Web Archives for Verifying Attribution in Twitter Screenshots

TL;DR

The paper tackles the problem of verifying attribution in Twitter screenshots to curb mis-/disinformation by combining OCR-based extraction of the handle, timestamp, and tweet text from images with evidence gathered from live web searches and the Internet Archive's Wayback Machine. It introduces a workflow that narrows archival search space using a ±$26$ hour window around the screenshot timestamp and a tweet ID prefix derived from TweetedAt, followed by text-overlap scoring to confirm attribution. Empirical results show high reliability: 100% timestamp extraction accuracy, 91% handle extraction accuracy on 4,504 screenshots, and an optimized 80% text-overlap threshold yielding an F1 score of 0.96 on 1,571 unique tweets, with archival evidence supporting real posts and limitations acknowledged for deleted or non-archived items. The work offers a practical, automated approach to surface archival evidence for tweet attribution in screenshots, with plans to extend to more platforms, complex screenshot types, and additional archives to strengthen misinformation detection and validation workflows.

Abstract

Screenshots of social media posts are a common approach for information sharing. Unfortunately, before sharing a screenshot, users rarely verify whether the attribution of the post is fake or real. There are numerous legitimate reasons to share screenshots. However, sharing screenshots of social media posts is also a vector for mis-/disinformation spread on social media. We are exploring methods to verify the attribution of a social media post shown in a screenshot, using resources found on the live web and in web archives. We focus on the use of web archives, since the attribution of non-deleted posts can be relatively easily verified using the live web. We show how information from a Twitter screenshot (Twitter handle, timestamp, and tweet text) can be extracted and used for locating potential archived tweets in the Internet Archive's Wayback Machine. We evaluate our method on a dataset of 1,571 single tweet screenshots.

Figures (17)

• Figure 1: Multiple people sharing screenshot of a fake tweet multiple times on different social media platforms made to appear to be by @tedcruz shared on Twitter. The Facebook post was shared over a thousand times, the Instagram post was liked over a thousand times, and the Twitter post was retweeted over 14 thousand times.
• Figure 2: Screenshot of a deleted tweet and its archived version.
• Figure 3: Because this humorous fake tweet never happened, it is not on the live web or web archives.
• Figure 4: An example fake tweet created using Tweetgen.
• Figure 5: Using tweet text to verify the attribution of the tweet in the screenshot using a fact-checking website.