Requests from Hugging Face accounts with an academic email address (for example .edu or .ac.uk) are approved automatically within 6 hours. All other requests are reviewed by hand.

SLOP is released for non-commercial research on detecting and characterizing AI-generated media, under the terms in LICENSE.md. Your Hugging Face username and account email are shared with the dataset maintainers.

⚠️WARNING

Request access through the form above. Requests from Hugging Face accounts whose email is an academic address (for example .edu or .ac.uk) are approved automatically within 6 hours; all other requests are reviewed by hand. If you are an academic using a personal email, change your account email to your institutional address before requesting access.

🔎SEE ALSO

SEAM, the detector introduced alongside SLOP, with code and checkpoints

SLOP (Social-platform Longitudinal Observed Posts) contains 5,925 real and AI-generated videos (80.7 hours) collected from Instagram, TikTok, and X, as they actually circulate on those platforms: re-encoded, cropped, captioned, and re-uploaded. The AI class comes from dedicated AI-persona, AI-influencer, and generator-showcase accounts; the real class comes from extensively documented human creators whose posting histories predate modern video generation. The dataset is close to balanced between AI (2,961) and real (2,964) videos, both overall and within each platform, and ships with account-disjoint train, validation, and test splits.

🌐PROJECT PAGE 💻CODE

Licensing

Access is granted at no cost to researchers at academic and other non-profit research institutions for non-commercial research on detecting or characterizing AI-generated media. Requests from academic email addresses are approved automatically; all others are reviewed by hand. Detectors can be used adversarially, to help build generators that evade them, so we check whether each requester has a legitimate evaluation purpose.

The videos remain the property of their original creators and are redistributed here solely for research. Do not re-host the videos, publish them outside of illustrative figures in research papers, or use them to train generative models. If you are the creator of a video in SLOP and would like it removed, contact us at maty@stanford.edu and we will remove it from the next version and notify existing users.

Any work utilizing the dataset should include appropriate attribution.

The creators of this dataset are not liable for any misuse or unintended consequences resulting from the use of the dataset. Users are responsible for ensuring that their use of the dataset complies with all applicable laws, platform terms, and ethical guidelines.

Getting Started

Install the datasets package (video decoding uses torchcodec):

pip install "datasets>=4.0" torchcodec

Request access above, then log into your Hugging Face account:

hf auth login

Load the dataset. It is downloaded on first use and cached afterwards (63.1 GB in total):

from datasets import load_dataset
dataset = load_dataset("matybohacek/slop")

Access the data individually or as an iterator:

sample = dataset["train"][0]
sample["label"], sample["platform"]   # "ai" / "real", "instagram" / "tiktok" / "x"
decoder = sample["video"]             # torchcodec VideoDecoder; frames are decoded lazily
frame = decoder[0]                    # uint8 tensor, C x H x W

for sample in dataset["test"]:
    # run your inference / analysis
    pass

If you prefer plain files, download the repository as a folder instead. Every split directory holds the videos and a metadata.csv, and manifest.csv at the root lists all videos:

from huggingface_hub import snapshot_download
snapshot_download("matybohacek/slop", repo_type="dataset", local_dir="slop")

Data

Our paper (preprint forthcoming; see the project page) describes data collection, label provenance, and a taxonomy of the content in detail. This card only contains a code-level description of the data.

Platform	AI	Real	Total
Instagram	1,000	1,000	2,000
TikTok	994	1,000	1,994
X	967	964	1,931
Total	2,961	2,964	5,925

The splits follow a 70 / 10 / 20 ratio and are account-disjoint: no account contributes videos to more than one split, so a detector cannot score well by recognizing creators.

Split	AI	Real	Total
`train`	2,072	2,129	4,201
`validation`	290	271	561
`test`	599	564	1,163

The videos come from 534 known accounts. For 1,316 videos, mostly legacy Instagram items, the posting account was not recorded: their handle is empty and each forms its own creator_id, so for these videos the splits are disjoint by video rather than verifiably by account.

Compared with the collection design of 1,000 videos per platform and class, this release:

replaces 96 TikTok videos that were deleted upstream with another video from the same account, in the same split;
replaces 89 AI TikTok videos whose accounts were deleted or made private with AI videos from other AI accounts found through generator hashtags. These accounts did not go through the account screening described in the paper; a stand-in from an account already in the dataset is placed in that account's split;
removes 69 repeated files, mostly X posts that attach the same media as another post by the same account.

The layout is:

train/metadata.csv
train/<platform>/<label>/<platform>_<video_id>.mp4
validation/...
test/...
manifest.csv

All datapoints come with the following attributes:

video: the decoded video (when loaded through datasets), or file_name, its path relative to the split directory;
label: "ai" or "real";
platform: "instagram", "tiktok", or "x";
split: "train", "validation", or "test";
video_id: the platform-native identifier (Instagram shortcode, TikTok video ID, or X <post id>_<media key>);
handle: the account that posted the video (empty when the account was not recorded);
creator_id: a stable, anonymized account key, the unit the splits are disjoint over;
post_url: a link to the original post, where it can be reconstructed;
posted_at: the posting time in UTC, decoded from the post ID (TikTok and X only);
duration_s, width, height, fps, video_codec, has_audio, bitrate_kbps, size_bytes: technical properties read with ffprobe;
sha256: checksum of the video file;
provenance: original, same_account_substitute, or cross_account_substitute (see above). Filter on it to exclude the stand-ins.

Labels are assigned by account provenance, not by per-video annotation. An account enters the AI class only if it posts exclusively AI-generated content, and the real class only if it belongs to a documented human creator; accounts with any evidence of mixed posting were removed whole. The criteria are documented in the preprint.

⚠️WARNING

Before training or evaluating, make sure to:

shuffle the data since, by default, it is ordered by platform and label;
use the provided splits, or split by creator_id, rather than splitting videos at random, since otherwise the same creators appear on both sides;
verify that your model is not learning to discriminate based on resolution, bitrate, or codec, since these differ between platforms and between classes (for example, over a third of the real Instagram videos have a short side of at most 480 px, against about a tenth of the AI Instagram videos);
verify that your model is not learning to discriminate based on file names, handles, or other metadata.

Data Samples

AI: Instagram	AI: TikTok	AI: X
AI: Instagram	AI: TikTok	AI: X
Real: Instagram	Real: TikTok	Real: X
Real: Instagram	Real: TikTok	Real: X

Citation

If you use SLOP, please cite:

@misc{susarla2026universal,
  title  = {Universal Representations for Robust AI-Generated Video Detection},
  author = {Susarla, Rhea and Bohacek, Maty},
  year   = {2026},
  note   = {Preprint forthcoming}
}

Contact

Questions, feedback, and removal requests: maty@stanford.edu.

Downloads last month: 15

Total file size:

63.2 GB