TrueVisLies Twitter Dataset

This dataset is used in the paper:

True (VIS) Lies: Analyzing How Generative AI Recognizes Intentionality, Rhetoric, and Misleadingness in Visualization Lies

It is a subset of the dataset collected in:

Maxim Lisnic, Cole Polychronis, Alexander Lex, and Marina Kogan. 2023. Misleading Beyond Visual Tricks: How People Actually Lie with Charts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, Article 817, 1–21. https://doi.org/10.1145/3544548.3580910

Dataset Overview

This dataset contains Twitter (now X) posts sharing data visualizations related to COVID-19, collected between January 2020 and August 2021. Each entry pairs an image of a chart or graph with metadata about the tweet and expert annotations indicating whether the visualization is misleading and, if so, what type of error it contains.

The dataset is intended for research on misleading visualization detection.

Image rights: Images are shared for research purposes only. All original rights belong to X (Twitter) and the respective content authors.

Errors in visualizations have been identified by Lisnic et al. (2023) https://doi.org/10.1145/3544548.3580910

---

Dataset Structure

The dataset consists of the following files:

• index.csv — The main annotation index. Each row corresponds to one image and contains the image_id, the associated tweet_id, a binary is_misleading label, and 14 binary columns indicating which specific error types (if any) are present.
• metadata.json — A JSON dictionary keyed by tweet_id (equivalent to image_id). Each entry includes the tweet author, publication date, tweet text, tweet URL, original image URL, image dimensions, and a structured errors field that subdivides annotations into visualization_design_violations and reasoning_errors.
• images/ — A folder containing all images in PNG format. Each file is named {image_id}.png, where image_id matches the corresponding entry in index.csv and metadata.json.

---

Statistics

Property	Value
Total samples	2,336
Misleading	1,168 (50%)
Not misleading	1,168 (50%)
Image format	PNG

Error Type Distribution

The table below reports the number of misleading images annotated with each error type. Note that a single image can be annotated with more than one error.

Error Type	Count
Causal inference	356
Cherry-picking	300
Value as area/volume	253
Setting an arbitrary threshold	243
Dual axis	224
Issues with data validity	67
Failure to account for statistical nuance	61
Truncated axis	54
Inverted axis	35
Unclear encoding	19
Misrepresentation of scientific studies	18
Inappropriate encoding	11
Uneven binning	3
Incorrect reading of chart	4

---

Fields Reference

index.csv

Column	Type	Description
image_id	int64	Unique image identifier (matches filename in images/ and key in metadata.json)
tweet_id	int64	Twitter post identifier (matches key in metadata.json)
is_misleading	int (0/1)	Whether the visualization is annotated as misleading
error[*]	int (0/1)	One column per error type; 1 if the error is present

Column Descriptions

• image_id unique identifier for the image (matches filename in images/ and key in metadata.json)
• tweet_id id of the tweet associated with the image (same as image_id)
• is_misleading int (0/1) | Whether the visualization is annotated as misleading
• error[Truncated axis] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Dual axis] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Value as area/volume] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Inverted axis] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Uneven binning] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Unclear encoding] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Inappropriate encoding] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Cherry-picking] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Setting an arbitrary threshold] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Causal inference] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Issues with data validity] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Failure to account for statistical nuance] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Misrepresentation of scientific studies] boolean, indicating whether the tweet contains a misleading visualization with this error.
• error[Incorrect reading of chart] boolean, indicating whether the tweet contains a misleading visualization with this error.

metadata.json

The file is a JSON dictionary keyed by tweet_id (equivalent to image_id). Each entry contains the following fields:

Field	Type	Description
tweet_id	string	Twitter post identifier
is_misleading	bool	Whether the visualization is misleading
author	string	Display name of the tweet author
date	string (ISO 8601)	Publication timestamp
text	string	Full text of the tweet
tweet_url	string	Permalink to the original tweet on X
image_url	string	URL of the original image as hosted by Twitter
image_width	int	Image width in pixels
image_height	int	Image height in pixels
errors.visualization_design_violations	list of strings	Design-level errors present in the visualization
errors.reasoning_errors	list of strings	Reasoning-level errors present in the visualization
image_id	string	Image identifier (matches image_id in index.csv)

---

Usage

import pandas as pd
import json
from PIL import Image

# Load annotations
index = pd.read_csv("index.csv")

# Load metadata
with open("metadata.json") as f:
    metadata = json.load(f)

# Load an image
image_id = index.iloc[0]["image_id"]
img = Image.open(f"images/{image_id}.png")

---

License and Rights

Images shared in this dataset are made available for research purposes only. All original rights to the images belong to X (Twitter) and the respective tweet authors.

---

Citation

If you use this dataset in your research, please cite the associated paper.