pith. sign in

arxiv: 2606.07613 · v1 · pith:4YKPFKI7new · submitted 2026-05-29 · 💻 cs.CV · cs.AI

Can You Trust What You See? Human and AI Detection of Synthetic Legal Evidence

Jinzhe Tan , Ali Ekber Cinar , Karim Benyekhlef This is my paper

Pith reviewed 2026-06-28 22:53 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords synthetic legal evidenceAI image detectionhuman perceptionmultimodal large language modelscivil disputestext-to-image generatorsauthenticity verification
0
0 comments X

The pith

Neither humans nor multimodal AI models can reliably tell authentic legal photos from AI-generated ones.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether people and frontier AI systems can spot synthetic images used as evidence in civil cases. It created a dataset of 200 real photos paired with 1200 fakes from six generators across ten common evidence categories. In experiments, human participants reached only 64.8 percent accuracy overall, falling to chance levels on the two strongest generators. The AI models never mistook a real image for fake but detected almost none of the harder synthetic images. Errors between humans and models were uncorrelated, suggesting they cannot serve as independent checks.

Core claim

In controlled tests on the SLED-1400 dataset, lay humans achieved 64.8% accuracy distinguishing real evidentiary photographs from AI-generated counterparts, dropping to 48.5% and 51.0% on outputs from Gemini-3-Pro-Image and Flux-2-Max. Four evaluated MLLMs showed 100% specificity but only 5.9% average detection rate on the hardest generator outputs. Human and MLLM errors were largely uncorrelated.

What carries the argument

The SLED-1400 dataset of 1400 images (200 authentic paired with 1200 synthetic from six text-to-image generators in ten object-centric categories) used in identical format for human experiments and MLLM evaluations.

If this is right

  • Visual evidence in legal proceedings should be treated as inherently contestable.
  • A workable procedural response must combine trained human review, MLLM screening, and provenance infrastructure such as C2PA Content Credentials.
  • Neither humans nor MLLMs alone provide sufficient authentication for synthetic legal evidence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar detection failures may occur in other high-stakes domains like journalism or medical imaging where authenticity matters.
  • Future work could test whether combining human and MLLM judgments improves overall detection rates beyond either alone.
  • Training humans on specific generator artifacts might raise accuracy above chance levels on advanced models.

Load-bearing premise

The chosen set of 200 authentic images and 136 lay web participants represent the kinds of evidence and decision-makers that appear in real civil disputes.

What would settle it

A study applying the same detection task to images actually submitted as evidence in ongoing civil cases would show whether the reported accuracy levels hold outside the controlled dataset.

Figures

Figures reproduced from arXiv: 2606.07613 by Ali Ekber Cinar, Jinzhe Tan, Karim Benyekhlef.

Figure 1
Figure 1. Figure 1: Overview of the experimental design. Real evidence images ( [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic-detection rate by generator (rows) and rater (columns). [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Human accuracy at each confidence level (1–7), with 95 % Wilson CIs [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: For each rater, the Truth: REAL bar shows responses on real images and the Truth: SYNTHETIC bar shows responses on synthetic images, broken into Predicted REAL (blue), Predicted SYNTHETIC (red), and Predicted UNSURE (grey). Right-hand columns: sensitivity 𝑑 ′ (higher = better separation between real and synthetic) and decision criterion 𝑐 (positive = bias toward calling an image real). 4.2.2 Stronger gener… view at source ↗
Figure 5
Figure 5. Figure 5: Per-rater accuracy on each of the ten evidence categories (real and [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Top: distribution of stated confidence for each MLLM (marker area [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pearson correlation between each pair of raters, computed across the 999 [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
read the original abstract

Visual evidence has long been treated as a reliable form of legal proof, but advances in artificial intelligence (AI) are undermining that assumption. This article asks how well humans and frontier multimodal large language models (MLLMs) can distinguish authentic evidentiary photographs from AI-generated counterparts in the object-centric scenarios typical of civil disputes. We built Synthetic Legal Evidence Detection (SLED-1400), a dataset of 200 authentic evidence images paired with 1,200 synthetic counterparts produced by six contemporary text-to-image generators across ten evidence categories. The same stimuli and response format were used in a controlled web experiment with 136 lay participants and in a standardized evaluation of four MLLMs (GPT-5.1, Gemini-3-Pro, Gemini-3-Flash, Qwen3-VL-235B). Human accuracy was 64.8% overall, and 48.5% and 51.0% on the two strongest generators (Gemini-3-Pro-Image and Flux-2-Max), indistinguishable from chance. MLLMs never misclassified an authentic image (100% specificity), but missed most synthetic outputs from the harder generators, with average MLLM detection at 5.9% on Gemini-3-Pro-Image outputs. Human and MLLM errors were largely uncorrelated, while the four MLLMs were strongly correlated with each other. Neither group is a reliable standalone authenticator. We argue that visual evidence in legal proceedings should be treated as inherently contestable, and that a workable procedural response must combine trained human review, MLLM screening, and provenance infrastructure such as C2PA Content Credentials.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper introduces the SLED-1400 dataset of 200 authentic evidentiary photographs paired with 1,200 synthetic outputs from six text-to-image generators across ten object-centric categories typical of civil disputes. It reports a controlled web experiment with 136 lay participants yielding 64.8% overall human accuracy (dropping to 48.5% and 51.0% on Gemini-3-Pro-Image and Flux-2-Max outputs, indistinguishable from chance) and a standardized evaluation of four MLLMs (GPT-5.1, Gemini-3-Pro, Gemini-3-Flash, Qwen3-VL-235B) showing 100% specificity but only 5.9% average detection on the hardest generator. Human and MLLM errors are largely uncorrelated while the MLLMs correlate strongly with each other. The authors conclude that neither humans nor MLLMs constitute reliable standalone authenticators and argue that visual legal evidence should be treated as inherently contestable, requiring combined trained review, MLLM screening, and provenance infrastructure such as C2PA.

Significance. If the empirical results hold, the work supplies concrete, falsifiable performance numbers on a purpose-built legal-evidence dataset and documents the complementary failure modes of humans and MLLMs. The new SLED-1400 stimuli and the direct head-to-head comparison constitute a useful public resource for subsequent research on detection reliability. The observation that errors are uncorrelated supplies a concrete rationale for multi-layered authentication protocols rather than reliance on any single method.

major comments (1)
  1. [Methods (Human Evaluation)] The experimental protocol (136 lay web participants, object-centric images presented without metadata, chain-of-custody cues, or adversarial legal framing) is described in the Methods section; because the headline claim is that neither humans nor MLLMs are reliable standalone authenticators in civil-dispute settings, the absence of any comparison to trained legal professionals or to stimuli that include additional evidentiary signals is load-bearing for the generalization from the controlled task to actual courtroom authentication.
minor comments (1)
  1. [MLLM Evaluation] The abstract states MLLM detection rates but does not specify the exact prompt templates or decision thresholds applied to the four models; adding these details (or a supplementary table) would improve reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the value of the SLED-1400 dataset and the head-to-head comparison. We address the single major comment below.

read point-by-point responses

  1. Referee: [Methods (Human Evaluation)] The experimental protocol (136 lay web participants, object-centric images presented without metadata, chain-of-custody cues, or adversarial legal framing) is described in the Methods section; because the headline claim is that neither humans nor MLLMs are reliable standalone authenticators in civil-dispute settings, the absence of any comparison to trained legal professionals or to stimuli that include additional evidentiary signals is load-bearing for the generalization from the controlled task to actual courtroom authentication.

    Authors: We agree that the controlled protocol with lay participants and stimuli lacking metadata or chain-of-custody cues limits direct generalization to full courtroom authentication by trained professionals. The experiment was intentionally designed to isolate unaided visual detection performance as a baseline, which remains relevant because evidentiary photographs in civil disputes are frequently presented without reliable accompanying signals. The observed drop to chance-level accuracy on strong generators already indicates that standalone visual inspection is unreliable even under simplified conditions; this supplies evidence for the broader claim that visual legal evidence should be treated as contestable. We did not test trained legal experts because the study targets general human capabilities (analogous to jury members) rather than specialist performance, but we accept that including such a comparison would strengthen external validity. We will revise the manuscript to add an explicit Limitations subsection clarifying the scope of the human evaluation and noting that future work should examine trained professionals and stimuli with additional evidentiary signals. revision: partial

Circularity Check

0 steps flagged

Empirical study with no derivation chain or self-referential predictions

full rationale

The paper is a data-collection and evaluation study: it constructs SLED-1400 (200 authentic + 1200 synthetic images), runs controlled web experiments with 136 participants, and evaluates four MLLMs on the same stimuli. No equations, fitted parameters, predictions derived from prior fits, or uniqueness theorems appear. Central claims (human accuracy 64.8 % overall, MLLM detection 5.9 % on hardest generator) are direct empirical measurements, not reductions to inputs by construction. Self-citations, if any, are not load-bearing for the reported results.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The study depends on standard assumptions about the representativeness of the stimuli and participants for the legal domain, with no free parameters or new entities introduced.

axioms (2)
  • domain assumption Lay participants in a web-based experiment can serve as proxies for individuals involved in civil disputes when assessing image authenticity.
    The human study uses 136 lay participants to measure detection ability.
  • domain assumption The outputs from the six text-to-image generators represent the current state of synthetic evidence that could be presented in legal contexts.
    The dataset uses these to test detection capabilities.

pith-pipeline@v0.9.1-grok · 5837 in / 1409 out tokens · 37841 ms · 2026-06-28T22:53:08.250178+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 41 canonical work pages

  1. [1]

    ACM Transactions on Multimedia Computing, Communications, and Applications , author =

    Revolutionizing. ACM Transactions on Multimedia Computing, Communications, and Applications , author =. 2024 , pages =. doi:10.1145/3689641 , abstract =

    work page doi:10.1145/3689641 2024

  2. [2]

    Rutgers Computer & Technology Law Journal , author =

    A. Rutgers Computer & Technology Law Journal , author =. 1992 , pages =

    1992

  3. [3]

    Yale Journal of Law & the Humanities , author =

    The image of truth:. Yale Journal of Law & the Humanities , author =. 1998 , pages =

    1998

  4. [4]

    IEEE Access , author =

    A. IEEE Access , author =. 2023 , keywords =. doi:10.1109/ACCESS.2023.3317083 , abstract =

    work page doi:10.1109/access.2023.3317083 2023

  5. [5]

    Chalkidis, Ilias and Androutsopoulos, Ion and Aletras, Nikolaos , editor =. Neural. Proceedings of the 57th. 2019 , pages =. doi:10.18653/v1/P19-1424 , abstract =

    work page doi:10.18653/v1/p19-1424 2019

  6. [6]

    , month = jul, year =

    Ashley, Kevin D. , month = jul, year =. Artificial. doi:10.1017/9781316761380 , abstract =

    work page doi:10.1017/9781316761380

  7. [7]

    The Law Teacher , author =

    Artificial intelligence in legal education: a scoping review , issn =. The Law Teacher , author =. 2026 , pages =. doi:10.1080/03069400.2025.2592440 , language =

    work page doi:10.1080/03069400.2025.2592440 2026

  8. [8]

    Changes in

    Vučić, Franjo , editor =. Changes in. Digital. 2023 , pages =. doi:10.1007/978-3-031-36833-2_12 , language =

    work page doi:10.1007/978-3-031-36833-2_12 2023

  9. [9]

    the language of the computer: a bilingual model of legal education in the age of technology and artificial intelligence , volume =

    The language of the law vs. the language of the computer: a bilingual model of legal education in the age of technology and artificial intelligence , volume =. Law, Innovation and Technology , author =. 2024 , pages =. doi:10.1080/17579961.2024.2392938 , language =

    work page doi:10.1080/17579961.2024.2392938 2024

  10. [10]

    , month = dec, year =

    Shen, Bingyu and RichardWebster, Brandon and O'Toole, Alice and Bowyer, Kevin and Scheirer, Walter J. , month = dec, year =. A. 2021 16th. doi:10.1109/FG52635.2021.9667066 , urldate =

    work page doi:10.1109/fg52635.2021.9667066 2021

  11. [11]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , author =

    Enhancing. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , author =. 2025 , pages =. doi:10.1609/aies.v8i1.36529 , abstract =

    work page doi:10.1609/aies.v8i1.36529 2025

  12. [12]

    IEEE Signal Processing Magazine , author =

    More. IEEE Signal Processing Magazine , author =. 2022 , pages =. doi:10.1109/MSP.2021.3120982 , number =

    work page doi:10.1109/msp.2021.3120982 2022

  13. [13]

    SN Social Sciences , author =

    Problematising reality: the promises and perils of synthetic media , volume =. SN Social Sciences , author =. 2020 , pages =. doi:10.1007/s43545-020-00010-8 , language =

    work page doi:10.1007/s43545-020-00010-8 2020

  14. [14]

    Perception inception: preparing for deepfakes and the synthetic media of tomorrow , shorttitle =

    Barraclough, Tom and Barnes, Curtis , month = may, year =. Perception inception: preparing for deepfakes and the synthetic media of tomorrow , shorttitle =

  15. [15]

    Frontiers in Artificial Intelligence , author =

    What you see is not what you get anymore: a mixed-methods approach on human perception of. Frontiers in Artificial Intelligence , author =. 2025 , pages =. doi:10.3389/frai.2025.1707336 , abstract =

    work page doi:10.3389/frai.2025.1707336 2025

  16. [16]

    The state of

    Labajová, Lucia , year =. The state of

  17. [17]

    Frank, Joel and Herbert, Franziska and Ricker, Jonas and Schönherr, Lea and Eisenhofer, Thorsten and Fischer, Asja and Dürmuth, Markus and Holz, Thorsten , month = may, year =. A. 2024. doi:10.1109/SP54263.2024.00159 , urldate =

    work page doi:10.1109/sp54263.2024.00159 2024

  18. [18]

    Dianey and Rojas-Velazquez, David and Gorelova, Aleksandra V

    Rueda-Arango, Y. Dianey and Rojas-Velazquez, David and Gorelova, Aleksandra V. and Lopez-Rincon, Alejandro , month = sep, year =. Exploring. 2024. doi:10.1109/ISTAS61960.2024.10732054 , urldate =

    work page doi:10.1109/istas61960.2024.10732054 2024

  19. [19]

    Seeing is not always believing: benchmarking human and model perception of

    Lu, Zeyu and Huang, Di and Bai, Lei and Qu, Jingjing and Wu, Chengyue and Liu, Xihui and Ouyang, Wanli , month = dec, year =. Seeing is not always believing: benchmarking human and model perception of. Proceedings of the 37th

  20. [20]

    IEEE Transactions on Pattern Analysis and Machine Intelligence46(8), 19 5625–5644 (2024) https://doi.org/10.1109/TPAMI.2024.3369699

    Vision-. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2024 , pages =. doi:10.1109/TPAMI.2024.3369699 , number =

    work page doi:10.1109/tpami.2024.3369699 2024

  21. [21]

    2023 , pages =

    Journal of Imaging , author =. 2023 , pages =. doi:10.3390/jimaging9100199 , abstract =

    work page doi:10.3390/jimaging9100199 2023

  22. [22]

    and Jain, Ishan and Wang, Oliver and Zhang, Richard , year =

    Epstein, David C. and Jain, Ishan and Wang, Oliver and Zhang, Richard , year =. Online

  23. [23]

    Cozzolino, Davide and Poggi, Giovanni and Nießner, Matthias and Verdoliva, Luisa , editor =. Zero-. Computer. 2025 , pages =. doi:10.1007/978-3-031-72649-1_4 , language =

    work page doi:10.1007/978-3-031-72649-1_4 2025

  24. [24]

    IEEE Access , author =

    Performance. IEEE Access , author =. 2024 , pages =. doi:10.1109/ACCESS.2024.3394250 , urldate =

    work page doi:10.1109/access.2024.3394250 2024

  25. [25]

    Computer Science Review , author =

    Methods and trends in detecting. Computer Science Review , author =. 2026 , pages =. doi:10.1016/j.cosrev.2026.100908 , language =

    work page doi:10.1016/j.cosrev.2026.100908 2026

  26. [26]

    Raising the

    Cozzolino, Davide and Poggi, Giovanni and Corvi, Riccardo and Nießner, Matthias and Verdoliva, Luisa , year =. Raising the

  27. [27]

    Houston Law Review , author =

    Navigating the. Houston Law Review , author =. 2025 , pages =

    2025

  28. [28]

    Hungarian Journal of Legal Studies , author =

    Large language models and their possible uses in law , volume =. Hungarian Journal of Legal Studies , author =. 2024 , pages =. doi:10.1556/2052.2023.00475 , abstract =

    work page doi:10.1556/2052.2023.00475 2024

  29. [29]

    Evaluating

    Bhambhoria, Rohan and Dahan, Samuel and Li, Jonathan and Zhu, Xiaodan , editor =. Evaluating. Compliance for. 2026 , pages =. doi:10.1007/978-3-032-12795-2_5 , language =

    work page doi:10.1007/978-3-032-12795-2_5 2026

  30. [30]

    Loyola of Los Angeles Law Review , author =

    Generative. Loyola of Los Angeles Law Review , author =. 2025 , pages =

    2025

  31. [31]

    Journal of Banking Regulation , author =

    Legal implications of automated suspicious transaction monitoring: enhancing integrity of. Journal of Banking Regulation , author =. 2024 , pages =. doi:10.1057/s41261-024-00233-2 , abstract =

    work page doi:10.1057/s41261-024-00233-2 2024

  32. [32]

    Global Times , file =

    Chinese court detects. Global Times , file =

  33. [33]

    2023 , pages =

    IEEE Access , author =. 2023 , pages =. doi:10.1109/ACCESS.2023.3324403 , urldate =

    work page doi:10.1109/access.2023.3324403 2023

  34. [34]

    Frontiers in Big Data , author =

    Audio deepfakes:. Frontiers in Big Data , author =. 2023 , pages =. doi:10.3389/fdata.2022.1001063 , abstract =

    work page doi:10.3389/fdata.2022.1001063 2023

  35. [35]

    IEEE Access , author =

    Audio. IEEE Access , author =. 2023 , pages =. doi:10.1109/ACCESS.2023.3333866 , urldate =

    work page doi:10.1109/access.2023.3333866 2023

  36. [36]

    Manipulating faces for identity theft via morphing and deepfake:

    Agarwal, Akshay and Ratha, Nalini , editor =. Manipulating faces for identity theft via morphing and deepfake:. Deep. 2023 , keywords =. doi:https://doi.org/10.1016/bs.host.2022.12.003 , abstract =

    work page doi:10.1016/bs.host.2022.12.003 2023

  37. [37]

    Brooklyn Journal of International Law , author =

    Deep. Brooklyn Journal of International Law , author =. 2023 , pages =

    2023

  38. [38]

    Social Media + Society , author =

    Deepfakes and. Social Media + Society , author =. 2020 , pages =. doi:10.1177/2056305120903408 , abstract =

    work page doi:10.1177/2056305120903408 2020

  39. [39]

    The Eurasia Proceedings of Science Technology Engineering and Mathematics , author =

    Impact of. The Eurasia Proceedings of Science Technology Engineering and Mathematics , author =. 2023 , pages =. doi:10.55549/epstem.1371792 , abstract =

    work page doi:10.55549/epstem.1371792 2023

  40. [40]

    Journal of Intellectual Property Law & Practice , author =

    Regulating deep fakes: legal and ethical considerations , volume =. Journal of Intellectual Property Law & Practice , author =. 2020 , pages =. doi:10.1093/jiplp/jpz167 , language =

    work page doi:10.1093/jiplp/jpz167 2020

  41. [41]

    Journal of Intellectual Property Law & Practice , author =

    Do deepfakes pose a golden opportunity?. Journal of Intellectual Property Law & Practice , author =. 2020 , pages =. doi:10.1093/jiplp/jpz139 , language =

    work page doi:10.1093/jiplp/jpz139 2020

  42. [42]

    Trauma, Violence, & Abuse , author =

    Legal. Trauma, Violence, & Abuse , author =. 2024 , pages =. doi:10.1177/15248380221143772 , abstract =

    work page doi:10.1177/15248380221143772 2024

  43. [43]

    Journal of Tort Law , author =

    Deepfake. Journal of Tort Law , author =. 2025 , pages =. doi:10.1515/jtl-2025-0032 , abstract =

    work page doi:10.1515/jtl-2025-0032 2025

  44. [44]

    Crime, Law and Social Change , author =

    Regulating deepfakes between. Crime, Law and Social Change , author =. 2025 , pages =. doi:10.1007/s10611-024-10197-z , language =

    work page doi:10.1007/s10611-024-10197-z 2025

  45. [45]

    Disrupting and

    Flynn, Asher and Clough, Jonathan and Cooke, Talani , editor =. Disrupting and. The. 2021 , pages =. doi:10.1007/978-3-030-83734-1_29 , language =

    work page doi:10.1007/978-3-030-83734-1_29 2021

  46. [46]

    Cyberspace,

    Montasari, Reza , year =. Cyberspace,. doi:10.1007/978-3-031-50454-9 , language =

    work page doi:10.1007/978-3-031-50454-9

  47. [47]

    IEEE Transactions on Technology and Society , author =

    Beyond. IEEE Transactions on Technology and Society , author =. 2024 , pages =. doi:10.1109/TTS.2024.3427816 , number =

    work page doi:10.1109/tts.2024.3427816 2024

  48. [48]

    Deceptive

    Jasserand, Catherine , month = sep, year =. Deceptive. 2024. doi:10.1109/BIOSIG61931.2024.10786729 , urldate =

    work page doi:10.1109/biosig61931.2024.10786729 2024

  49. [49]

    International Review of Law, Computers & Technology , author =

    Generative. International Review of Law, Computers & Technology , author =. 2024 , pages =. doi:10.1080/13600869.2024.2324540 , language =

    work page doi:10.1080/13600869.2024.2324540 2024

  50. [50]

    2026 , pages =

    AI & Society , author =. 2026 , pages =. doi:10.1007/s00146-025-02755-3 , language =

    work page doi:10.1007/s00146-025-02755-3 2026

  51. [51]

    Face/off: "

    Gerstner, Erik , year =. Face/off: ". Defense Counsel Journal , publisher =

  52. [52]

    Richmond Journal of Law & Technology , author =

    Out of. Richmond Journal of Law & Technology , author =. 2021 , pages =

    2021

  53. [53]

    Northwestern University Law Review , author =

    Deepfake. Northwestern University Law Review , author =. 2021 , pages =

    2021

  54. [54]

    Mitchell Hamline Law Review , author =

    A. Mitchell Hamline Law Review , author =. 2022 , pages =

    2022

  55. [55]

    Virginia Law Review Online , author =

    Deepfakes,. Virginia Law Review Online , author =

  56. [56]

    Hastings Law Journal , author =

    Deepfakes on. Hastings Law Journal , author =. 2023 , pages =

    2023

  57. [57]

    2026 , file =

    Tampa Bay 28 , author =. 2026 , file =

    2026

  58. [58]

    Judges say they're not ready

    AI-generated evidence is showing up in court. Judges say they're not ready. , url =. NBC News , author =. 2025 , day =

    2025

  59. [59]

    Delfino, Rebecca , year =. The. Ohio State Law Journal , publisher =

  60. [60]

    Deeply dehumanizing, degrading, and violating:

    Pascale, Emily , year =. Deeply dehumanizing, degrading, and violating:. Syracuse Law Review , publisher =

  61. [61]

    California Western International Law Journal , author =

    A. California Western International Law Journal , author =. 2024 , pages =

    2024

  62. [62]

    The Business, Entrepreneurship & Tax Law Review , author =

    Synthetic. The Business, Entrepreneurship & Tax Law Review , author =. 2024 , pages =

    2024

  63. [63]

    Science and Technology Law Review , author =

    Judicial. Science and Technology Law Review , author =. 2025 , pages =. doi:10.52214/stlr.v26i2.13890 , abstract =

    work page doi:10.52214/stlr.v26i2.13890 2025

  64. [64]

    , year =

    Murray, Michael D. , year =. Visual. SMU Science and Technology Law Review , publisher =

  65. [65]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    A style-based generator architecture for generative adversarial networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  66. [66]

    arXiv preprint arXiv:2204.06125 , year=

    Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , year=

    Pith/arXiv arXiv

  67. [67]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  68. [68]

    arXiv preprint arXiv:2006.07397 , year=

    The deepfake detection challenge (dfdc) dataset , author=. arXiv preprint arXiv:2006.07397 , year=

    Pith/arXiv arXiv 2006

  69. [69]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Faceforensics++: Learning to detect manipulated facial images , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  70. [70]

    arXiv preprint arXiv:2509.09495 , year=

    OpenFake: An Open Dataset and Platform Toward Real-World Deepfake Detection , author=. arXiv preprint arXiv:2509.09495 , year=

    arXiv

  71. [71]

    arXiv preprint arXiv:2412.15260 , year=

    Analyzing Images of Legal Documents: Toward Multi-Modal LLMs for Access to Justice , author=. arXiv preprint arXiv:2412.15260 , year=

    arXiv

  72. [72]

    arXiv preprint arXiv:2511.13772 , year=

    Can LLMs Create Legally Relevant Summaries and Analyses of Videos? , author=. arXiv preprint arXiv:2511.13772 , year=

    arXiv

  73. [73]

    Proceedings of the 33rd

    Qiang, Chenhui and Wei, Zhaoyang and Han, Xumeng and Wang, Zipeng and Li, Siyao and Lan, Xiangyuan and Jiao, Jianbin and Han, Zhenjun , month = oct, year =. Proceedings of the 33rd. doi:10.1145/3746027.3758208 , language =

    work page doi:10.1145/3746027.3758208

  74. [74]

    Forensic Science International: Digital Investigation , volume=

    Digital forensics in law enforcement: A case study of LLM-driven evidence analysis , author=. Forensic Science International: Digital Investigation , volume=. 2025 , publisher=

    2025

  75. [75]

    Forty-first International Conference on Machine Learning , year=

    Chatbot arena: An open platform for evaluating llms by human preference , author=. Forty-first International Conference on Machine Learning , year=

  76. [76]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  77. [77]

    parallax , volume=

    Visual images in the courtroom: A historical perspective , author=. parallax , volume=. 2008 , publisher=

    2008

  78. [78]

    China Daily , url =

    Li, Hongyang , title =. China Daily , url =. 2025 , month = dec, date =

    2025

  79. [79]

    California Law Review , volume=

    Deep fakes: A looming challenge for privacy, democracy, and national security , author=. California Law Review , volume=

  80. [80]

    AI4AJ@ ICAIL , pages=

    Chatgpt as an artificial lawyer? , author=. AI4AJ@ ICAIL , pages=

Showing first 80 references.