acceptodds
Under review as a conference paper at ICLR 2027

FraudBench: A Multimodal Benchmark for Detecting AI-Generated Fraudulent Refund Evidence

Abstract

Artificial Intelligence (AI)-generated images have become increasingly realistic and readily adaptable to real-world claims, creating new challenges for verifying visual evidence. A concrete emerging risk is AI-generated refund fraud, in which manipulated or synthetic images are used to support claims about damaged products, poor delivery conditions, or service-related defects. Existing AI-generated image detection benchmarks mainly evaluate standalone authenticity classification, cross-generator transfer, or forensic localization, leaving claim-conditioned fraudulent evidence detection underexplored. To bridge this gap, we introduce , a multimodal benchmark for detecting AI-generated fraudulent refund evidence. is constructed from real-world user-review evidence across e-commerce, food delivery, and travel-service scenarios. We curate real evidence images together with their associated review and product metadata, identify genuine damaged and undamaged evidence through MLLM-assisted filtering and human annotation, and synthesize fake-damaged evidence from genuine undamaged reference images using 12 state-of-the-art image editing and generation models. Using , we evaluate MLLMs, specialized AI-generated image detectors, and human evaluators on the same evaluation images. Experiments show that current MLLMs often recognize real-damaged evidence but fail on many fake-damaged subsets, with fake-damage detection rates (TPR) far below the 50% baseline on most generator subsets. Specialized detectors generally perform better but remain inconsistent across generators and can produce false positives on real-damaged samples, revealing a clear gap between generic AI image detection and reliable claim-conditioned refund-evidence verification. An anonymized version of is available at https://dataverse.harvard.edu/previewurl.xhtml?token=0bf91c2e-daeb-408f-b08d-0826200fcc4d.

Primary area:
datasets and benchmarks
Keywords:
Multimodal Benchmark, Fraudulent Refund Evidence, AI-Generated Image Detection, MLLM Evaluation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ASFdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FraudBench, a multimodal benchmark for detecting AI-generated fraudulent refund evidence in e-commerce and related transaction scenarios. The authors collect real-world review data from four sources and use 12 image generation and editing models to synthesize fraudulent damage evidence, along with corresponding fake review texts. The benchmark further defines five evaluation dimensions to assess the reliability of detection methods under different settings. The authors evaluate MLLMs and specialized AI-generated image detectors. Experimental results show that MLLMs generally preserve genuine evidence but fail to detect most synthetic damage, while specialized detectors achieve better fake-image detection performance but exhibit substantial sensitivity to generators and model checkpoints, with some producing high FP rates on genuine evidence.

Critical strengths
  • FraudBench addresses a practical and important refund fraud problem in real-world e-commerce scenarios.
  • The benchmark goes beyond single-image detection by introducing a claim-centered perspective on evidence verification. Images submitted in refund scenarios should be evaluated in relation to the specific claims they support, rather than being treated merely as independent Real/Fake images.
  • The benchmark considers several important aspects of realistic refund scenarios, including diverse review contexts (e.g., travel services, food delivery, dine-in, and pickup), multilingual reviews, and both single-image and multi-image.
  • The synthetic evidence generation pipeline is well-motivated. Rather than generating fake-damaged samples arbitrarily, the authors first mine damage patterns from real damaged evidence and then synthesize defects based on these patterns. This improves the reliability of FraudBench.
Critical weaknesses
  • The evaluation metrics mainly focus on the final verdict, with limited attention to the quality and faithfulness of the explanations generated by MLLMs. In practical refund verification, whether the provided explanation accurately supports the decision is also important.

  • The authors argue that refund fraud detection differs from conventional AI-generated image detection by additionally requiring multi-view visual consistency and alignment between visual evidence and the user's stated claim. However, the fake review texts are generated based on the corresponding synthetic damage evidence, making the images and claims generally aligned by construction. Consequently, the benchmark does not sufficiently evaluate cases where visual evidence contradicts the user's claim. Although Appendix D.4.3 includes an image-text mismatch experiment, the mismatched reviews are sampled from different product categories, rather than more challenging cases involving the same product but inconsistent damage claims. This somewhat limits the distinction between the proposed task and conventional AI-generated image detection.

  • Specialized detectors such as CO-SPY [ProGAN] and ForgeLens [GenImage] achieve relatively strong results without using textual information, while MLLMs show only modest improvements when review text is provided. This raises the question of whether the main challenge in fraudulent refund evidence detection still lies in visual authenticity detection, and to what extent claim-centered information provides additional benefits. A more detailed analysis of the contribution of claim information would help better establish the value of the proposed claim-centered evaluation framework.

Questions for authors

As discussed in the second and third points of Critical Weaknesses, I have concerns about the necessity of a claim-centered perspective in fraudulent refund evidence detection, as the authenticity of the evidence images appears to be more important.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ASFd (1/2)

We sincerely appreciate your recognition of our contributions, in particular the practical importance of the refund-fraud problem and the claim-centered perspective. Your questions led to some of the most informative new experiments in this rebuttal. We would like to address each concern below.


Weakness 1: Quality and Faithfulness of Explanations

"...In practical refund verification, whether the provided explanation accurately supports the decision is also important."

Response 1:

Thank you for this suggestion. We have added two analyses of explanation quality, and both show that most correct verdicts point at the actual edit. Together with the localization study in the appendix, explanations are now evaluated in three ways:

1. Explanations Name the Right Damage (New)

For fake-damaged images that a model correctly flags (single image, no review text), we checked whether its explanation names the damage type that was actually synthesized. Across seven models, this holds for 68% to 83% of correct verdicts (76% on average). When the damage labels are shuffled among the same model's correct verdicts within each category, the rate falls to 37% to 51% (44% on average).

2. Explanations Point to the Edited Region (New)

We sampled 300 of these correct verdicts (50 per generator) and outlined the edited region on each image. Two MLLM judges that are not among the seven models (GPT-6 Sol and Qwen3.8-Flash) then decided whether the explanation refers to something inside the outline. As controls, the same explanation was checked against the same outline moved to another part of the image, and the true outline was checked against the explanation of a different image:

The Explanation Refers to Outlined RegionYes or PartlyYes
True Edited Region, Own Explanation0.8970.513
Same Outline Moved Elsewhere (Control)0.3070.025
True Edited Region, Explanation of Another Image (Control)0.2450.025

The two judges agree on 96% of the verdicts (Cohen's κ = 0.79). The rate is high for every generator (0.77 to 0.97, against 0.16 to 0.46 for the moved outline), and lowest for GPT Image 2, which is also the hardest generator to detect. The judges also find the cited evidence visible in the image for 98% of the explanations, against 29% when the explanation of another image is used. Most correct verdicts therefore point at the edited defect, not only at generic artifacts.

3. Localization (Appendix D.4.7)

Boxes predicted by Gemini 3 Flash reach a mean IoU of 0.208 with the edited region, compared with 0.087 for FakeShield masks on the same 786 images, although part of this advantage comes from the boxes being coarser.

We would add these explanation metrics to the evaluation section.


Weakness 2: Image–Claim Alignment by Construction

"...the mismatched reviews are sampled from different product categories, rather than more challenging cases involving the same product but inconsistent damage claims. ..."

Response 2:

We ran the same-category mismatch you suggested. MLLMs ignore the claim when asked whether an image is authentic, but they check it well when asked directly whether the photos support it.

1. Contradicting Claims Under the Authenticity Question

Every review of both classes receives a genuine 1-star complaint written for another product in the same category. The complaint describes either the same damage type as the image (matched) or a different one (contradicting). Since both classes draw from the same pool of genuine text, the text carries no label information, and the two conditions differ only in whether the claim agrees with the image. The single-image results cover 410 images (364 fake and 46 real) from the paper's stratified sample with 12 MLLMs, and the multi-image results cover all 1,668 reviews with at least two images with the seven main MLLMs. Both tables report balanced accuracy; the "None" rows are the runs reported in the paper, and the other rows are new runs from the same week.

Single Image:

Complaint Shown with ImageGPT-6 AstraGPT-6 SolGPT-6 LunaGPT-5.6 TerraGrok 4.6Kimi K3DeepSeek V4.1 FlashQwen3-VL-PlusQwen3.5-Omni-PlusQwen3.5-Omni-FlashQwen3.8-27BQwen3.8-FlashAverage
None0.6550.6140.6460.6350.5740.5680.6790.5230.5520.5690.6070.6700.608
Genuine, Other Category0.6860.6750.6000.6460.5730.5620.6970.5170.5460.5250.5880.6340.604
Genuine, Same Category, Matched Damage0.7000.6720.6000.5970.5680.5580.7010.5130.5480.5300.5860.6590.603
Genuine, Same Category, Contradicting Damage0.6960.6190.5990.6270.5810.5590.6630.5140.5630.5340.5880.6620.600
AuthorsOpenReview ↗

Response to Reviewer ASFd (2/2)

Multiple Images:

Complaint Shown with ImagesGPT-6 AstraGPT-5.6 TerraGemini 3.8 FlashGrok 4.6DeepSeek V4.1 FlashKimi K3Qwen3.8-MaxAverage
None0.7730.6490.6120.5700.6370.5630.6940.642
Genuine, Other Category0.7700.6530.6100.5700.7150.5710.6670.651
Genuine, Same Category, Matched Damage0.7810.6350.6150.5670.7050.5640.6410.644
Genuine, Same Category, Contradicting Damage0.7850.6480.6110.5660.7180.5700.6550.651

Contradicting claims barely change the result (0.600 vs. 0.603 with single images, and 0.651 vs. 0.644 with multiple images). With single images, they make three models somewhat more suspicious of both classes (TNR from 0.992 to 0.970 and TPR from 0.214 to 0.229 on average), which does not help separate them. When asked whether an image is authentic, current MLLMs do not check it against the claim.

2. Claim Verification with Same-Product Contradicting Claims

When asked directly, MLLMs check the claim well. Each of 208 real-damaged reviews is shown twice with the same photos: once with its complaint reworded by an LLM, and once with the complaint rewritten by the same LLM to describe a clearly different type of damage to the same product (for example, a cracked drawer claimed to be burnt). Reviews whose complaint names several damage types were excluded. The models are asked whether the photos show the problem described.

Since all photos are authentic and both complaints are LLM-written, neither an image-only detector nor a writing-style cue can do better than chance (0.5). Across 14 MLLMs, balanced accuracy is 0.871 on average (0.789 to 0.944) with an AUROC of 0.921: the models accept 91% of the matching complaints and reject 84% of the contradicting ones. We will add this control to the benchmark as a claim-verification track.


Weakness 3: Is Visual Authenticity the Main Challenge?

"...This raises the question of whether the main challenge in fraudulent refund evidence detection still lies in visual authenticity detection, and to what extent claim-centered information provides additional benefits. ..."

Response 3:

Yes, visual authenticity remains the main challenge, and the claim adds a different kind of benefit.

For deciding whether the damage is AI-generated, adding a genuine claim changes accuracy by less than 0.01. For deciding whether the photos support the claim, the claim gives image-only detectors nothing, while it raises MLLMs from chance to 0.871. Balanced accuracy with and without the claim, on the same images (from the tables in Response 2):

QuestionModelsImage OnlyImage + ClaimGain from the Claim
Is the Damage AI-Generated? (Single Image)12 MLLMs0.6080.603−0.005
Is the Damage AI-Generated? (Multiple Images)7 MLLMs0.6420.644+0.002
Do the Photos Show the Claimed Problem?Specialized Detectors0.500Cannot Read the Claim--
Do the Photos Show the Claimed Problem?14 MLLMs0.5000.871+0.371

1. Visual Authenticity Is the Main Challenge

In the first two rows, each image receives a genuine 1-star complaint from the same category that describes the same damage type, so the text carries no label information. Accuracy does not change, so deciding whether the damage is fabricated depends on the images. This image task remains hard: in the paper's single-image results (Table 1), the 16 MLLMs reach a balanced accuracy of only 0.584 on average and detect only 23.0% of the fake-damaged images.

2. The Claim Adds a Check That Only MLLMs Can Perform

In the claim-verification control, the same authentic photos are shown once with a matching and once with a contradicting complaint, so without the claim the two cases cannot be told apart (0.500 by construction). With the claim, 14 MLLMs reach 0.871 on average (AUROC 0.921) over 208 reviews. Specialized detectors take only the image, so the claim gives them no benefit.

We will state this scope clearly in the revision and add the claim-verification control to the benchmark as its own track.


Thank you again for your valuable comments! We hope that our clarifications and new results address your concerns, and we would be glad to continue the discussion if any questions remain.

Authors of Submission #6087

I appreciate the authors' clarification and the additional experiments. However, the submitted PDF appears unchanged: it does not include these new results or the corresponding revisions. As these additions are essential to the paper's completeness. This may affect my rating.

Reviewer_2Q3AOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

FraudBench constructs synthetic damaged-product and service evidence from real review images, pairing it with textual complaints and evaluating MLLMs, specialized detectors, and humans. It finds substantial difficulty detecting convincing synthetic manipulation.

Critical strengths
  1. This paper employs genuinely damaged images as negative examples, making the task more meaningful than distinguishing synthetic damage from pristine products.
  2. The paper reviews text and multiple photographs, which support richer evidence-verification questions.
  3. The supporting analyses are very detailed and convincing.
Critical weaknesses
  1. Image authenticity, consistency with a complaint, and whether a refund claim is fraudulent are different questions. The benchmark seems to mix these together. Please define more precisely what the benchmark measures.
  2. In the introduction, it is claimed that the synthesis pipeline "forces" cross-modal and multi-view reasoning. But the Appendix D.4.5 acknowledges that real and synthetic images differ in size and format, potentially influencing detector scores. The claim that performance measures semantic verification rather than properties of the construction pipeline should be further considered.
  3. Low fake-image recall alone does not establish below-chance classification. The broad failure narratives, such as below random, should be corrected.
Questions for authors

Please refer to the weakness section.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2Q3A (1/2)

We sincerely thank the reviewer for the precise and constructive comments, and for finding our supporting analyses detailed and convincing. We would like to address each concern below.


Weakness 1: What FraudBench Measures

"Image authenticity, consistency with a complaint, and whether a refund claim is fraudulent are different questions. The benchmark seems to mix these together. ..."

Response 1:

Thank you for drawing this distinction. FraudBench measures the first of the three questions: whether the damage shown in the images is genuine or produced by AI editing. We will make this explicit with a task definition in Section 3:

"Given the evidence images EE of one review, optionally together with the claim text cc, decide whether the damage shown in EE was produced by AI image editing or generation (fake-damaged) or is genuine (real-damaged)."

The three questions relate to the benchmark as follows:

1. Image Authenticity: The Label of the Benchmark

This is the label of every setting, and it applies to both MLLMs and specialized detectors.

2. Consistency with the Complaint: Applies to Different Models

Specialized detectors take only the image as input, so they cannot use the complaint at all. MLLMs read the complaint and can check whether the photos agree with it. In a new claim-verification control, the same authentic photos are shown once with a complaint that matches the damage and once with a complaint rewritten to describe a clearly different damage to the same product, and models are asked whether the photos support the complaint. Since the photos are identical, any image-only detector stays at chance level by construction, whereas 14 MLLMs reach a balanced accuracy of 0.871 on average (AUROC 0.921) over 208 reviews.

3. Whether a Refund Claim Is Fraudulent: Outside Our Scope

This also involves intent, eligibility, and platform policy, and an authentic photo can still come with a false claim.

Accordingly, we will name the main task "detection of AI-fabricated damage evidence" and report the claim experiments as a separate claim-verification track.


Weakness 2: "Forces" Versus the Size and Format Difference

"...The claim that performance measures semantic verification rather than properties of the construction pipeline should be further considered."

Response 2:

Thank you for raising this. We tested it directly by changing the size and the format of both classes in the same way: the MLLM results change little, while the specialized detectors change a lot. The MLLM results therefore reflect the image content, whereas the detector scores depended on file properties, as Appendix D.4.5 suspected.

On a stratified subset of 579 images (498 fake-damaged and 81 real-damaged), we resized both classes to a 1024-pixel long edge and saved them losslessly (PNG), re-saved both classes as JPEG (quality 90) at their original size, or did both, and ran the same models on the same files in the same week. Balanced accuracy:

ModelOriginal FilesResized Only (PNG)Re-saved Only (JPEG)Both (FraudBench-N)Largest Change
GPT-6 Astra0.6400.6120.6340.6020.038
GPT-5.6 Terra0.6250.6080.6220.6030.022
GPT-6 Luna0.6470.6370.5870.5570.090
DeepSeek V4.1 Flash0.6380.5720.6570.5540.103
Qwen3.8-Max0.6700.6530.6410.6500.029
MLLM Average0.6440.6160.6280.5930.051
ForgeLens [GenImage]0.8980.8310.4930.5000.405
ForgeLens [ProGAN]0.8160.7240.6190.5480.268
Effort [SD-v1.4]0.7770.7900.5720.4950.295
CO-SPY [ProGAN]0.7880.7750.5330.5250.263
IAPL [SD-v1.4]0.7170.7450.6110.5910.154

The other three detector checkpoints are weak in all four versions (0.433–0.642).

AuthorsOpenReview ↗

Response to Reviewer 2Q3A (2/2)

1. MLLMs Change Little

Across the four versions, each MLLM changes by 0.022 to 0.103 and their average by 0.051, whereas the five detectors change by 0.154 to 0.405. The multi-view gains do not depend on file properties either: on FraudBench-N and without any review text, using several photos raises the average of seven MLLMs from 0.577 (single image) to 0.614 (multi-image) and 0.686 (multi-step).

2. Why the Detectors Change

The detectors do not see the images at their original size: Effort resizes every image to 224×224 pixels, and IAPL to 256×256 before a 224-pixel center crop. Resizing alone therefore barely changes them (Effort 0.777 → 0.790, IAPL 0.717 → 0.745, CO-SPY [ProGAN] 0.788 → 0.775). ForgeLens takes a 224-pixel crop at full resolution, so resizing changes the scale it sees and lowers it somewhat (0.898 → 0.831). Re-saving both classes as JPEG, however, brings all of them down to 0.493–0.619. In the original files, the real-damaged photos were compressed by the platforms, while four of the six generators return uncompressed PNG files; saving both classes losslessly keeps this difference, and re-saving both as JPEG removes it. The detectors' high scores therefore came from this difference in file compression rather than from the visible damage.

In the revision, we will rephrase the "forces" sentence in the introduction, report this test, and add FraudBench-N, which applies both changes to the full benchmark, as an evaluation protocol.


Weakness 3: Low Recall Versus Below Chance

"Low fake-image recall alone does not establish below-chance classification. The broad failure narratives, such as below random, should be corrected."

Response 3:

We agree, and we will correct this. Low recall alone does not mean below-chance classification. In the revision, we will:

  • remove "below random guessing" and "far below the 50% baseline" from the abstract, introduction, and conclusion;
  • describe the MLLM results as a conservative decision threshold (high TNR, low TPR) rather than a failure to separate the two classes;
  • report balanced accuracy and threshold-free metrics (AUROC) alongside TPR.

Thank you again for your valuable comments! We would be grateful to know whether the results above address your soundness concern, and we would be glad to continue the discussion if any questions remain.

Authors of Submission #6087

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

The paper introduces FraudBench, a benchmark for detecting AI-generated “fake-damage” evidence submitted with refund claims. Real review images are drawn from Amazon Reviews 2023 (590 reviews, 27 categories), Trip.com and GrabFood (225 reviews), plus 7 self-collected records. Using star rating as the initial signal, followed by MLLM screening and verification by two of the authors, the data are split into 988 real-undamaged images (422 reviews) and 1,012 real-damaged images (400 reviews). Each real-undamaged image is edited by 12 commercial image editing/generation models with an MLLM-constructed, category-specific damage prompt, yielding 11,856 fake-damaged images, each paired with an LLM-written fake complaint. The task is binary: fake-damaged = positive, real-damaged = negative. Evaluation covers five “dimensions”: single- vs. multi-image input, review-free vs. review-conditioned, multi-step (per-image turns), prompt sensitivity, and real-image preservation (TNR). The authors evaluate 30 API-based MLLMs, 4 AI-generated-image detectors (8 checkpoints), 4 image-forgery detection/localization (IFDL) methods, and 3 human evaluators. Headline findings: MLLMs are “over-credulous”: average TNR 0.938 but average fake-damage TPR 0.230 in the single-image setting. Specialized detectors are stronger but checkpoint-dependent (ForgeLens [GenImage] reaches 0.904 balanced accuracy; Effort [Chameleon] has 0.051 TNR). Multiple images and multi-step prompting help more than review text. Humans reach 0.686 balanced accuracy. The appendix adds machine-derived edit masks, threshold-free detector metrics, a localization study, an image–text mismatch control, and a complementarity analysis. Every reported evaluation uses 6 of the 12 generators (6,940 images: 5,928 fake + 1,012 real).

Critical strengths
  1. A well-posed task with a hard, realistic negative class and a controlled synthesis design. Unlike generic real-vs-fake benchmarks, the negatives are genuinely damaged user photos, so a detector cannot simply equate “damage” with “fake”. Each fake comes from a real undamaged photo, and a single edit instruction is shared across all generators. This enables controlled per-generator comparisons and lets the authors derive region-level masks from source/edit residuals (Appx. C.4). This matters to my assessment because claim-conditioned refund evidence is a real and under-studied threat, and the paired design is a reusable asset even if the current evaluation needs repair.
  2. Broad, transparent, and internally consistent evaluation. The evaluation spans 30 MLLMs, 8 detector checkpoints, 6 IFDL checkpoints and humans. The paper documents exact model IDs and endpoints, every prompt, the retry/downscale protocol, and API call counts. The threshold-free analysis (Tab. 58) usefully corrects the default-threshold picture: IAPL [SD-v1.4] falls from 0.956 TPR to 0.275 at 5% FPR, and Effort [Chameleon] has AUROC 0.399. I spot-checked Tab. 1: the 16-model averages (TNR 0.938, TPR 0.230; per-generator TPRs 0.104 / 0.144 / … / 0.387) and per-model balanced accuracies reproduce from the reported per-generator values. As a dated snapshot of current-model behavior, this is valuable.
Critical weaknesses
  1. Uncontrolled non-semantic differences between the classes undermine what the benchmark measures. • File-level properties. The positives are generator outputs that re-render the whole frame at native resolutions (e.g., 1536×1024 in Fig. 24). The negatives are platform JPEGs (Amazon SL1600 images, filtered to ≥1024 px). The authors concede that detector scores “may partly reflect” size and format differences (D.4.5). Yet nothing in the paper normalizes for this (e.g., resizing and JPEG re-encoding both classes identically), and there is no trivial metadata-only baseline. The introduction claims the pipeline prevents shortcuts such as “resolution drops”, but this is never shown. The specialized-detector results (e.g., ForgeLens [GenImage] at 1.000 TPR on Nano Banana 2 and Qwen-Image-2.0) and the conclusion that “detectors are stronger than MLLMs” may partly reflect this cue. In addition, the MLLM downscaling protocol (D.1.1) triggers on payload size, so the two classes can receive different re-encoding. • Review text. Fake complaints are LLM-generated; genuine ones are human 1-star reviews. The mismatch control (D.4.3) shows the single-image gain from review text disappears once both classes receive genuine reviews, and the authors acknowledge it may reflect style. In the multi-image setting, however, review text adds much more: GPT-6 Astra goes from 0.773 to 0.930, DeepSeek-V4.1-Flash from 0.637 to 0.748, and Gemini 3.8 Flash from 0.612 to 0.742 (Tab. 14 vs. 15). The text-style control was never run there. • Multi-image construction. Damage selection runs independently per image (C.3), so different photos in one fake review can show different damage types and locations. Cross-view inconsistency then becomes an easy tell, which a real fraudster would avoid. This likely inflates the reported multi-image and multi-step gains. • Source content. Positives come from photos in 5-star reviews and negatives from 1-star reviews (B.4), so photographic style and scene type differ systematically between the classes.
  2. Several headline claims conflict with the paper’s own results and scope. • The abstract and introduction advertise 12 generators and 12,868 evaluation images. All reported evaluations use 6 generators and 6,940 images. The six later-added generators are never evaluated, including GPT Image 2.5, which the paper calls “a recent milestone in photorealistic image synthesis”. • The paper claims “human evaluators outperform all evaluated MLLMs”. The human figure (0.686) is single-image and review-free only. GPT-6 Astra reaches 0.930 balanced accuracy with TNR 1.000 in the multi-image review-conditioned setting (Tab. 15), above both humans and the best specialized detector (0.904). Yet GPT-6 Astra is left out of the “representative” models in Tab. 2. Even in the single-image setting, the margin over Qwen3.8-Max (0.664) is small. Individual humans range from 0.610 to 0.772, and the size of the human-evaluation subset is not reported. • The paper frames MLLM TPR as “far below the 50% baseline”. TPR alone is not comparable to random guessing: every MLLM in Tab. 1 except GPT-6 Luna has balanced accuracy above 0.5. “Over-credulity” mostly describes a conservative operating point. A score-based comparison (e.g., AUROC from the self-reported confidence) is needed before concluding that MLLMs cannot separate the classes. • Fig. 1 says generated evidence is “further verified by human annotators”, but C.5 says humans examine only MLLM-flagged cases, and the region annotations are marked human_verified: false. No edit-success rate is reported. Failed edits that are labeled fake but show no visible damage would be counted as MLLM “misses”.

Post-Rebuttal Update: In their rebuttal, the authors introduced FraudBench-N (normalizing resolution and JPEG compression), ran the multi-image text-style controls, evaluated the remaining six generators, and removed user identifiers from the metadata. These new results resolve my main soundness concerns, though the main text still needs to be updated in the camera-ready version to reflect the reversed detector-vs-MLLM findings.

Questions for authors

Clarification requests:

  1. 6 vs. 12 generators. Why are only six generators evaluated? Will results for the other six (GPT Image 2.5 Flare/Sunburst, Grok Imagine Image 2.0, Nano Banana 2 Lite, Qwen-Image-3.0/-Pro) be reported? If not, please correct the counts in the abstract and introduction.
  2. File-level confounds. Please report the resolution, aspect-ratio, format and JPEG-quality distributions per class and per generator. Were generator outputs re-encoded to JPEG (the filenames are .jpg)? Please also report (a) the main results after normalizing both classes to a common size and JPEG quality, and (b) a metadata-only baseline (e.g., a logistic regression on width, height and estimated JPEG quality). If such a baseline achieves high balanced accuracy, the detector conclusions change.
  3. Multi-image fakes and review text. In a fake multi-image review, is every image edited, and are damage types chosen independently per image? What review text is given in the multi-image setting: one comment, or a concatenation of per-image comments? Please run the genuine/mismatched-review control in the multi-image setting and explain GPT-6 Astra’s jump from 0.773 to 0.930 (Tab. 14 → 15).
  4. Edit verification. What fraction of the generated images was inspected by a human? What is the per-generator edit-failure rate (no visible damage, unchanged image, wrong object)?
  5. Human study. How many images were evaluated, and how were they sampled? What is the inter-rater agreement? Given Tab. 15, please qualify the “outperform all MLLMs” claim.
  6. Model selection. What was the criterion for the “representative” MLLMs in Tab. 1–3, and why is GPT-6 Astra absent from Tab. 2?
  7. Negative-label noise. The real-damaged photos come from 2025–2026 platform reviews, which could themselves contain AI-edited fraud — the paper’s own premise. How was this ruled out? Suggestions for future work (not decisive): add mixed real/fake multi-image reviews (an attacker who edits only one of several photos), add claim–image mismatch hard negatives, and provide a train/test split so that fine-tuned detectors can be evaluated.
6 comments
AuthorsOpenReview ↗

Response to Reviewer eqnu (1/5)

We sincerely thank the reviewer for the careful and detailed review, and for pointing us to the exact sections involved, which made it much easier for us to check each concern. We would like to address each concern below.


Weakness 1.1 (Question 2): File-Level Properties

"...Yet nothing in the paper normalizes for this (e.g., resizing and JPEG re-encoding both classes identically), and there is no trivial metadata-only baseline. ..."

Response 1.1:

The two classes did differ in file format. This explains the strong results of the specialized detectors, but not those of the MLLMs. To answer Q2, generator outputs were stored exactly as returned and were not re-encoded. Real-damaged images are platform JPEGs (99%, median quality 75), four of the six generators return PNG data, and no image size appears in both classes. The metadata-only baseline you proposed (a logistic regression on width, height, and estimated JPEG quality, with leave-one-category-out cross-validation) reaches a balanced accuracy of 0.987. Following your suggestion, we then normalized both classes identically, which turned out to be one of our most informative new experiments:

1. FraudBench-N Removes the File-Level Difference

Every evaluated image is decoded, brought to a standard aspect ratio, resized to a 1024-pixel long edge, and saved as a quality-90 JPEG without metadata. On FraudBench-N, the same metadata baseline drops to 0.583, and to exactly 0.500 on the 768×1024 subset (754 real-damaged and 3,479 fake-damaged images), where all images share the same size, format, and quality.

2. All Eight Detector Checkpoints Fall to Chance Level

We re-ran all eight checkpoints with the official code. On the original files, the re-run reproduces Table 1 image by image (r = 1.000 for ForgeLens, Effort, and IAPL; CO-SPY is reproduced exactly with the code version used for the paper). Applying the two normalization steps separately shows where the detectors' advantage came from:

Detector (Balanced Accuracy)Original FilesJPEG Re-encoding OnlyResizing OnlyFraudBench-N
ForgeLens [GenImage]0.9040.5050.8260.498
ForgeLens [ProGAN]0.8160.6060.7100.541
Effort [SD-v1.4]0.7950.5880.8010.502
CO-SPY [ProGAN]0.7830.5360.7770.518
IAPL [SD-v1.4]0.7270.6240.7420.597
All Eight Checkpoints0.452–0.9040.455–0.6240.454–0.8260.472–0.597

JPEG re-encoding alone brings them to chance level, while resizing alone barely matters. This matches a known bias of these detectors: their training sets store real images as JPEG and generated images as PNG, so they learn compression as a cue [1, 2], and JPEG compression is known to sharply lower their recall [3, 4].

3. MLLMs Are Barely Affected

To compare like with like, we ran 16 MLLMs on both versions in the same week (the 15 Table 1 models that are still available, plus GPT-5.6 Terra). Their average balanced accuracy moves only from 0.596 to 0.571, and TNR stays the same (0.987 vs. 0.984). On FraudBench-N, five MLLMs outperform every detector checkpoint (best: Qwen3.8-Max, 0.632). The gap is largest at the low false-positive rates that refund platforms need: at a 5% false-positive rate, GPT-6 Astra detects 39.0% of fake-damaged images (GPT-6 Sol 38.6%, GPT-5.6 Terra 36.6%), against at most 15.7% for any detector checkpoint.

4. Changes in the Revision

The detectors' advantage on the original files comes from compression and disappears once the photos are re-encoded, which is what refund platforms do to uploaded photos. Without this cue, MLLMs are the strongest model family, and the task still leaves clear room for improvement. We will add FraudBench-N as an evaluation protocol, rewrite the detector-versus-MLLM conclusion based on it, and remove the introduction's statement about shortcuts.

5. Downscaling Protocol

Downscaling is applied only after a provider rejects a request or drops the connection, so it affects few requests. Even counting every re-send (which also includes temporary errors), it affected 0.7% of fake-image and 0.6% of real-image requests in the single-image setting, and 2.5% of fake reviews and 0.03% of real reviews in the multi-image setting without review text. FraudBench-N images are far below every provider's size limit, so this difference no longer exists there.


AuthorsOpenReview ↗

Response to Reviewer eqnu (2/5)


Weakness 1.2 (Question 3): Review Text

"...In the multi-image setting, however, review text adds much more ... The text-style control was never run there."

Response 1.2:

We ran the control you suggested in the multi-image setting. The gain from review text came from how the prompt was built and from the style of the generated complaints, while the gain from using several photos holds without any text. To answer Q3, in the multi-image prompt a fake review lists one generated comment per image, while a genuine review has one shared review text. The multi-step dialogue likewise adds a new comment at each turn only for fake reviews. Balanced accuracy of the seven main MLLMs on the 1,668 reviews with at least two images (the first two rows are the runs reported in the paper, and the last two are new runs from the same week):

Review Text (Multi-Image Setting)GPT-6 AstraGPT-5.6 TerraGemini 3.8 FlashGrok 4.6DeepSeek V4.1 FlashKimi K3Qwen3.8-MaxAverage
No Review0.7730.6490.6120.5700.6370.5630.6940.642
Original Reviews0.9300.6880.7420.5910.7480.6480.6890.719
Same Structure for Both Classes0.8850.6710.7340.5950.7440.6030.6650.700
Genuine Reviews from Another Category for Both Classes0.7700.6530.6100.5700.7150.5710.6670.651

1. The Text Gain Came from Prompt Structure and Writing Style

With the same structure for both classes, GPT-6 Astra drops from 0.930 to 0.885. With genuine reviews for both classes, it returns to 0.770, the same as without any review, and the seven-model average returns to 0.651 (0.642 without any review). In line with this, 98% of Astra's correct fake-review verdicts in the original setting cite the comments. Writing style is a cue as well: generated complaints are shorter than genuine reviews (median 17 versus 34 words), and a rule based on length alone reaches 0.870. With single images, the genuine-review control gives 0.604, against 0.608 without any review (12 MLLMs).

2. The Multi-Image Gain Does Not Depend on Text

Without any review text, using several photos raises the seven-model average on FraudBench-N from 0.577 (single image) to 0.614 (multi-image) and 0.686 (multi-step), close to the gains on the original files (+0.043 and +0.108). Multi-image improves all seven models, and multi-step improves six of them.

In the revision, we will base the multi-view conclusions on the review-free settings and report the review-conditioned results together with these controls. To test whether models use the claim itself without such cues, we have also added a claim-verification control: the same authentic photos are shown once with a matching and once with a contradicting complaint, and 14 MLLMs tell them apart with a balanced accuracy of 0.871 on average, while an image-only detector is at chance. Details are in our response to Reviewer ASFd (Weakness 2).


Weakness 1.3 (Question 3): Multi-Image Construction

"...Cross-view inconsistency then becomes an easy tell, which a real fraudster would avoid. ..."

Response 1.3:

This is a fair concern, and we checked it: inconsistency across images helps the models only slightly, and the multi-image gain holds on reviews whose images all show the same damage.

1. Both Cases Occur in Practice

A careful fraudster may well keep the damage consistent, and we evaluate this case separately below. Inconsistent evidence, however, also occurs in real fraud. In a recent case in China (China Daily, December 4, 2025), a buyer claimed that six of eight live crabs had arrived dead and submitted two AI-generated videos: one showed two male and four female crabs, the other three of each, although four of each had been shipped. The refund was still paid, the discrepancy was noticed only afterwards, and the police later confirmed that the videos were synthetic.

2. Inconsistent Reviews Are Only Slightly Easier

Every image of a fake review is edited, and the damage type is chosen independently for each image (Appendix C.3), so 46% of the 239 multi-image source reviews contain more than one damage type. Without review text and averaged over models, review-level TPR on these reviews is higher than on consistent ones by only 0.019 (multi-image) and 0.038 (multi-step).

3. The Multi-Image Gain Holds on Consistent Reviews

On the 129 source reviews whose images all share one damage type (774 fake reviews across six generators), the six Table 2 models still reach a multi-step TPR of 0.411, compared with 0.210 per image with single images, while keeping a TNR of 0.977 on genuine reviews. We will release this subset as a separate split and mark damage consistency for every review, so that both cases can be evaluated directly.


AuthorsOpenReview ↗

Response to Reviewer eqnu (3/5)


Weakness 1.4: Source Content

"...Positives come from photos in 5-star reviews and negatives from 1-star reviews (B.4), so photographic style and scene type differ systematically between the classes."

Response 1.4:

The models react to the edit, not to the photo style: they flag only 3.2% of the unedited source photos, against 15.8% of the fakes made from them. The difference in style follows from the task itself: genuine damage is photographed by customers who complain, while an edit has to start from an undamaged photo, and each edit keeps the scene of its source. We will describe this difference in the limitations.

To test this, we ran 16 MLLMs on the 988 real-undamaged source photos in FraudBench-N. These photos have exactly the style of the fakes but contain no edit. The models flag 3.2% of them as AI-generated, close to the 1.6% on real-damaged images and far below the 15.8% on fake-damaged images.


Weakness 2.1 (Question 1): Six Versus Twelve Generators

"The abstract and introduction advertise 12 generators and 12,868 evaluation images. All reported evaluations use 6 generators and 6,940 images. ..."

Response 2.1:

The paper evaluates a representative six of the twelve generators, and following your question, we have now evaluated the other six as well. We keep FraudBench up to date as new image generators are released (now 12 generators; 12,868 real-damaged and fake-damaged images), and we add new detectors over time. Re-running every model on every new generator is not feasible within our compute and API budget, so the paper evaluates a representative subset (6 generators; 6,940 images). The abstract and introduction will state this distinction explicitly.

The new results support the description of GPT Image 2.5 as a milestone from two independent sides:

  • Generation Quality. GPT Image 2.5 currently ranks first on the Arena leaderboards: GPT Image 2.5 Sunburst is first in both image editing and text-to-image generation, and GPT Image 2.5 Flare is second in image editing.
  • Detection Difficulty. On FraudBench, the six later generators are clearly harder to detect. With all 16 MLLMs and against the same real images, their average TPR is 0.093, compared with 0.158 for the first six. GPT Image 2.5 Flare and Sunburst are the hardest of the six, with a TPR of only 0.045 and 0.043.

Weakness 2.2 (Questions 5 & 6): Human Comparison and Model Selection

"...The paper claims 'human evaluators outperform all evaluated MLLMs'. ... and the size of the human-evaluation subset is not reported."

Response 2.2:

1. Human Comparison

With a single image, humans are comparable to the best MLLMs; with several images, MLLMs clearly exceed them. Computed the same way as Table 1, the human average in the single-image, review-free setting is 0.684, close to the best of the 30 MLLMs (Gemini 3 Flash, 0.674), and we will describe it this way. Without any review text, MLLMs exceed this level once they see several photos, for example Qwen3.8-Max at 0.818 in the multi-step setting. As your Weakness 1 points out, GPT-6 Astra's 0.930 partly reflects a cue in the review text (Response 1.2), so we use the review-free results for this comparison.

2. Size of the Human Evaluation

Each of the three evaluators labeled all 6,940 evaluation images, the same set used for the MLLMs in Table 1, one image at a time and without review text. The evaluators were not involved in building the benchmark and had not seen any generated image before the study. We apologize that the paper did not report this: "small-scale" in Section 4.1 refers to the number of evaluators, and the revision will state both. The evaluators differ widely in how readily they call an image fake (TNR from 0.61 to 0.97), much like the MLLMs, which shows how hard the task is even for attentive humans. Their majority vote reaches a balanced accuracy of 0.749, and we will report the full inter-rater agreement in the appendix.

3. Model Selection

Table 2 shows one model per provider, chosen for a balance between accuracy and API cost, because every Table 2 model is run in all three settings, with and without review text, on the full benchmark. Table 1 lists our 16 latest MLLMs, and the 14 earlier ones follow the same protocol in the appendix. GPT-6 Astra was left out because it is by far the most expensive model we use (10 USD and 50 USD per million input and output tokens, four to five times the price of GPT-5.6 Terra), so its multi-step runs on the original files were incomplete. For this rebuttal, we completed its multi-step evaluation on FraudBench-N for all 1,668 multi-image reviews: it reaches 0.765, compared with 0.599 for single images on the same reviews. We will add GPT-6 Astra to Table 2 and state the selection criterion in the caption.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all three reviewers for their thoughtful and constructive feedback. Their comments motivated us to conduct substantial additional experiments, refine our task definition, and revisit several conclusions in the original paper. We would like to briefly summarize the key findings:


1. Benchmark Reliability and File-Level Properties (Reviewers eqnu and 2Q3A).

We identified a file-level shortcut in the original benchmark: a metadata-only baseline achieved 0.987 balanced accuracy. After applying identical resizing and JPEG re-encoding to both classes (FraudBench-N), this dropped to 0.583 (0.500 on a fixed-resolution subset). All eight specialized-detector checkpoints also dropped to 0.472–0.597, compared with up to 0.904 on the original files. In contrast, the average balanced accuracy of 16 MLLMs changed only modestly, from 0.596 to 0.571. We have completed the evaluation on FraudBench-N and will revise our interpretation of the relative performance of specialized detectors and MLLMs accordingly.


2. Multi-View Evidence and Claim Verification (Reviewers eqnu, 2Q3A, and ASFd).

Our new controls in the multi-image setting show that the apparent benefit of review text largely reflects prompt-structure and writing-style differences. Importantly, multi-view gains remain on FraudBench-N without review text: average balanced accuracy improves from 0.577 (single-image) to 0.614 (multi-image) and 0.686 (multi-step). We have also clarified that FraudBench primarily evaluates the detection of AI-fabricated damage evidence, rather than determining whether a refund claim is fraudulent. In a separate claim-verification control involving matching and same-product contradicting complaints, 14 MLLMs achieved 0.871 balanced accuracy (AUROC 0.921) across 208 authentic reviews, whereas image-only detectors could not distinguish the paired cases because the images were identical.


3. Explanation Quality and Data Verification (Reviewers ASFd and eqnu).

For 300 correctly detected fake-image cases, two independent MLLM judges found that 89.7% of explanations referred at least partially to the actual edited region, compared with 30.7% and 24.5% in two control conditions. We have also clarified the human verification procedure. Pixel-difference analysis found changes exceeding 0.15% of image area in 97.6% of evaluated fakes; excluding the remaining 2.4% barely changed average MLLM recall (0.230 to 0.231).


4. Evaluation Scope, Corrections, and Responsible Release (Reviewers eqnu and 2Q3A).

We have additionally evaluated the six generators not included in the original main evaluation. We have also clarified that all three human evaluators assessed the full set of 6,940 images. We will qualify human–MLLM comparisons to matched input settings and remove overstatements such as "below random guessing" and claims that the benchmark necessarily "forces" multimodal reasoning.

Regarding ethics, the paper and repositories include explicit misuse warnings. For the final release, we plan to implement gated access under research-only terms prohibiting redistribution and re-identification, and update the corresponding ethics documentation.


As the discussion period is coming to a close, we sincerely hope that these additional experiments and revisions have addressed the concerns you raised during the review process. We would be very grateful if you could let us know if any further clarification is needed before the discussion concludes.

Thank you again for your time and consideration!

Sincerely,

Authors of Submission #6087

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.