🤖 AI

Beyond Seeing Is Believing: On Crowdsourced Detection of Audiovisual Deepfakes

This paper evaluates crowdsourced detection of audiovisual deepfakes, finding that while aggregating multiple worker judgments can reliably screen for video authenticity, crowdsourcing struggles to consistently identify specific manipulation types and timestamps, particularly for complex audio-video cases.

Original authors: Michael Soprano, Andrea Cioci, Stefano Mizzaro

Published 2026-05-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Michael Soprano, Andrea Cioci, Stefano Mizzaro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the editor of a massive newsroom. Suddenly, a flood of videos arrives, some real and some "deepfakes"—videos that look and sound real but have been secretly altered by AI. Your job is to sort them out. But you can't watch every single one alone, so you hire a team of 240 regular people (a "crowd") to help you spot the fakes.

This paper is the report card on how well that team performed. Here is what they found, explained simply:

The Setup: Two Different "Training Camps"

The researchers didn't just test the crowd once; they ran two separate experiments using two different sets of videos:

The "Hard" Camp (AV-Deepfake1M): These videos were short, tricky, and very realistic.
The "Easier" Camp (TMC): These videos were longer and slightly easier to spot.

In both camps, the crowd had to do three things for each video:

Spot the Fake: Is this real or doctored?
Name the Cheat: Did they mess with the audio (voice), the video (face), or both?
Pinpoint the Time: Exactly when did the cheating start?

The Results: What the Crowd Got Right (and Wrong)

1. The "Safety Net" Effect (RQ1)
The crowd was excellent at not crying wolf. If a video was truly real, the workers almost never called it a fake. They were very careful not to accuse innocent videos.

The Problem: They were terrible at catching the actual fakes. It was like a security guard who is very good at letting real people in, but misses almost all the thieves trying to sneak in.
The Difference: The crowd caught about twice as many fakes in the "Easier" camp (TMC) compared to the "Hard" camp. This proves that the type of video matters a lot.

2. The "Herd Mentality" (RQ2)
When the researchers asked, "Did everyone agree?", the answer was "Not really."

For any single video, the workers often disagreed with each other. One person might think it's real, while another thinks it's fake.
However, when you take the majority vote (what most people said), the signal became much clearer. It's like asking a room full of people to guess the weight of a cow; one person might be way off, but the average of 10 people is usually pretty close.
The Catch: Even with the majority vote, if the video was a really tricky fake, the crowd still missed it. You can't fix a blind spot just by asking more people.

3. The "Blind Spot" on Details (RQ3)
This was the hardest part. Even when the crowd did spot a fake, they were terrible at guessing how it was faked.

The Mix-Up: If a video had both a fake voice and a fake face, the crowd usually just guessed "Fake Voice" or "Fake Face," rarely getting the "Both" answer right.
The Silver Lining (Timestamps): While they couldn't agree on what was faked, they were surprisingly good at agreeing on when it happened. If they thought a video was fake, they could usually point to the same 5-second chunk where the weirdness started.

The Takeaway: A Two-Step Strategy

The paper suggests a practical way to use this "crowd" system, using a simple analogy:

Think of the crowd as a metal detector at an airport.

Step 1 (Screening): The metal detector (the crowd) is great at scanning thousands of bags quickly. It rarely flags a bag that has nothing in it (low false alarms), but it might miss some small, hidden items (missed fakes).
Step 2 (Expert Review): If the detector beeps, you don't ask the detector what the item is or where exactly it is inside the bag. You send that bag to a human expert (or a smart AI) to open it up and inspect it closely.

In short: The crowd is a great "first filter" to catch obvious fakes and flag suspicious videos for experts to look at. But don't ask them to be the final judge on how the video was faked, because they will likely get that wrong.

One Small Warning

The researchers noted that some workers might have started the test with their volume turned down. Since they couldn't hear the audio, they might have missed "audio-only" fakes. This is a reminder that even in a simple task, small details (like volume) can change the results.

Bottom Line: Crowdsourcing is a useful tool for spotting "something is wrong" in videos, but it needs help to figure out exactly "what is wrong."

Technical Summary: Beyond Seeing Is Believing: On Crowdsourced Detection of Audiovisual Deepfakes

Problem Statement
As deepfake technology becomes increasingly accessible and realistic, the reliability of human judgment in distinguishing authentic from manipulated audiovisual media is a critical concern for misinformation mitigation. While automated detectors exist, they often suffer from domain shifts and limited robustness to unseen manipulation techniques. Conversely, human perception is often unreliable, with accuracy frequently hovering near chance levels under online conditions. This paper investigates the role of crowdsourcing as a scalable mechanism for assessing audiovisual authenticity. Specifically, it examines whether non-expert workers can consistently distinguish authentic videos from manipulated ones, agree on their judgments, and accurately identify the specific type of manipulation (audio-only, video-only, or joint audio-video) and its temporal location.

Methodology
The authors conducted two matched crowdsourcing studies using the Prolific platform, employing a single experimental protocol across two distinct benchmarks: AV-Deepfake1M and the Trusted Media Challenge (TMC) dataset.

Data Sampling: From each dataset, the researchers stratified random sampling to select 48 videos (12 authentic, 12 audio-only manipulated, 12 video-only manipulated, and 12 joint audio-video manipulated), resulting in a total of 96 videos.
Task Design: Workers were assigned work units containing four videos. For each video, they first judged authenticity (Real vs. Manipulated). If "Manipulated" was selected, they identified the type (Audio, Video, or Both) and marked an approximate timestamp.
Aggregation Strategies: To derive video-level labels, the study employed two aggregation methods:
1. Majority Vote: Selecting the label with the highest number of votes (breaking ties in favor of "Authentic" to minimize false positives).
2. Dempster-Shafer (DS) Theory: A belief-function approach that weights evidence based on worker reliability (computed via leave-one-out validation) to handle uncertainty and combine imperfect judgments.
Evaluation Metrics: The study measured classification performance (Precision, Recall, F1, False Positive/Negative Rates), inter-annotator agreement (Krippendorff's $\alpha$ , majority agreement, pairwise agreement), and timestamp consistency (median, IQR, and agreement scores within a $\pm$ 5% window). Statistical significance was assessed using non-parametric tests (Mann-Whitney U, McNemar's, Kruskal-Wallis) with appropriate corrections for multiple comparisons.

Key Contributions

Dataset and Protocol Release: The authors released two matched crowdsourcing studies on AV-Deepfake1M and TMC, providing a full dataset of 960 judgments (10 per video) including authenticity, manipulation type, and timestamp annotations, alongside the complete task configuration.
Comparative Analysis: The study provides a systematic analysis of accuracy, agreement, and error profiles across two different datasets and three manipulation types under a unified protocol, assessing the internal consistency of reported timestamps.
Aggregation Trade-offs: The paper compares Majority Vote and Dempster-Shafer aggregation, quantifying the trade-off between missed manipulations (False Negatives) and false alarms (False Positives) when using crowd judgments as a screening signal.

Results

Authenticity Detection (RQ1): Crowd workers rarely misclassify authentic videos as manipulated (low False Positive Rate), but they frequently miss actual manipulations (high False Negative Rate). Performance varied significantly by dataset: TMC yielded higher accuracy and recall than AV-Deepfake1M. Aggregation (both Majority Vote and DS) stabilized the authenticity signal but could not recover manipulations that the majority of workers consistently missed. DS aggregation increased sensitivity to manipulations (higher recall) but at the cost of increased false positives.
Agreement (RQ2): Internal consistency among workers was limited overall but significantly higher for TMC than for AV-Deepfake1M. Krippendorff's $\alpha$ was 0.113 for AV-Deepfake1M and 0.353 for TMC, indicating that the stability of the crowd signal is heavily dependent on dataset characteristics.
Manipulation Type Identification (RQ3): Identifying the specific manipulation type was substantially noisier than detecting authenticity. Even when workers correctly flagged a video as manipulated, they struggled to attribute the anomaly to the correct modality. Joint audio-video manipulations were particularly difficult to recognize, often being reduced to a single modality (usually audio-only). However, when workers did flag a manipulation, their timestamp reports showed reasonable convergence on the manipulated segment, especially in the TMC dataset.

Significance and Claims
The paper concludes that crowdsourcing can serve as a scalable screening signal for audiovisual authenticity, particularly for filtering out clearly authentic content, but it is not a standalone solution for reliable detection or modality attribution. The authors emphasize that:

Dataset Dependency: The effectiveness of crowd judgments is not uniform; it varies significantly based on the characteristics of the underlying dataset (e.g., video length, manipulation quality).
Modality Attribution is an Open Challenge: While detecting that a video is fake is feasible with aggregation, determining how it was faked (audio vs. video vs. both) remains unreliable, especially for complex joint manipulations.
Practical Workflow: The findings support a two-stage workflow where crowdsourcing is used for initial, scalable screening to prioritize content for further review, while modality attribution and final verification are deferred to expert review or model-assisted interfaces. The study does not claim that crowdsourcing can replace automated detectors or expert fact-checkers but rather suggests its utility as a component in a human-in-the-loop system for misinformation management.

The Setup: Two Different "Training Camps"

The Results: What the Crowd Got Right (and Wrong)

The Takeaway: A Two-Step Strategy

One Small Warning

Technical Summary: Beyond Seeing Is Believing: On Crowdsourced Detection of Audiovisual Deepfakes

More like this