acceptodds
Under review as a conference paper at ICLR 2027

Beyond Visual Artifacts: Knowledge-Intensive Detection of AI-Generated Images

Abstract

Detectors that rely on pixel-level artifacts or semantic visual flaws can struggle with increasingly realistic AI-generated images (AIGIs). Prior AIGI detectors have not systematically exploited a complementary signal: knowledge inconsistencies between implicit visual claims and real-world facts. Such inconsistencies can arise from generator hallucinations or deliberate creator choices. To evaluate this signal, we introduce K-AIGI-Bench, to our knowledge the first benchmark specifically designed for knowledge-intensive AIGI detection. It contains carefully human-curated, high-quality images produced by a state-of-the-art image generator, each with a deliberately embedded knowledge inconsistency, alongside topic-matched real images collected from the Internet. To address this task, we further propose K-VERA (Knowledge-grounded Visual Evidence Retrieval and Adjudication), a training-free, retrieval-augmented multi-agent framework. It identifies the scene and entities, turns implicit visual claims into independent verification tasks, and retrieves webpages and reference images to verify them. It then checks source support and resolves conflicting findings to produce an evidence-grounded verdict. Experiments on K-AIGI-Bench show that K-VERA outperforms all evaluated dedicated AIGI detectors by over 15% in balanced accuracy, while producing higher-quality, evidence-grounded explanations. Combining K-VERA with a pixel-level detector further improves performance on three in-the-wild benchmarks. These results establish knowledge-intensive detection as an important enhancement for detecting increasingly realistic AIGIs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Multimodal large language models, Retrieval augmented generation, Multi-agent systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1QVTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how factual inconsistencies can help detect AI-generated images that look realistic. For example, an image may show a product with features that the real version does not have. The paper calls this knowledge-intensive detection and introduces: (i) a task formulation to detect knowledge inconsistencies; (ii) K-AIGI-Bench with about 1,000 images and (iii) K-VERA, a training-free multi-agent framework that uses several vision language agents without additional task-specific training. Starting from an image, it identifies the scene and objects, formulates questions about the depicted details and retrieves webpages and reference images to check them. It then checks whether the sources support the findings and produces a judgment with an explanation and supporting references. The paper evaluates detection accuracy and explanation quality using ablations and a human evaluation of explanations. K-VERA is reported to perform better than the compared detectors. Combining it with an existing pixel-level detector also improves detection on three additional benchmarks. This shows the benefits of using factual verification alongside conventional detection cues.

Critical strengths
  1. The paper introduces K-AIGI-Bench, a new benchmark to detect factual errors in realistic images. K-AIGI-Bench covers eight knowledge topics and records the intended error in each generated image. This enables one to test if a detector identifies the actual inconsistency rather than simply testing if the real/fake label is right.

  2. The complete system performs well and gives explanations that can be checked. Combining it with DDA also improves detection over either method alone on all three external benchmarks.

Critical weaknesses
  1. K-VERA uses separate agents to perform web searches, factual checks and visual error detection. The paper does not compare it with a single model using the same search tools and processing budget. As a result, the paper does not show that the proposed workflow is needed.

  2. On the external benchmarks, the paper does not separate detections based on factual inconsistencies from those based on visual flaws. The results show that the complete system works but not that the proposed factual checking contributes outside the constructed benchmark.

Questions for authors
  1. How does K-VERA compare with a single agent using the same underlying model, search tools and a processing budget? This would show that dividing the checking process among several agents improves detection beyond providing access to external evidence.

  2. On the external benchmarks, how many correct fake image detections are attributed to factual inconsistencies and how many to visual flaws? Have the claimed factual inconsistencies been checked against the images and cited sources? This would show that gains are due to factual checking.

  3. Have human reviewers also checked whether the factual errors are visible in the images and supported by the cited webpages or reference images? If so, how often were both conditions satisfied?

4 comments

We thank the reviewer for the helpful comments. We address your concerns below.

Q1. Single agent with the same model and tools. We added Qwen3.7-Plus w/ tools to Table 1 (L337, L363–365): a single agent with the same backbone and all five tools. It reaches 0.739 BAcc, between the tool-free backbone (0.685) and K-VERA (0.754). K-VERA also produces clearly better explanations: GT-KI Recall 0.562 vs. 0.528, and EQS 2.17/2.71/2.82 vs. 1.83/2.64/2.65 (L398–403). Both external evidence and the multi-agent decomposition therefore contribute. Results with three further backbones are in Table 9 (Appendix D, L820–833).

Q2. Factual versus visual detections on external benchmarks. Table 14 (Appendix M, L1462–1476) gives the verdict distribution.

  • Knowledge-Inconsistent verdicts account for 10.5%, 6.1%, and 19.3% of fake images on So-Fake-OOD, WildRF, and GPT-Image-2 in the Wild.
  • AIGI verdicts account for 46.5%, 76.7%, and 43.8%.

Many AIGI verdicts are themselves grounded in web evidence rather than visual flaws (Appendix M, Fig. 15): closely related images found online that are explicitly labeled as AI-generated, or generator watermarks. On K-AIGI-Bench, where the ground-truth inconsistency is known, GT-KI Recall and EQS check whether the claimed issue is correct, visible in the image, and supported by the cited evidence. For EQS, the evaluators inspect the cited webpages and images (Table 1; Appendix H.1).

Q3. Human verification. In the user study, participants were asked to verify whether each claimed inconsistency is actually present in the image and factually correct, using web search when needed (Sec. 5.5, L471–473). For fairness across methods, K-VERA's reference webpages and images were withheld. Whether the cited sources support each claim is evaluated by the EQS evaluators, who inspect these sources directly (Appendix H.1).

Please note that we have updated the paper. In the revised PDF, revised text is shown in blue; for newly added sections, only the section titles are blue. Line numbers here refer to the revised PDF.

Reviewer_1QVTOpenReview ↗

Response to authors

Thank you for the new results. Was the single agent given the same processing budget as K-VERA?

An image can be fake even if K-VERA gives the wrong reason. On the external benchmarks, did anyone check the images to confirm that the factual errors reported by K-VERA were actually errors

Thank you for the follow-up questions.

Q1. Processing budget of the single agent. The single agent (Qwen3.7-Plus w/ tools) was allowed up to 32 tool-use rounds (Sec. 5.1, L361–363). Default K-VERA allows up to 48 rounds across its tool-enabled roles. Importantly, with a total budget of 24 rounds, i.e., at most 4 tool-use rounds per tool-enabled agent and at most 4 Actors, K-VERA still achieves 0.753 balanced accuracy, outperforming the single agent with up to 32 rounds (0.739) (Table 7, L786; Table 1, L337).

Q2. Verification of knowledge inconsistencies on the external benchmarks. Yes. We manually inspected 50 images from the in-the-wild benchmarks that are labeled fake and for which K-VERA returns "Knowledge-Inconsistent". In every one of them, we confirmed at least one knowledge inconsistency explicitly identified by K-VERA. We have added two such examples to the appendix (Appendix M, L1474–1478, Fig. 16).

Reviewer_1QVTOpenReview ↗

Response to authors from 1QVT

Thank you for the clarification. My main concerns have been addressed.

Reviewer_w3k5OpenReview ↗
Rating
4(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces the task of knowledge-intensive AIGI detection to identify visually plausible synthetic images that conflict with real-world facts. The authors present K-AIGI-Bench as a specialized benchmark for evaluating factual inconsistencies across eight knowledge topics and propose K-VERA, a training-free multi-agent framework that performs retrieval-augmented visual evidence verification. By resolving image context and retrieving multimodal web evidence to adjudicate implicit visual claims, K-VERA outperforms existing detectors in balanced accuracy and provides evidence-grounded explanations for its verdicts.

Critical strengths
  1. Novel Benchmark for Knowledge-Intensive Detection: The authors introduce K-AIGI-Bench, the first benchmark specifically designed to evaluate AI-generated images through the lens of knowledge inconsistencies. By curating approximately 1,000 high-quality images across eight diverse topics such as Biology & Nature and Geography & Architecture, this dataset provides a foundation for testing whether detectors can identify discrepancies between implicit visual claims and real-world facts rather than relying on pixel-level artifacts.

  2. The proposed K-VERA framework offers a training-free approach to AIGC detection by leveraging a multi-agent framework. By decomposing the forensic workflow into specialized roles including a Context Resolver, Planner, Actors, and an Adjudicator, the system effectively mimics human expert reasoning to resolve image context and verify complex factual claims without requiring expensive retraining for new generative models.

  3. Unlike traditional black-box detectors, K-VERA produces high-quality explanations supported by retrieved multimodal web evidence and external reference links. The experimental results demonstrate that this retrieval-augmented approach outperforms dedicated AIGI detectors by 17.4% in balanced accuracy on K-AIGI-Bench and successfully complements existing pixel-level methods to improve performance on various in-the-wild benchmarks.

Critical weaknesses
  1. Insufficient Evaluation of Performance Across Knowledge Required Dates: The K-AIGI-Bench contains a significant distribution of images requiring very recent knowledge, yet the experiments do not analyze how K-VERA performs as the "required knowledge date" approaches or exceeds the backbone MLLM training cutoff. While the authors implement a search cutoff for retrieval, the lack of a per-topic or per-date performance breakdown leaves it unclear whether the framework’s success is driven by the internal knowledge of the MLLM or the actual effectiveness of the retrieval-augmented agents for time-sensitive facts.

  2. Contextual Reliance and Error Propagation in the Multi-agent Pipeline: The framework follows a linear forensic workflow where the Planner and Actors rely heavily on the initial scene description provided by the Context Resolver. The current results do not quantify the impact of contextual hallucinations during the first stage, where an incorrect entity identification by the Context Resolver might constrain the subsequent verification tasks. Without an ablation study or error analysis focusing on how initial context errors propagate through the multi-agent hierarchy, the robustness of the "uncertainty-aware" design remains unverified.

Questions for authors
  1. For weekness 1, could the authors provide a performance breakdown categorized by the required knowledge date intervals to demonstrate K-VERA’s specific gain on events occurring after the MLLM’s knowledge cutoff?

  2. For weekness 2, can the authors conduct an experiment where K-VERA is tested with and without the Context Resolver's output to quantify how much the final adjudication depends on the accuracy of the initial scene identification? If that's the vital part design in your pipeline, can you please analyze the contextual content in the Context Resolver in the appendix. I feel interested in this point.

I will increase my score after you settle my concerns.

5 comments

We sincerely thank the reviewer for the strong interest in our work. We address your concerns below.

Q1. Breakdown by required knowledge date and topic. Table 5 (Sec. 5.6, L486–495, discussed at L480–490) reports BAcc and fake accuracy per date interval for K-VERA and the tool-free Qwen3.7-Plus.

  • K-VERA improves in every interval, and overall by +0.069 BAcc and +0.146 fake accuracy.
  • The largest gains are for the most recent facts: +0.180 for "now" and +0.115 for 2026.
  • Earlier intervals also improve (e.g., +0.138 for 2016–2020). Retrieval supplements long-tail knowledge that the backbone has not memorized, not only time-sensitive facts.

We also added a per-topic BAcc breakdown (Table 6, discussed at L497–501). The gains concentrate on entity-specific and time-sensitive topics: Events & Developments +0.129, Geography & Architecture +0.123, and Software & Interfaces +0.113.

Q2. Context Resolver. Removing the Context Resolver lowers BAcc from 0.754 to 0.738 (Table 2). As requested, we added a dedicated analysis of the Context Resolver (Appendix E, L836–961; referenced in the main text at L458–459).

Hallucinations (L838–854, Fig. 6). Among the 190 AIGIs in K-AIGI-Bench that K-VERA misclassifies, the Context Resolver produces obvious hallucinations in 27, and 19 of these are strongly related to the incorrect final verdict. Manual inspection shows that the hallucinations fall into four types:

  • Contradicting the image. In our Raspberry Pi example, the board in the image has a single full-size HDMI port, yet the Context Resolver describes dual micro-HDMI ports, as on the real Raspberry Pi 5.
  • Misinterpreting the defect. In our tuatara example, the image shows an ear opening behind the eye, which real tuatara do not have. The Context Resolver explains it as the parietal "third eye" on top of the head, which does not match its appearance or position in the image either.
  • Misidentifying the entity. In our Petra example, the image depicts Al-Khazneh, a facade carved directly into a sandstone cliff in Petra, as a free-standing building. The Context Resolver identifies it as Al-Deir, a different monument in Petra, although Al-Deir does not match the image either.
  • Forced matching. In our Statue of Liberty example, the Context Resolver finds a social-media photo that, like the input, shows the Statue of Liberty and One World Trade Center, and therefore strongly leans toward judging the image as real. However, a closer look shows that the relative positions of the statue and One World Trade Center differ between the reference and the input.

Role of the context (L855–863, Fig. 7). We inspected the samples that K-VERA classifies correctly with the Context Resolver but incorrectly without it, and found that the Context Resolver plays a key role in these samples. It describes the objects in the scene in text and provides relevant background. Besides supplying background information before the Planner starts planning, its benefit may also stem from the observation in related work that MLLMs often reason more effectively over text than directly over images (e.g., Wang et al., NeurIPS 2024, arXiv:2406.14852; arXiv:2510.22836). For example, in our WhatsApp example, the Context Resolver output explicitly records three key pieces of information in the screenshot: the open context menu containing "Delete for everyone", the date of the message, and the system-clock date. This directly helps the Planner find the conflict, since this option is visible only within 2.5 days after a message is sent. Without the Context Resolver, the Planner does not connect these three elements.

Please note that we have updated the paper. In the revised PDF, revised text is shown in blue; for newly added sections, only the section titles are blue. Line numbers here refer to the revised PDF.

Reviewer_w3k5OpenReview ↗

Re: Response to Q1

Thank you for your further breakdowns. K-VERA improves in every date interval, and the gains are largest for the most recent facts ("now": +0.180; 2026: +0.115). This shows that retrieval (rather than the backbone's internal knowledge) drives the gains on recent facts, which addresses my W1.

However, I am confused that some values in Table 5 appear inconsistent with the calculation of balanced accuracy. Let RR and FF denote real and fake accuracy. Following Table 1 (e.g., K-VERA: (0.887+0.622)/2=0.754(0.887 + 0.622)/2 = 0.754), BAcc is their mean, so R=2×BAcc−FR = 2 \times \mathrm{BAcc} - F. For K-VERA in 2016–2020 in your Table 5, this gives

R=2×0.879−0.714=1.044>1R = 2 \times 0.879 - 0.714 = 1.044 > 1

It's impossible, right? The same problem appears in 2000–2015 (K-VERA: 2×0.823−0.588=1.0582 \times 0.823 - 0.588 = 1.058; Qwen3.7-Plus: 2×0.790−0.412=1.1682 \times 0.790 - 0.412 = 1.168). And the rebuttal's example of gains ("+0.138 for 2016–2020") comes from one of these rows.

Could the authors clarify this?

We sincerely apologize for this error, and we thank you for checking Table 5 so carefully.

In Table 5, we mistakenly reported the overall accuracy in the BAcc. columns. Because real and fake images are not balanced within each date interval, the overall accuracy differs from the balanced accuracy. The previously reported overall accuracies are consistent with the per-interval numbers of real and fake images in Figure 10 (Required knowledge date distributions, already included in our original submission) and with the overall accuracy on the full K-AIGI-Bench, which almost equals the BAcc because the full benchmark is class-balanced. You are welcome to verify this. The Fake Acc. columns were correct and remain unchanged.

We have replaced these values with the correct BAcc in the revised paper (Table 5, L486–495). The table below compares the correct BAcc with the overall accuracy reported before:

DateK-VERA BAccK-VERA Overall AccQwen3.7-Plus BAccQwen3.7-Plus Overall Acc
<20000.6830.6850.6540.655
2000–20150.7500.8230.6730.790
2016–20200.8230.8790.6590.741
2021–20230.6830.6980.6050.638
2024–20250.7280.7560.6620.697
20260.8920.8860.7830.771
now0.8710.8710.6910.691
Overall0.7540.7540.6850.685

The "now" interval contains only AIGIs, so both values equal its fake accuracy.

With the correct BAcc, the conclusion that K-VERA outperforms its backbone in every date interval remains unchanged, and the largest gain is still for "now" (+0.180).

We apologize again for the confusion caused by this error.

Reviewer_1KgAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies "knowledge-intensive" AI-generated image (AIGI) detection: detecting synthetic images whose content conflicts with real-world facts even when no visual artifacts are present. It contributes (i) K-AIGI-Bench, about 1,000 images: 504 GPT-Image-2 images, each with one deliberately embedded knowledge inconsistency across eight topics, and 506 topic-matched real photographs; and (ii) K-VERA, a training-free multi-agent pipeline (Context Resolver, Planner, Actors, Adjudicator) that verifies implicit visual claims with web, image and reverse-image search and maps any verified inconsistency to "AI-generated". On K-AIGI-Bench, K-VERA reaches 0.754 balanced accuracy, compared with 0.685 for its tool-free backbone and 0.580 for the best dedicated detector. It also obtains higher LLM-judged and human-rated explanation scores. An OR-combination with DDA improves results on three in-the-wild benchmarks.

Critical strengths
  1. A useful complementary signal. Verifying whether depicted content can be true complements artifact-based forensics. The low correlation with DDA (Table 9) and the OR-combination gains (Table 4) support this.
  2. Explanation-level benchmark. Each fake carries a ground-truth inconsistency, which allows evaluating whether a detector identifies the correct reason. The matched tool-free baseline gives a controlled estimate of the retrieval gain.
Critical weaknesses
  1. Knowledge inconsistency does not imply AI generation. Photoshop composites, illustrations, film stills and satire routinely contain such inconsistencies, yet K-VERA labels every inconsistent image as AI-generated. If these images count as real, the false-positive rate on them is never measured, because the negative set contains only camera photographs. If they count as fake, the task is misinformation detection rather than AIGI detection.

  2. Web evidence is assumed to be reliable. The web increasingly contains fabricated and AI-generated content (Fig. 3 itself cites an "AI stock" image), and the method has no source verification. The benchmark cannot test this, because its fakes were never published online.

  3. The benchmark design largely determines the headline result. Every fake contains a search-verifiable inconsistency by construction, so the gain over dedicated detectors mainly reflects generator shift. The in-the-wild results are unexplained, and provenance lookup may contribute to them.

  4. Limited empirical rigor. There is no single-agent tool-using baseline, and all results come from single runs. The user-study mean of 4.76/5, with zero assigned to wrong verdicts, is incompatible with K-VERA's fake accuracy of 0.622.

Questions for authors
  1. Should a human-made image containing a knowledge inconsistency (Photoshop composite, illustration, CGI film still, satirical poster) be labeled real or AI-generated in your task? What is K-VERA's false-positive rate on a set of such images?
  2. How does accuracy change when fabricated supporting pages or AI-generated reference images are injected into the retrieved evidence? In the reported runs, what fraction of cited reference images are themselves AI-generated, and how was the "AI stock" image in Fig. 3 used?
  3. On the in-the-wild benchmarks, what are the real/fake accuracies and the verdict distribution, and how often do the cited sources include the image's original post or an explicit AI-generation label? K-VERA alone reaches 0.631 fake accuracy on self-reported GPT-Image-2 posts, which suggests that provenance lookup may contribute.
  4. Please report a single-agent ReAct baseline with the same tools and budget, tool-free results for the stronger backbones in Table 7, and repeated runs or confidence intervals for the ablations.
5 comments

We thank the reviewer for the detailed comments.

We would first like to clarify a general point. No AIGI detector is perfect; every method can produce false positives and false negatives (e.g., most vision-only detectors are not robust to degraded images). In practice, a method is effective when it clearly improves over existing baselines in the wild. K-VERA does so: combined with DDA, it achieves the highest score on all three in-the-wild benchmarks and outperforms either component alone (Table 4). We address the specific questions below.

Q1. Human-made images with knowledge inconsistencies. Under our definition, any image containing a knowledge inconsistency should be classified as Knowledge-Inconsistent (Sec. 4.1, L225–227). We additionally ran K-VERA on 25 Photoshop composites, 25 illustrations, and 25 CGI film stills (Appendix J, L1296–1333, Table 11, Fig. 13). Manual inspection explains the results:

  • Illustrations: 24/25 are judged Real or Unsure, as K-VERA recognizes them as human-created artworks.
  • CGI film stills: 24/25 are judged Real or Unsure, as K-VERA finds their source films.
  • Photoshop composites: 14/25 are flagged, mostly as Knowledge-Inconsistent, as most evidently contain implausible content.

These results are intuitive. In practice, K-VERA's verdicts can also be combined with other systems, e.g., with a binary detector's confidence rather than only an OR rule to reach a better overall decision.

Q2. Reliability of web evidence.

  • Among the 1,295 reference images cited in K-VERA's final outputs, only 6.5% (84) are classified as fake by DDA, many of which come from credible news media and official websites (Fig. 14). Meanwhile, the source contexts of 0.3% (4) indicate AI generation (Sec. 5.6, L504–508; Appendix K, L1376–1383).
  • Only 3.2% (32) of the 1,010 samples contain conflicting web evidence, i.e., two web-supported findings about the same entity that cannot both be true (Sec. 5.6, L502–504; Appendix K, L1346–1353, Table 12).
  • Thus, the retrieved web evidence is largely consistent, and K-VERA's reference images rarely include AI-generated images when searching for photorealistic images, the dominant style of K-AIGI-Bench.

The "AI stock" image in Fig. 3 was cited by the Adjudicator only as an example of a common generator error (a perching loon). The decisive evidence was the cited webpage stating that loons cannot perch.

Q3. In-the-wild verdicts and provenance. Table 14 (Appendix M, L1462–1476) reports the full verdict distributions. Although "AIGI" verdicts account for most correctly detected fake images, Knowledge-Inconsistent verdicts still contribute substantially: on So-Fake-OOD, they cover 10.5% of fake images (vs. 46.5% for "AIGI"), and on GPT-Image-2 in the Wild, 19.3%.

An "AIGI" verdict does not mean that K-VERA located the original post, which is impossible because search results matching the input are removed using crop-resistant perceptual hashes (Sec. 4.2, L248–250; Appendix G, L1118–1119). Manual inspection shows that, besides conspicuous semantic visual flaws, "AIGI" verdicts mainly rest on (Sec. 4.1, L229–231; Appendix M, Fig. 15):

  • very similar images found online that are explicitly labeled as AI-generated (the most common pattern);
  • generator watermarks, which these benchmarks retain.

Q4. Rigor.

  • Single-agent baseline. We added Qwen3.7-Plus w/ tools, a single agent with the same backbone and all five tools, to Table 1 (L337, L363–365). K-VERA is better in BAcc (0.754 vs. 0.739) and clearly better in explanations: GT-KI Recall 0.562 vs. 0.528 and higher EQS under all three evaluators (L398–403).
  • Other backbones. Table 9 (Appendix D, L820–833) adds tool-free and single-agent results for GPT-5.6-Sol, Gemini 3.5 Flash, and Claude Opus 4.8. K-VERA outperforms the tool-free setting with every backbone and the single agent with tools for two of the three additional backbones.
  • Repeated runs. Two additional runs on K-AIGI-Bench give 0.761 and 0.745 BAcc; with the reported 0.754, the mean is 0.753 and the standard deviation is 0.008. The decoding temperature is 0.01.
  • User study. Samples for which all three methods return incorrect verdicts are excluded, so a method with the highest fake accuracy can score close to 5. This differs from the computation of EQS; we have clarified this in the main text (Sec. 5.5, L469–470).

Please note that we have updated the paper. In the revised PDF, revised text is shown in blue; for newly added sections, only the section titles are blue. Line numbers here refer to the revised PDF.

Thank you for the response and the additional experiments. The repeated runs, the single-agent baseline and the user-study clarification address my concerns about rigor. However, the answers to Q1–Q3 confirm my main concerns, so I maintain my rating.

Q1. If every knowledge-inconsistent image is classified as fake, human-made composites and illustrations become positives, and the task is no longer AIGI detection. The new results also contradict this definition: illustrations and CGI stills judged Real are presented as correct. If human-made images are negatives instead, then 14/25 Photoshop composites are false positives. The paper should commit to one label and align its claims accordingly.

Q2. DDA detects only 22–31% of modern fakes (Tables 1 and 4), so it cannot reliably audit reference images. The low conflict rate shows that the evidence is consistent, not that it is correct. The contamination experiment remains missing.

Q3. Most correct in-the-wild detections rely on online images labeled as AI-generated and on generator watermarks, while Knowledge-Inconsistent verdicts cover only 10.5–19.3% of fakes. This is provenance and watermark lookup, so Table 4 provides little evidence for knowledge-intensive detection.

Q4. The gain over the single agent (0.754 vs. 0.739 BAcc) is modest and does not hold for all backbones. The contribution lies mainly in retrieval rather than in the multi-agent design.

AuthorsOpenReview ↗

Response to Reviewer 1KgA

Thank you for acknowledging that the repeated runs, the single-agent baseline, and the user-study clarification address your concerns about rigor. We further clarify Q1–Q4 below.

Q1. Label of human-made images. The AIGI benchmarks use origin-based labels. A verified knowledge inconsistency is one signal K-VERA uses to predict the label; it does not change the ground truth. Table 11 separately reports how the detector responds to human-made imagery: it flags 14/25 composites, 1/25 illustrations, and 1/25 CGI stills. A flagged human-made composite is not counted as a correctly detected AIGI.

This does not turn the task into misinformation detection. The fine-grained verdicts keep "Knowledge-Inconsistent" and "AIGI" separate (Sec. 4.1), and every benchmark we evaluate on is an AIGI benchmark. The false-positive rate on real images is also measured directly: real in-the-wild images from So-Fake-OOD and WildRF receive a Knowledge-Inconsistent verdict in only 4.4% and 3.0% of cases (Table 14, L1479–1486).

Q2. Reliability of web evidence.

  • Our conclusion does not rely on DDA. The primary check is the source context of each reference image, which indicates AI generation for 4 of the 1,295 images (0.3%, Appendix K, L1376–1383). DDA is only a supplementary check, and many of its flags are images from credible news media and official websites (Fig. 14). Its low fake accuracy (0.313 and 0.215; Tables 1 and 4) is measured only on GPT-Image-2 images; on So-Fake-OOD and WildRF, which cover other generators, DDA reaches 0.800 and 0.887 BAcc (Table 4).
  • Retrieval does not noticeably increase false positives on real images: K-VERA's real accuracy is 0.887 vs. 0.893 for the tool-free backbone (Table 1), which would not be the case if the retrieved evidence were frequently wrong.
  • The in-the-wild benchmarks already test K-VERA on the real, uncontrolled web, which contains fabricated and AI-generated content. There, fakes are rarely judged Real (8.5% on So-Fake-OOD, 4.9% on WildRF; Table 14). The Adjudicator also checks source quality and support (Sec. 4.3, L268–270).

Q3. In-the-wild verdicts.

  • Knowledge-Inconsistent verdicts alone cover 19.3% of the fakes in GPT-Image-2 in the Wild, nearly as many as DDA detects in total (0.215, Table 4). We also manually inspected 50 such fakes and confirmed at least one inconsistency identified by K-VERA in every one of them (Appendix M, L1474–1478, Fig. 16).
  • The controlled evidence for knowledge-intensive detection is K-AIGI-Bench, where every fake has a ground-truth inconsistency (Table 1). Table 4 serves a different purpose: it shows that this knowledge-based signal complements pixel-level detection in practice (L426–428).

Q4. Contribution of the multi-agent design.

  • The multi-agent design mainly improves evidence-grounded explanations, a central goal of our task: GT-KI Recall 0.562 vs. 0.528, and EQS 2.17/2.71/2.82 vs. 1.83/2.64/2.65 (Table 1).
  • Our main contribution is proposing knowledge-intensive AIGI detection and showing that it can be combined with conventional detectors to provide a complementary detection signal (Table 4).

We therefore respectfully disagree that these answers confirm the concerns: the labels are origin-based and unaffected by the knowledge-inconsistency signal, knowledge inconsistencies also contribute clearly to the recall of fakes in the wild, and the multi-agent design improves accuracy and explanations. Thank you again for the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.