pith. sign in

arxiv: 2605.20787 · v3 · pith:A7R72KRSnew · submitted 2026-05-20 · 💻 cs.CV

Findings of the Counter Turing Test: AI-Generated Image Detection

Pith reviewed 2026-05-22 09:59 UTC · model grok-4.3

classification 💻 cs.CV
keywords AI-generated image detectiongenerative model attributionimage classificationcomputer visiondeep learning detectorssynthetic mediamodel fingerprinting
0
0 comments X

The pith

AI-generated images can be detected with high accuracy but identifying the exact generative model remains difficult.

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

The paper reports results from testing whether images can be classified as real or AI-generated and, for synthetic images, which model created them. Participants applied convolutional networks, vision transformers, frequency analysis and other methods to a dataset pairing real photographs with synthetic images from several generators. Strong results on the binary task indicate that current tools can reliably flag synthetic content, which would help address misinformation if the performance holds. Weaker results on model identification show that different generators produce overlapping traces that are hard to separate.

Core claim

Participants achieved F1-scores above 0.83 for classifying images as real or AI-generated using strategies such as convolutional neural networks, vision transformers, frequency-based analysis, contrastive learning, and multimodal techniques. In contrast, the highest F1-score for identifying which particular generative model produced a given image reached only 0.4986. The evaluation relied on a dataset combining real images with 50,000 synthetic images produced by multiple generative models.

What carries the argument

A dual-task benchmark requiring first binary classification of images as real or synthetic and second attribution of synthetic images to their source generative model.

Load-bearing premise

The collected set of synthetic images from several current generative models paired with real images captures the range of visual properties that detectors will encounter in everyday use.

What would settle it

Testing the top binary and model-identification systems on images produced by generative models absent from the original dataset and measuring whether F1 scores fall below 0.7 would determine whether the reported performance generalizes.

Figures

Figures reproduced from arXiv: 2605.20787 by Aishwarya Naresh Reganti, Aman Chadha, Amitava Das, Amit Sheth, Ashhar Aziz, Gurpreet Singh, Kapil Wanaskar, Nasrin Imanpour, Nilesh Ranjan Pal, Parth Patwa, Rajarshi Roy, Ritvik Garimella, Shashwat Bajpai, Shreyas Dixit, Shwetangshu Biswas, Subhankar Ghosh, Vasu Sharma, Vinija Jain, Vipula Rawte.

Figure 1
Figure 1. Figure 1: Baseline workflow. The input image is first transformed into its frequency domain represen￾tation and then passed through a ResNet-50 CNN classifier to predict whether it is real or fake. 4. Participating Systems The challenge utilized the MS COCOAI dataset, an extension of the MS COCO dataset, comprising 50,000 images generated by models such as DALL-E 3, Stable Diffusion, and Midjourney. Participants aim… view at source ↗
Figure 1
Figure 1. Figure 1: Baseline workflow. The input image is first transformed into its frequency domain represen￾tation and then passed through a ResNet-50 CNN classifier to predict whether it is real or fake. 4. Participating Systems The challenge utilized the MS COCOAI dataset, an extension of the MS COCO dataset, comprising 50,000 images generated by models such as DALL-E 3, Stable Diffusion, and Midjourney. Participants aim… view at source ↗
read the original abstract

The rapid advancements in generative AI technologies, such as Stable Diffusion, DALL-E, and Midjourney, have significantly transformed the creation of synthetic visual content. While these models enable innovation across industries, they also pose serious challenges, including misinformation, disinformation, and biased content generation. The increasing realism of AI-generated images makes their detection a pressing concern for researchers, policymakers, and industry stakeholders. In this paper, we present the findings of the Defactify 4.0 workshop, which introduced the Counter Turing Test (CT2) for AI-Generated Image Detection. The competition consisted of two key tasks: (1) binary classification of images as either AI-generated or real and (2) identification of the specific generative model responsible for an AI-generated image. To support both tasks, we employed the MS COCOAI dataset, a benchmark of 96000 real and synthetic images generated by five state-of-the-art models alongside real images from MS COCO. Participants employed diverse detection strategies, including convolutional neural networks (CNNs), Vision Transformers (ViTs), frequency-based analysis, contrastive learning, and multimodal techniques. The results demonstrated that while AI-generated images can be detected with high accuracy (F1-score > 0.83), identifying the exact model used remains significantly more challenging (highest F1-score: 0.4986). These findings highlight the need for improved model fingerprinting, adversarial robustness, and real-time detection mechanisms.

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 / 2 minor

Summary. The manuscript reports the findings of the Defactify 4.0 workshop's Counter Turing Test (CT2) competition on AI-generated image detection. It introduces the MS COCOAI dataset (50,000 synthetic images from multiple generative models paired with real MS COCO images) and describes participant submissions using CNNs, ViTs, frequency analysis, contrastive learning, and multimodal methods. The central empirical results are F1 > 0.83 on binary AI-vs-real classification and a best F1 of 0.4986 on identifying the specific generative model.

Significance. If the benchmark proves robust, the results establish that binary detection is practically feasible with existing architectures while model attribution remains substantially harder, providing a concrete empirical baseline that can guide future work on fingerprinting and adversarial robustness.

major comments (1)
  1. [§3] §3 (MS COCOAI dataset construction): The dataset is described only at a high level (50k synthetic images from 'multiple generative models' plus MS COCO reals). No information is given on exact model versions, generation hyperparameters, prompt sampling, resizing/upsampling kernels, or compression steps. This detail is load-bearing for the headline claim of F1 > 0.83, because without it the performance cannot be distinguished from exploitation of dataset-specific artifacts (fixed kernels, prompt biases, or train-test leakage) as noted in the stress-test concern.
minor comments (2)
  1. [Abstract] Abstract: The statement 'F1-score > 0.83' should specify whether this is the single best submission, the mean across teams, or a threshold; the same clarification is needed for the model-identification F1 of 0.4986.
  2. [Results] Results section: Add per-team breakdowns, number of submissions, and any statistical significance or variance measures for the reported F1 scores to allow readers to assess stability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript describing the findings of the Defactify 4.0 Counter Turing Test competition. We address the single major comment below and will incorporate the requested details to strengthen the paper's reproducibility and address concerns about potential dataset artifacts.

read point-by-point responses
  1. Referee: [§3] §3 (MS COCOAI dataset construction): The dataset is described only at a high level (50k synthetic images from 'multiple generative models' plus MS COCO reals). No information is given on exact model versions, generation hyperparameters, prompt sampling, resizing/upsampling kernels, or compression steps. This detail is load-bearing for the headline claim of F1 > 0.83, because without it the performance cannot be distinguished from exploitation of dataset-specific artifacts (fixed kernels, prompt biases, or train-test leakage) as noted in the stress-test concern.

    Authors: We agree that the current high-level description of the MS COCOAI dataset in Section 3 is insufficient for full reproducibility and does not adequately address potential concerns about dataset-specific artifacts. In the revised manuscript we will expand Section 3 with a dedicated subsection that specifies the exact generative models and versions employed, the generation hyperparameters, the prompt sampling procedure (including how MS COCO captions were selected and diversified), and all post-processing steps such as resizing kernels, upsampling methods, and compression. We will also add a brief discussion of steps taken during dataset construction to mitigate common artifacts, such as prompt diversity and standardized pipelines. These additions will allow readers to better evaluate the robustness of the reported F1 scores (>0.83 for binary detection) and will clarify that the competition dataset was designed as a standardized benchmark rather than an artifact-prone test set. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical competition report with no derivation chain

full rationale

The paper is a report on competition results for binary AI-vs-real classification and model identification using the MS COCOAI dataset. It contains no equations, mathematical derivations, fitted parameters, or self-citation chains that reduce any claimed performance metric to the input data by construction. All reported F1 scores are direct empirical outcomes from participant submissions evaluated on the held-out test split; the analysis is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

No free parameters, invented entities, or non-standard axioms; the work rests on standard machine-learning evaluation assumptions such as representative train/test splits and i.i.d. sampling.

axioms (1)
  • domain assumption Standard machine-learning assumptions of i.i.d. data and representative sampling hold for the MS COCOAI dataset
    Implicit in any competition-based performance claim

pith-pipeline@v0.9.0 · 5878 in / 1201 out tokens · 50420 ms · 2026-05-22T09:59:07.595063+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    Participants employed diverse detection strategies, including convolutional neural networks (CNNs), Vision Transformers (ViTs), frequency-based analysis, contrastive learning, and multimodal techniques. The results demonstrated that while AI-generated images can be detected with high accuracy (F1-score > 0.83)

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.