehensive Evaluation of AI Configurations and Human Baselines

The primary evaluation across 600 ICLR 2026 papers (2,231 human reviews, 3,600 AI reviews) yields the following performance profiles:

Base Model Agent Workflow Decision Acc. Accept F1 Actionable (Act) ↑ Unsupported (Unsup) ↓ Professionalism (Prof) ↑
GPT-5.5 DeepReviewer2 0.870 0.901 0.386 0.095 0.998
GPT-5.5 ReviewGrounder 0.805 0.863 0.791 0.110 1.000
GPT-5.5 AgentReview 0.538 0.638 0.576 0.225 0.992
DeepSeek-V4-Pro DeepReviewer2 0.769 0.905 0.420 0.246 0.922
DeepSeek-V4-Pro ReviewGrounder 0.718 0.723 0.848 0.143 1.000
DeepSeek-V4-Pro AgentReview 0.530 0.623 0.875 0.215 0.931
Human Baseline Official Reviews 0.726 0.711 0.770 0.379 0.938

In the controlled human-in-the-loop study (10 experienced reviewers, 40 papers), Human+AI collaboration in Review Copilot achieved superior artifact quality while maintaining human-calibrated scoring:

Review Group Decision Acc. Accept F1 Actionable (Act) ↑ Unsupported (Unsup) ↓ Professionalism (Prof) ↑ Score MAE ↓ Mean Shift Scale Ratio
AI Alone 0.742 0.739 0.634 0.097 0.995 1.19 +0.85 0.44
Human Alone 0.831 0.831 0.769 0.194 0.956 1.15 0.00 1.00
Human+AI (Copilot) 0.800 0.800 0.952 0.019 1.000 1.11 +0.14 0.74

Panel Diversity and Grounded Synthesis Diagnostics

The within-paper semantic similarity analysis (Table A) and support-scaled residual volume metrics (Table B) validate the mechanism of human-mediated semantic distillation:

Section Human Similarity AI Similarity Difference \(\Delta\) (AI - Human) Framework Contribution \(\Delta_{F-M}\)
Summary 0.815 ± 0.085 0.835 ± 0.037 +0.020 +0.066
Strengths 0.648 ± 0.109 0.835 ± 0.033 +0.187 +0.016
Weaknesses 0.631 ± 0.102 0.835 ± 0.041 +0.203 +0.021
Questions 0.503 ± 0.149 0.776 ± 0.050 +0.274 +0.065
Evaluation Group Support-Scaled Residual Volume (SSRV) ↑ Grounded Weighted Eff. Dim. (\(\text{GWDeff}\)) ↓ Synthesis Behavior
Human Baseline 0.182 7.737 Broad evaluative dispersion driven by unstated domain priors
AI Baseline (Mean) 0.201 8.149 Broad evidence coverage but dispersed, unadjudicated directions
Human+AI (Copilot) 0.243 5.802 Semantic distillation: high evidence-anchored volume concentrated in fewer strong directions

Key Findings

  • Separation of Artifact Quality from Decision Authority: LLMs can produce professional, actionable critiques with low factual hallucination rates, but their numerical scores exhibit upward inflation (+0.85) and variance collapse, confirming that artifact fluency cannot justify delegating scoring authority.
  • The Pseudo-Panel Fallacy: AI reviews exhibit high evaluative convergence (\(\Delta = 0.203\) on Weaknesses), driven primarily by shared agent workflows rather than base models (\(\Delta_{F-M} > 0\)). Multi-agent review configurations do not replicate human panel diversity.
  • Evidence of Semantic Distillation: Review Copilot reduces unsupported assertions to a negligible 1.9% while increasing SSRV to 0.243 and compressing effective dimensions (\(\text{GWDeff} = 5.802\)). Human reviewers actively edit and reject candidate points (averaging 2.33 strength and 2.13 weakness edits/rejections per review), distilling provisional evidence into consolidated, robust judgments.

Highlights & Insights

  • From System Competition to Responsibility Formalization: Shifts the academic discourse from building autonomous AI reviewers to rigorously quantifying governance boundaries across usability, calibration, diversity, and synthesis.
  • Geometric Disentanglement of Evidence and Judgment: Introduces an elegant embedding-space methodology that separates local manuscript text anchoring (\(M_i\)) from post-evidence evaluative residual volume (\(V_{\text{unit}}^{\text{res}}\)), offering an objective lens for auditing critique substance.
  • Interaction Constraints Preventing Automation Bias: Implements strict HCI safeguards in Review Copilot—such as mandatory 7-round evidence Q&A and point locking—that systematically prevent passive rubber-stamping of automated drafts.

Limitations & Future Work

  • Single-Domain Primary Benchmark: The primary 600-paper corpus is sourced from ICLR 2026; despite cross-era validation on NeurIPS 2022, generalizability across non-CS domains (e.g., medicine, life sciences) warrants further investigation.
  • LLM-as-Judge Dependencies: Granular diagnostic metrics partially rely on GPT-5.5 as an automated judge; while validated via blinded human annotations (\(n=70\), Cohen's \(\kappa \in [0.61, 0.85]\)), minor evaluator priors could remain.
  • Future Directions: Extending the evidence-assistant paradigm to author rebuttal cross-examination and Area Chair meta-review synthesis workflows.
  • vs Autonomous Review Agents (DeepReviewer2, AgentReview, ReviewGrounder): While previous systems endeavor to automate full review generation, this work demonstrates their calibration and diversity limitations, repurposing them as specialized evidence retrieval engines.
  • vs Conference AI Pilots (ICLR 2025, AAAI 2026): Where conference pilots focused on coarse post-hoc author/reviewer perception surveys, this study provides fine-grained representational diagnostics and controlled human-in-the-loop metrics.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ [Pioneering formalization of the human-AI responsibility boundary in peer review via geometric representation analysis and empirical auditing]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [600 ICLR 2026 + 100 NeurIPS 2022 papers, 6 AI configurations, controlled 40-paper user study, and 275-reviewer observational deployment]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Exceptionally rigorous narrative, clear mathematical formulations, and polished structural exposition]
  • Value: ⭐⭐⭐⭐⭐ [Crucial empirical benchmark and conceptual foundation for scientific publishing governance and AI peer-review policy]