acceptodds
Under review as a conference paper at ICLR 2027

GenTell-Gaze: A Dataset and Benchmark for Localising Physical Anomalies in AI-Generated Video with Human Gaze

Abstract

Generated AI videos are becoming increasingly realistic, making the identification of when and how they diverge from real-world behaviour not only harder, but also more important. An AI video failure that is still relatively common is incorrectly representing physical laws and dynamics, leading to anomalous artefacts in the videos. Correctly identifying and locating these can help create explainable and more trustworthy fake-video detectors, as well as help evaluate vision-language models’ ability to understand physics. We introduce the GenTell-Gaze dataset: 380 generated videos spanning 80 physics-based scenes and six video generators, with 1,070 manually annotated spatio-temporal anomaly `tubes', alongside eye-tracking data from 129 participants. We develop GenTell-Bench, a benchmark for evaluating vision-language model (VLM) reasoning capabilities. This includes classifying videos as real or AI-generated, as well as recognising, classifying, and localising any present physical anomalies. Using the eye-tracking data, we are able to evaluate whether human visual attention can improve VLM video reasoning and anomaly localisation. Across all models tested, performance falls substantially as the tasks progress from video classification to temporal and spatio-temporal localisation, revealing a persistent gap between recognising generated content and identifying the responsible anomalies with physical reasoning.

Primary area:
datasets and benchmarks
Keywords:
AI-generated video detection, vision-language models, anomaly localization, eye tracking, physical reasoning

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_bDQkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a dataset, GenTell-Gaze, that provides spatio-temporal anomaly tubes (trajectories of bounding boxes) and corresponding human gaze data for AI-generated videos. The proposed benchmark (GenTell-Bench) evaluates video classification, anomaly count and type prediction, temporal localization, and spatio-temporal localization. Experimental results show that open-weight and frontier VLMs underperform humans in real-vs.-AI classification and exhibit weak anomaly localization performance. When models use human gaze data, it improves the temporal localization performance of some VLMs but does not show consistent improvements overall.

Critical strengths
  1. Practical and insightful combination of human gaze data and annotations of physical artifacts in AI-generated videos. This annotation potentially enables us to understand how humans identify AI-generated videos and why they classify them that way.

  2. Interesting finding regarding the number of anomalies. Humans can more easily classify AI-generated videos with a larger number of artifacts, but VLMs do not show the same clear trend. This could provide useful insights for future model development.

Critical weaknesses
  1. Limited analysis of human gaze data. How much do humans actually look at anomaly regions? Does looking at these regions help them correctly identify AI-generated videos?

  2. Gaze patterns might depend on individual participant characteristics, but these are not presented in the paper. For example, age, gender, familiarity with AI or AI-generated images/videos, eyesight, etc. Although I understand that the authors did not collect personal or demographic data, some of these factors should have been addressed, at least in the participant recruitment criteria.

Questions for authors
  1. How often did participants actually look at the annotated anomaly regions? Is there any relationship between gaze–anomaly alignment and the accuracy of human AI-generated video detection?

  2. Could the authors clarify the participant recruitment and screening criteria, particularly regarding eyesight and familiarity with AI-generated content? How might individual differences (e.g., age, gender, or prior AI experience) affect the gaze patterns and the interpretation of the results?

2 comments

Thank you very much for your time and effort evaluating our work. Answers below.

  1. We did run a more in-depth analysis of the human gaze before submitting the paper (it was actually one of the first things we did) but decided not to include it in the paper. Clearly, the reviewer indicated there is much interest, so these will be included. For your questions: On the first watch, people, on average, fixated on 33% of the annotated anomalies. Yes, this does help, as people whose gaze was fixed on the anomaly, meaning they saw it (which does not guarantee they realised it was an anomaly), had a higher accuracy at correctly detecting AI videos (66% when they didn’t fixate on an anomaly, 78% when they fixated in one watch, and 83% when they fixated on the same anomaly on both watches). The remaining 17% in the latter are people who saw the anomaly but must have decided it was not an anomaly, similar to the phenomenon we observed with Astra.

  2. Regarding demographics, we did not collect race/nationally/etc, but we did make a record of gender, age and eyesight, as well as prior experience with generative AI. We have now checked all these characteristics against both accuracy and gaze on anomalies and found no detectable effect. Participants were recruited via flyers within the university and were mostly students, although there were few older participants. All participants had normal or corrected-to-normal vision. No participants were screened based on these criteria. We will add summary statistics of these characteristics and the effect analysis to the paper as well.

Dear Reviewer,

We have now uploaded a revised version of the paper. Additional gaze analysis is included in Section 5.1 and Appendix I, and a summary table of the participants' characteristics available to us, including a confirmation that these do not impact gaze data in a non-negligible manner, has been added to Appendix E.

Many thanks!

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

This paper describes GenTell-Gaze, a benchmark for evaluating vision language model (VLM) capabilities by classifying videos as real or AI-generated. The authors identify spatio-temporal anomalies in videos as a marker for AI-generated videos and label these in such videos. In addition to evaluating models on their ability to identify these videos and localize the artifacts, they also record human eye-tracking data on these videos to understand whether human visual attention can help with VLMs on this task. Their findings indicate that this is still a challenging task for existing VLMs, and that even the best performing models do not/cannot use human gaze to improve their performance.

My opinion overall is that the paper is readable, asks an interesting question, and discovers an interesting result. The value of this paper is mainly in the benchmark and future work that may use this benchmark to help train VLMs to better use gaze and video evidence grounding for decision making. Therefore, even though I do not think there is a lot of novelty, I lean towards acceptance.

Critical strengths

Well motivated benchmark, paper, and evaluation experiments. For each possible evaluation setting and condition, multiple options are considered making which helps build confidence in the findings.

Critical weaknesses

The major finding is that the biggest contributor in the performance was the baseline bias of a VLM at this task, which is striking, but does not actually show how VLMs can get get competitive with humans on this task.

Questions for authors

Is it possible to evaluate (beyond the reasoning trace) by using counterfactuals, whether certain aspects (video evidence, spatio-temporal evidence) contribute to VLM reasoning on this task?

2 comments

Thank you very much for your time and effort evaluating our work. Answers below.

For your question – yes, there is to some extent. First, we do see that for all models, when the model correctly identifies an anomaly, they judge the video as AI, which is a strong indication that their reasoning link is there between anomaly → fake. For the correct identification, the traces are very useful; however, we can also direct the VLMs directly to the anomalous area, ask them to describe what is happening there and use that as a replacement for the traces. We can also test how much this shifts the accuracy. To some extent, this was done by providing the models with the gaze data (although less on the nose), but at least without training, the models didn’t seem to be able to extract this information from this data. For most of the models (outside Astra), the issue was more that they detected anomalies where there were no human-detectable anomalies or no anomalies at all (on the real videos). So while the issue with Astra was not identifying the area where anomalies were, or not recognising them as actual anomalies, the other had the issue of “detecting” anomalies where there aren’t any; therefore, asking for each anomaly to be described is likely to improve our understanding of their current physical reasoning. In addition, with the code we will include the full benchmark runs, which provide substantial individual examples for further insight into interpreting the model's reasoning.

Reviewer_oPPqOpenReview ↗

Response to rebuttal

Thank you to the authors for their response. It clarifies my question to some extent. I will keep my score.

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

This paper studies human vs AI fake video classification and diagnostics from a physics perspective. It contributes a dataset on human vs AI video classification and diagnostics. Most notably, this work makes a valuable contribution of two types of annotations, including spatial-temporal bounding boxes (tubes) and human gaze data.

Critical strengths
  1. This paper contributes to addressing a severe lack of spatial-temporal data in video understanding-related tasks. The authors collect spatial-temporal tubes - bounding boxes over time. This could be made clearer in the writing, including whether this is novel, and if so, why this representation?

  2. The gaze data is also valuable.

  3. The overall potential of this work is undeniable, but it lacks critical analysis and findings, which should be 70% of this manuscript in different sections. If sincerely addressed, this would improve both informativeness and soundness substantially.

Critical weaknesses
  1. Soundness: What do the gaze data vs the annotator labels tell, respectively? Currently, the experiments and findings don't immediately connect them.

    a. After reading the paper, I can't help but ask the question of "why collect annotated tubes?" and "why collect annotated gaze?"

    b. There are no findings and examples based on annotated tubes and gaze. Currently, the paper is sitting at fewer than 9 pages and there are a lot of compression opportunities. You should feature findings based on human vs AI tubes and gazes front and center, as that's the unique advantage of your work that cannot be easily replicated by others.

    c. In fact, with the above implemented, you should reposition your paper as a finding paper, e.g., "AIs and Humans Attend to Physics Differently" or change "Differently" to your actual finding. This would be incredibly strong.

  2. Soundness: you should complete the missing batch protocols in lines 246-249, unless there is a convincing reason not to.

  3. Soundness and writing re: Section 4. "The GenTell Benchmark."

    a. You listed many tasks: video classification, anomaly detection, temporal localization, and spatio-temporal localization. However, it's unclear which ones have human baselines. If they are based on models alone, then there may be a soundness concern.

    b. This section should not be written like this. Currently, the "The GenTell Benchmark." indicates that this is supposed to describe the benchmark, but the current content is describing your model-based study section describing tasks and metrics for baseline models. Either move it to a subsection in Section 3 or move it to Section 5.

  4. Writing: the bounding box study designs are underspecified:

    a. What were the human annotator cohorts? Did you hire the same cohorts for each group of generated videos, i.e., did the same annotators work on the same videos? Why the variation? b. Were they also shown the real videos for a blind control? Did they label any anomalies for the real videos in addition to the fake videos? This is a critical and a significant control potentially leading to findings. This control would also likely strengthen your existing findings.

  5. Writing: tube is underspecified. Are tubes standard representations? What do they look like? Do others use it? How do you make sure that there may be an abnormal bbox but not necessarily a tube?

  6. Writing (stylistic). I'm sure you are aware by now of some fixable issues: a. The caption for Table 1 (lines 157-161) is LLM-generated and conveys little meaning. b. line 236: "University of [Anonymised]" c. The paper isn't full (fewer than 9 pages).

Questions for authors
  1. Broadly, what are the connections between your dataset/the broader problem and your annotated tubes data and gaze data?

  2. What are your findings regarding humans vs AI for tubes and gaze?

  3. What were your IRB considerations since this involves recruited human annotations for tubes and especially gaze which is necessarily in-person?

3 comments

Thank you very much for your time and effort evaluating our work. Answers below.

  1. I am not sure I understand this question but will try my best to answer it based on other questions within this review. When we generated the AI videos which correspond (i.e., show the same scene) to real videos, we have ground truth for which video is real and which is generated. As the generated videos received a still image from the real video within the prompt, they look visually extremely similar to the real video. The main differentiator between AI-generated and real videos is physical anomalies in the former. For these anomalies, we do not have an immediate ground truth, as these were not seeded/injected. Instead, for anomaly detection, maximal-effort human baseline is taken as the ground truth. We therefore collect annotated tubes to be able to know exactly where and when there are human-detectable anomalies within the videos. For the definition of a tube, it is a time-series of bounding boxes which captures the anomalous area. These were chosen to both contain rich information compared to a static image or time, and to be a human-interpretable annotation. A couple of other works use a similar annotation framework (see Table 1 in the paper). Tubes were collected through a clear protocol, with no time limits, and the ability to rewatch the videos, pause and zoom with no restrictions. Gaze data, on the other hand, is collected from first and second watch of the video only, to allow for a meaningful comparison between gaze data of different viewers. When putting the gaze and tubes together, one can answer the question of whether the viewer looked at the anomalous area of the video (in the spatiotemporal sense). That in itself does not tell us if the viewer recognised that what they are seeing is an anomaly, but this can be estimated based on their judgment of whether the video was real or fake. When comparing to AI, the tubes can be used to verify if a model can correctly locate (in the spatiotemporal sense) human-detectable anomalies within the video. Using the gaze, we can also compare how many people looked at that anomaly and compare it to the model’s ability to correctly locate it.
  2. This comparison should be handled a little carefully, as AI and humans don’t “view” videos the same mechanistically. We can make two comparisons – AI anomaly detection compared to human maximal effort (currently reported) and AI versus human first/second watch (added). In the latter, we find several findings: First, on the first watch, people, on average, fixated on 33% of the annotated anomalies, while AI models detected up to 12% (time-only) and 4% (spatiotemporal) of anomalies. Second, people whose gaze was fixed on the anomaly, meaning they saw it (which does not guarantee they realised it was an anomaly), had a higher accuracy at correctly detecting AI videos (66% when they didn’t fixate on an anomaly, 78% when they fixated in one watch, and 83% when they fixated on the same anomaly on both watches). The remaining 17% in the latter are people who saw the anomaly but decided it was not an anomaly, similar to the phenomenon we observed with Astra. Finally, anomalies which more people were fixated on were also, on average, more likely to be detected by AI.
  3. We indeed conducted in-person experiments with participants who were recruited via flyers across the university. Participants were told they would be shown AI and real videos (so no deception), they gave written informed consent and knew their gaze data would be collected and included in a dataset, and were able to pause or stop at any time. None of the videos showed distressing content, and participants got a comfort break after 10 clips. Participants were paid for their time. There is no risk of de-identification from the gaze data. For tube annotating, the task was not time-bound, and payment was hourly. Overall, this required 38 hours of effort, with a mean of 6.2min for each 5 sec clip. Annotators did see the full datasets, including real videos and no anomalies were annotated on real clips. Up to two annotators covered each video; different generators were annotated by the same annotators, so there is no bias within the labelling.

1. Structure and completeness of the rebuttal. The authors' response does not follow the conventions of a structured academic rebuttal. Rather than systematically addressing individual reviewer concerns, it reads more like an informal conversational exchange. It is often unclear which specific concerns are being addressed, what evidence or revisions support each response, and whether all points have been adequately covered. I encourage the authors to organize their response point by point, explicitly identifying each concern and the corresponding clarification, evidence, or manuscript revision. This would make it substantially easier to evaluate the completeness and adequacy of the rebuttal.

2. Implications of the new exploratory findings. I appreciate the addition of Section 5.1 and Appendix I, which substantially strengthen the empirical contribution of the paper. However, the manuscript primarily reports these findings descriptively, with limited discussion of their broader scientific and practical implications.

In particular, what do the observed differences and similarities between human attention and AI anomaly localization imply for our understanding of VLM physical reasoning? How might these findings inform the development of better AI-generated video detectors, particularly regarding temporal reasoning, attention mechanisms, and human-AI complementarity?

I encourage the authors to expand the discussion of these implications, distinguishing conclusions supported by the current evidence from promising directions for future investigation. This would help establish the significance of the new findings beyond their descriptive value. Although I have raised my score, I expect the authors to continue working towards the goal of pushing this paper even further, to better answer the question "what does each finding imply?"

Thank you very much for your time and valuable feedback!

Point 1: Our intention was to be concise, but we appreciate that the format you suggest might be more convenient. Point-by-point from the original review above:

  1. (a), (b), and partly (c) are addressed in the new section Sec. 5.1 and Appendix I, and the primary results are also reported in (1) the rebuttal comment above.
  2. More details on the protocols were added in Sec. 3.3 and in Appendix E.1.
  3. Section 4 was rewritten following your suggestion. We further clarified which are the human baselines in Sec. 5.1 (first paragraph) in line with the explanation given in (2) in the rebuttal comment above.
  4. Bounding-box study design -- further details were added to Appendix F. As stated in (3) in the rebuttal comment above, annotators did see real clips as well, and no anomalies were annotated in those.
  5. Tube definition added to sec. 3.2, and illustrative examples of tubes were added to Appendix F.
  6. Stylistic: (a) Table 1 caption — rewritten for improved clarity; it is still long, but we felt the information there is needed to understand the comparison with other works. (b) "University of [Anonymised]" — removed from Sec. 3.3. (c) Length — main text now concludes at the end of p.10.

Point 2: We really appreciate this! These are important questions which we will carefully consider and work towards including them, together with an expanded discussion, in the final version of the paper.

Thank you very much again for engaging with us and helping us to improve this paper!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Revision

Dear Reviewers and AC, thank you for your helpful feedback. We have now uploaded the rebuttal revision. In the spirit that this is a rebuttal and not a revise-and-resubmit, we aimed to keep the changes to a minimum. The biggest change is the addition of section 5.1 and Appendix I, which analyses participants' attention to anomalies (based on the gaze data) and makes the comparison to AI. We have also added a summary table of the participants' characteristics available to us, including a confirmation that these do not impact gaze data in a non-negligible manner (Appendix E). We also added details on the experimental protocols (Appendix E). In addition, we added the tube definition and added examples (Sec. 3.2; Appendix F), and restructured Section 4, following the suggestion from reviewer woKQ.

We thank you again for your time and efforts in evaluating our paper!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.