MIC: Explaining Image-Claim Inconsistencies in AI-Generated Multimodal Misinformation

Ruihong Zeng; Jonathan Tonglet; Preslav Nakov; Iryna Gurevych

cs.CLSep 27, 2026

MIC: Explaining Image-Claim Inconsistencies in AI-Generated Multimodal Misinformation

Authors: Ruihong Zeng, Jonathan Tonglet, Preslav Nakov, Iryna Gurevych

Organizations: Mohamed bin Zayed University of Artificial Intelligence, UAE · UKP Lab · Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science, TU Darmstadt and National Research Center for Applied Cybersecurity ATHENE, Germany · Department of Electrical Engineering, KU Leuven, Belgium · Department of Computer Science, KU Leuven, Belgium

Abstract

Claims paired with AI-generated images are a rapidly growing form of misinformation. Existing automated fact-checking (AFC) methods mainly treat this as a provenance problem, detecting low-level synthesis artifacts to decide whether an image is AI-generated. However, such methods do not verify what human fact-checkers often check: whether an image's content is consistent with the context implied by its accompanying claim. To address this gap, we introduce MIC (Multimodal Inconsistency Checking), an AFC framework that assists human fact-checkers by detecting AI-generated multimodal misinformation and explaining inconsistencies using world knowledge. MIC first uses supervised fine-tuning (SFT) for task adaptation and then applies Group Relative Policy Optimization (GRPO) to directly optimize component-level verifiable rewards for verdict prediction, inconsistency type classification, visual evidence description, and world-knowledge explanation. We further introduce MIC-Bench, a benchmark comprising 8,812 image-claim instances derived from 4,406 claims, where each claim is paired with an authentic image and an AI-generated counterpart that introduces a controlled contextual inconsistency. Compared with SFT alone, GRPO further improves Macro-F1 by 4.67 and 4.11 points in the in-distribution and out-of-distribution settings, respectively, while also improving the semantic similarity of visual evidence descriptions and world-knowledge explanations to reference annotations. Our code and data are available at https://github.com/UKPLab/arxiv2026-mic.

Figures & tables

Figure 1: (A) A real-world example of AI-generated multimodal misinformation fact-checked by a human expert ( Awosoro, 2026 ) . (B) A MIC-Bench example of AI-generated multimodal misinformation fact-checked by MIC using visual evidence and world knowledge.

Figure 2: Overview of our MIC-Bench construction pipeline.

Inconsistency type	Representative visual elements
clothing	Uniforms, traditional clothing, religious attire
flag	National flags, regional flags, protest flags
gesture	Cultural greetings, public-speech gestures
signage	Storefront signs, government signs, street signs
architecture	Religious buildings, civic landmarks, arches, facades
infrastructure	Road signs, transit signs, traffic signals

Table 1: The nine inconsistency types in MIC-Bench.

	In-Distribution Replacement				Out-of-Distribution Replacement
Model	Macro-F1	TypeAcc	VisualSim	ExplSim	Macro-F1	TypeAcc	VisualSim	ExplSim
CLIP	39.91	–	–	–	40.72	–	–	–
MiRAGe	53.13	–	–	–	51.70	–	–	–
SNIFFER	46.69	–	–	10.66	44.07	–	–	8.49
GPT-5.4-mini	53.89	11.52	20.14	17.88	53.00	12.18	19.43	17.02
Qwen2.5-VL-3B-Instruct	40.47	15.79	13.83	29.11	38.98	13.32	11.83	28.08

Table 2: Performance (in %) of MIC and baselines on MIC-Bench. In-Distribution Replacement uses edited-in visual elements seen during training, whereas Out-of-Distribution Replacement uses edited-in visual elements held out from training. The best performance per dataset per metric is marked in boldface.

Figure 3: Verdict prediction performance on the test set.

Figure 4: Ablation study of the MIC training stages.

Figure 5: Impact of visual and textual inputs of MIC.

Figure 6: Fine-grained performance of MIC across different inconsistency types.

Figure 7: Generalization performance of MIC on images generated by FLUX.2-klein-9B.

Figure 8: Generalization performance of MIC on recent 2026 news images.

Failure mode	Count
Missed cultural edit	11
Missed object or scene edit	4
Evidence describes feature not in image	5
Evidence describes unedited element	1
Wrong country or brand attribution	9

Table 3: Manual analysis of 30 MIC failure cases from the MIC-Bench test splits.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Input	Retained	Removed	Criterion
Source-image URL and resolution filtering	16,000	13,507	2,493	Valid URL and resolution $\geq 1024\times 1024$
Sample screening for editing (GPT-4o-mini)	13,507	10,301	3,206	At least one feasible inconsistency type
Visual element visibility check (GPT-5.5)	10,301	5,984	4,317	Clearly visible target without caption-only inference
Edit generation and post-generation filtering	5,984	5,818	166	Successful generation passing moderation and size checks
Human verification (test subset)	1,533	1,401	132	Both annotators approve realism, edit fidelity, and the factual correctness of the explanation
Final split construction	5,818	4,406	1,412	Temporal and entity-disjoint splits with the verified test set

Appendix

Table 4: Filtering statistics and decision criteria for the MIC-Bench construction pipeline.

Split	#Claims	#Consistent	#Inconsistent	Total
Train	2,561	2,561	2,561	5,122
Validation	444	444	444	888
Test (In-distribution)	703	703	703	1,406
Test (Out-of-Distribution)	698	698	698	1,396
Total	4,406	4,406	4,406	8,812

Appendix

Table 5: Dataset split statistics for MIC-Bench.

Inconsistency type	Train	Validation	Test (In-distribution)	Test (Out-of-Distribution)
clothing	295	59	68	73
flag	244	45	100	97
gesture	263	35	79	77
signage	263	52	88	89
architecture	293	59	80	77
infrastructure	301	56	75	64

Appendix

Table 6: Distribution of inconsistency types in MIC-Bench across data splits.

Type	Edits	Distinct elements	Most frequent element (share)	OOD held-out entities
Clothing	495	189	Pakistani shalwar kameez ( $3.8\%$ )	59
Flag	486	48	Chinese flag ( $12.3\%$ )	17
Gesture	454	48	Indian namaste ( $12.8\%$ )	22
Signage	492	99	Spanish text ( $7.1\%$ )	36
Architecture	509	175	Mughal architecture ( $8.8\%$ )	57
Infrastructure	496	245	Pakistani motorway sign ( $6.9\%$ )	60

Appendix

Table 7: Introduced-element statistics at the canonical entity level for each inconsistency type.

Benchmark	Primary task	Misinformation input	Inconsistency structure	Controlled edit	Inconsistency annotations
MiRAGeNews	Real vs. generated news	Fictional claim + AI-generated image	No controlled image–claim inconsistency	No	No
OOC benchmarks	Consistent vs. inconsistent	Different event/context + authentic image	Broad contextual mismatch	No	No
MIC-Bench	Consistency and explanation	Same real-event claim + AI-generated image	Localized inconsistency and broader event preserved	Yes	9 types, visual evidence, and world-knowledge explanation

Appendix

Table 8: Comparison of MIC-Bench with related multimodal misinformation benchmarks.

Inconsistent Type	Claim	What Changed	Why Contradicts
clothing	Muhammadu Buhari Wins Second Term as Nigeria’s President. President Muhammadu Buhari of Nigeria after casting his vote in his hometown Daura on Saturday.	Buhari’s light blue Nigerian traditional robe and patterned cap $\rightarrow$ Ghanaian Ashanti kente ceremonial garment.	The caption identifies Nigeria’s president voting in his hometown, so dressing him in distinctly Ghanaian ceremonial clothing conflicts with the Nigerian electoral context.
flag	‘Suffering and Hardship Belong to Me’: A Voice From a Chinese Prison. A protester holding a photo of Xu Zhiyong outside the Chinese liaison office in Hong Kong in 2014.	Chinese five-star red flag $\rightarrow$ South African multicolor Y-design flag.	The caption identifies the scene as outside the Chinese liaison office in Hong Kong in 2014, where a Chinese national flag would be contextually expected, not South Africa’s flag.
gesture	Retracing the Photographic Steps of a 1951 New York City Shoot. Central Park. April 2, 1951.	Man filming communal meal $\rightarrow$ man and nearby diners exchanging Indian namaste greeting.	A formal South Asian namaste-style greeting changes the scene from a 1951 Central Park social dining gathering into a culturally different ceremonial interaction that does not fit the caption’s New York context.
signage	Gandhi Biographer Arrested as Protests Over Citizenship Law Sweep India. Protesters denouncing a new citizenship law near the historic Red Fort in New Delhi on Thursday.	English protest placard text “No CAA. & NRC.” $\rightarrow$ German placard text “Nein zu CAA und NRC”.	A German-language anti-CAA/NRC protest sign is inconsistent with a street protest described as taking place near the Red Fort in New Delhi, where such signage would ordinarily be in English, Hindi, or Urdu rather than German.
architecture	Italy’s Populist Parties, on Precipice of Power, Fail to Form Government. Italy’s prime minister-designate, Giuseppe Conte, center, leaving a meeting with Italy’s president, Sergio Mattarella, on Sunday in Rome.	Italian official interior doorway/wall detailing $\rightarrow$ Brazilian Portuguese-colonial blue-and-white azulejo-tiled doorway surround with shallow colonial arch.	The caption places the scene in Rome at Italy’s presidential residence, but the edited architecture implies a Brazilian Portuguese-colonial governmental interior instead of an Italian one.
infrastructure	Snowstorm That Wasn’t Finds City Well Prepared. A city truck spread salt on 18th Avenue in Brooklyn on Friday.	Brooklyn U.S. parking regulation sign $\rightarrow$ Mexico City Metrobus-style Spanish transit-lane sign.	A Mexican transit-corridor sign is inconsistent with a caption placing the scene on 18th Avenue in Brooklyn, New York City.

Appendix

Table 9: Representative visual examples for all nine inconsistency types in MIC-Bench.

Figure 9: Type-level confusion matrices of MIC on the in-distribution and out-of-distribution test splits.

Figure 10: Precision-Recall curves and AUPRC of MIC and CLIP on the combined in-distribution and out-of-distribution test set.

/* Qwen3-Embedding-0.6B */
	In-Distribution		Out-of-Distribution
Model	VisualSim	ExplSim	VisualSim	ExplSim
GPT-5.4-mini	20.14	17.88	19.43	17.02
Qwen3-VL-4B-Instruct	23.16	25.24	24.18	27.18
MIC	83.04	82.54	72.09	75.78
/* all-mpnet-base-v2 */

Appendix

Table 10: Robustness of VisualSim and ExplSim to the choice of evaluation embedding model. The best performance per dataset per metric is marked in boldface.

Figure 11: Impact of reward weight on MIC.

	In-Distribution Replacement				Out-of-Distribution Replacement
Model	VisualSim	ExplSim	VisualJudge	ExplJudge	VisualSim	ExplSim	VisualJudge	ExplJudge
SNIFFER	–	10.66	–	3.50	–	8.49	–	2.32
GPT-5.4-mini	20.14	17.88	21.05	20.74	19.43	17.02	19.89	19.71
Qwen2.5-VL-3B-Instruct	13.83	29.11	10.87	9.87	11.83	28.08	8.60	8.22
Qwen3-VL-4B-Instruct	23.16	25.24	23.53	22.84	24.18	27.18	25.42	24.21
Qwen3-VL-8B-Instruct	26.74	28.55	27.68	26.60	25.13	26.88	25.16	25.21

Appendix

Table 11: Reference-based evaluation of MIC on MIC-Bench. The best performance per dataset per metric is marked in boldface.

Model	$F_{1}^{\mathrm{inc}}$	$F_{1}^{\mathrm{con}}$	Macro-F1
CLIP	65.31	3.77	34.54
GPT-5.4-mini	38.24	68.18	53.21
Qwen2.5-VL-3B-Instruct	50.56	59.55	55.05
Qwen3-VL-8B-Instruct	53.12	68.75	60.94
InternVL3-8B	57.82	68.98	63.40
MIC	85.98	83.87	84.93

Appendix

Table 12: Robustness to the AI-generation shortcut.

Claim	Benign Addition	Why It Remains Consistent
Gov. Greg Abbott of Texas said on Thursday that access to clean water remained a problem even as power was restored.	a weathered pickup truck	Pickup trucks are common in Texas, so the addition is plausible and does not alter the depicted event.
Police officers in tactical gear left the site of a shooting in Bangkok on Friday.	a street vendor cart with snacks	Street vendor carts are common in Bangkok and add a local element without changing the depicted event.
President Trump with Prime Minister Narendra Modi at Motera Stadium in Ahmedabad, India.	an Indian flag on a pole	An Indian flag is expected at an event held in India, introducing no contradiction with the claim.
An M.T.A. worker disinfected a subway car at the Coney Island Yard in Brooklyn on Tuesday.	a small yellow traffic cone	Traffic cones are routine in a subway maintenance yard, matching the claimed context.
Michael R. Bloomberg speaking in Clarksville, Tenn., on Friday.	a small American flag pin	Flag pins are a common accessory for U.S. politicians at public speaking events.
Activists gathered outside City Hall calling for an extension of the eviction moratorium in New York on Monday.	a yellow taxi cab	Yellow taxis are ubiquitous in New York City, so the addition fits the setting.

Appendix

Table 13: Representative AI-generated consistent examples used in the AI-generation shortcut control. Each edited image contains a benign addition that preserves consistency with the claim.

Model	Macro-F1	TypeAcc	VisualSim	ExplSim
Qwen3-VL-4B-Instruct	65.20	43.00	39.90	43.50
GPT-5.4-mini	50.70	9.00	9.70	11.60
Qwen3-VL-8B-Instruct	54.00	54.00	50.90	54.80
Qwen3-VL-8B-Thinking	68.50	21.00	17.50	18.80
MIC	95.50	65.00	52.50	59.90

Appendix

Table 14: Performance (in %) on the two-inconsistency evaluation set. The best performance per metric is marked in boldface.

/* In-Distribution Replacement */
Method	Macro-F1	TypeAcc	VisualSim	ExplSim
Zero-shot CoT	58.90	19.06	23.16	25.24
SFT	90.21	86.77	77.52	78.98
SFT + DPO	75.08	84.64	80.40	83.91
SFT + GRPO (MIC)	94.88	89.76	83.04	82.54
/* Out-of-Distribution Replacement */

Appendix

Table 15: Comparison of training strategies. DPO and GRPO are initialized from the same SFT checkpoint.

Model	$F_{1}^{\mathrm{con}}$	$F_{1}^{\mathrm{inc}}$	Macro-F1
Qwen2.5-VL-3B-Instruct	7.70	67.60	37.60
InternVL3-8B	51.20	66.10	58.70
Qwen3-VL-4B-Instruct	79.10	82.60	80.80
Qwen3-VL-8B-Instruct	83.90	86.00	84.90
GPT-5.4-mini	93.50	92.50	93.00
MIC	95.20	94.70	95.00

Appendix

Table 16: Cross-benchmark consistency evaluation on a balanced, human-verified subset of MiRAGeNews. Results are reported in %.

Model	Macro-F1	TypeAcc	VisualSim	ExplSim
Qwen2.5-VL-3B-Instruct	36.52	12.50	10.66	13.70
LLaVA-OneVision-7B	53.43	13.75	24.07	27.93
InternVL3-8B	60.03	31.75	23.31	27.67
GPT-5.4-mini	59.94	29.00	21.53	23.67
Qwen3-VL-8B-Instruct	63.42	38.25	32.33	34.48
Qwen3-VL-4B-Instruct	59.53	36.00	24.01	29.80

Appendix

Table 17: Performance (in %) on the held-out flag and signage types. The best performance per metric is marked in boldface.

Failure Mode	Claim	MIC Output	Why This Is Incorrect
Missed cultural edit	President Trump defied the governor of Nevada by holding an indoor rally near Las Vegas.	<verdict> : CONSISTENT <type> : None <visual> : None <explanation> : None	The candidate’s dark suit has been replaced by a Filipino barong Tagalog , an attire that does not fit a U.S. campaign rally. MIC fails to register the cultural mismatch and incorrectly predicts CONSISTENT .
Missed object or scene edit	Residents of Kenosha, Wis., watched President Trump’s motorcade this month.	<verdict> : CONSISTENT <type> : None <visual> : None <explanation> : None	The sidewalk in a Wisconsin street scene has been replaced with shallow muddy water and rice-paddy seedlings, an environmental cue that does not fit an urban Midwestern setting. MIC overlooks the scene-level edit entirely and predicts CONSISTENT .
Evidence describes feature not in image	Activists were arrested for planning an anti-corruption demonstration in Manara Square, Ramallah.	<verdict> : INCONSISTENT <type> : branding <visual> : Dell-branded storefront/logo <explanation> : A Dell storefront logo introduces a distinctly American consumer-tech brand into a Palestinian protest scene, creating a cultural mismatch.	MIC reaches the correct verdict but justifies it using hallucinated evidence: the cited Dell storefront does not exist in the image. The actual edit, Arabic storefront signs replaced by Urdu Nastaliq script, is never identified.
Evidence describes unedited element	Firefighters pronounced three children dead at an apartment complex in Los Angeles.	<verdict> : INCONSISTENT <type> : architecture <visual> : Spanish colonial-style arched stucco facade with terracotta tile roofline <explanation> : The edited architecture suggests a Latin American Spanish colonial setting rather than a typical California urban residential building.	MIC bases its verdict on an unedited architectural feature already present in the original image. The actual edit, an apartment leasing sign replaced by an Occupy Wall Street banner, is missed.
Wrong country/brand attribution	Andrew Young, the American ambassador to Burkina Faso, during the opening ceremony of a joint military exercise.	<verdict> : INCONSISTENT <type> : technology <visual> : Russian AK-12 assault rifles <explanation> : Equipping the ceremonial honor guard with Russian-service AK-12 rifles shifts the implied military equipment and alliance context away from the Burkinabé setting in the caption.	MIC attends to the correct edited object, the ceremonial rifles, and reaches the correct verdict, but misattributes their identity: the actual edit introduces Italian Beretta ARX160 rifles, not Russian AK-12s. The verdict and visual region are correct, but the entity-to-country link fails.

Appendix

Table 18: Representative MIC failure cases across different error modes. MIC may miss edited evidence, hallucinate unsupported evidence, attend to unedited regions, or misattribute edited entities.

Inconsistency Type	Claim	What Changed	Why Contradicts
clothing	Meeting of French mountain troops and Danish soldiers as part of the Arctic endurance mission.	Danish-style red patch with white cross $\rightarrow$ Indonesian TNI red-over-white flag patches and “TNI” identifiers.	The caption frames the scene as a meeting between French and Danish forces. Replacing the visible uniform identifiers with Indonesian TNI patches changes the represented military affiliation.
flag	U.S. Marine Corps Maj. Sean Gunn speaks to Philippine Air Force leadership.	Philippine flag $\rightarrow$ Indian tricolor flag with navy Ashoka Chakra.	The claim anchors the scene in a Philippine Air Force meeting. Replacing the Philippine flag with India’s national flag misidentifies the host-country context.
gesture	Kaja Kallas and S. Jaishankar signing a Security and Defence Partnership in New Delhi.	Two officials signing documents $\rightarrow$ two officials performing a mutual Japanese deep bow while seated.	The claim describes a diplomatic signing ceremony in India. The edited body language changes the action into a Japanese-style bowing greeting, so the depicted interaction no longer matches the claimed event.
signage	Aircraft attached to Carrier Air Wing (CVW) 9 sit on the flight deck of Nimitz-class aircraft carrier USS Abraham Lincoln (CVN 72) in support of Operation Epic Fury.	Aircraft fuselage text “VFA-41” $\rightarrow$ Japanese Air Self-Defense Force text.	The claim describes U.S. Navy aircraft aboard the USS Abraham Lincoln. The edited Japanese military marking assigns the aircraft to a different national force, creating a direct mismatch with the carrier-air-wing context.
architecture	Korpaskhas evacuates victims of the Cisarua landslide in West Bandung.	Damaged slanted greenhouse roofs $\rightarrow$ damaged Japanese shrine/temple-style curved kawara tile roofs.	The claim places the landslide site in West Bandung, Indonesia, but the edited roof forms suggest Japanese religious or traditional architecture rather than local agricultural structures.
infrastructure	The front of the march in support of Ukraine with flags and banners; behind the trees is the Washington Monument on 17th Street NW, Washington, DC.	Washington, DC road markings $\rightarrow$ Japanese Tokyo-style zebra crosswalk and diamond warning marking.	The claim identifies the scene as 17th Street NW in Washington, DC, while the edited road markings are characteristic of Japanese urban traffic infrastructure.

Appendix

Table 19: Representative edited examples from recent real-world news images from 2026.

Explore similar work

SynCred-Bench: Benchmarking Synthetic Credibility in AI-Generated Visual Misinformation
Jun 2, 2026Junxiao Yang, Minghao Zhang, Xiaoce Wang +4Credibility Human Annotators
The Conversational AI (CoAI) group, DCST, Tsinghua University
SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data
Aug 13, 2026Yicheng Bao, Xiahui Guo, Xuhong Wang +1Ai-Generated Image Detection Adversarial Examples
1East China Normal University · 2Shanghai Artificial Intelligence Laboratory
TextFake: Benchmarking AI-Generated Image Detection on Text-Rich Images
May 31, 2026Yuning Zhang, Changtao Miao, Mingyu Liao +5Ai-Generated Image Detection Forgeries
School of Cyber Science and Technology, University of Science and Technology of China · Anhui Province Key Laboratory of Digital Security · Individual Researcher

Jun 2, 2026, cs.CVJ/K move · Enter open · S save

SynCred-Bench: Benchmarking Synthetic Credibility in AI-Generated Visual Misinformation

Junxiao Yang, Minghao Zhang, Xiaoce Wang, Haoran Liu+3

The Conversational AI (CoAI) group, DCST, Tsinghua University

Credibility Human Annotators Generative Models

Open paper AskPDF arXiv

Recent generative models can now produce visual artifacts with realistic embedded text and layouts, creating a new misinformation threat: synthetic credibility. We introduce SYNCRED-Bench, a benchmark of 600 AI-generated misinformation images balanced across six credible-form categories and seven fine-grained circulation styles, together with FP450, a real-image negative set for measuring false positives. Extensive evaluation shows that existing systems remain unreliable: under a 5% false-positive-rate constraint, 15 MLLMs achieve only 10.5% true positive rate (TPR), open-source AIGC detectors achieve less than 5%, and commercial APIs reach 57.6%. Human annotators also struggled to identify synthetic credibility, reaching only 63% TPR. These findings establish synthetic credibility as a severe and underexplored visual misinformation challenge, and provide a benchmark for developing detectors that reason beyond superficial credibility cues.

Abstract

Explore similar work

SynCred-Bench: Benchmarking Synthetic Credibility in AI-Generated Visual Misinformation

SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data

TextFake: Benchmarking AI-Generated Image Detection on Text-Rich Images