Why AI models are bad at verifying photographs

The Tow Center for Digital Journalism systematically tested the ability of seven AI models (ChatGPT o4-mini-high and 5-Thinking, Perplexity, Grok, Gemini, Claude, and Copilot) on a basic fact-checker’s task: confirming that an image is real, and identifying its location, date, and source. The goal was to evaluate how well these tools function as fact-checkers, observe patterns in how they reason through photographs, and place their role in the fact-checking process.

[Read the piece to learn about our findings] ()

Models Tested

• ChatGPT 5 Thinking
• ChatGPT o4-mini-high
• Perplexity Pro Search
• Grok3
• Copilot
• Claude Opus 4.1
• Gemini 2.5 Flash

Each LLM was tested in both standard and deep research modes, using both downloaded image files and screenshots, for a total of 280 queries per prompt.

Research Questions

1. Provenance
   Prompt: “Where was this taken? When? By whom?”

2. Authenticity
   Prompt: “Is this image real or AI-generated? Why?”

📂Dataset

Access the data here: (https://docs.google.com/spreadsheets/d/e/2PACX-1vQObLR46X3MqUplYiL2VRcYHmFCVbu1vVSHoKONY8q5fCApviHrLfmnzi5UojmVcSpEd9akySHj_sTa/pubhtml)

• There are three tabs:

1. Information about each photograph we tested with
2. Results from the query to test provenance
3. Results from the query to test authenticity

Baseline Comparison

As a benchmark, all 10 images were also run through Google Reverse Image Search, a primary tool used by visual investigators. Only images that Google correctly identified in the top results were included in the study.

Data Collection & Structure

Rows: Unique Model × Mode × Image × Format × Prompt

Provenance Data Table

Column	Description
Event	Photo identifier
Platform	The LLM platform (e.g., ChatGPT, Claude, Gemini) and the model that was tested
Image Type	Whether the input was a downloaded file or a screenshot
Model Type	Standard vs. deep research mode
Full Response	Raw text output generated by the model (often includes reasoning, cited sources, etc.).
Correctness	Overall evaluation of the response (e.g., Completely Correct, Partially Correct, Partially Wrong, Completely Wrong, No Answer)
Location of Photograph	Location identified by model
Confidence_Location	Confidence level in the location answer (self-reported or inferred)
Correctness_Location	Correctness of the location answer. As long as the location was factually correct, it was marked as correct even if the precise location was not provided. Correctness wasn't granted for precision if the answer wasn't factually right ie, in most instances, the answer was marked as correct at the city/town level but if the LLM guessed the region too and the region was incorrect, the answer was marked as Incorrect, even if the city was correct.
Location Accuracy	How precise the location identified by the model was relative to the correct answer. This allows partial credit when the model was directionally correct but not exact
Date of Photograph	Date identified by the model
Confidence_Date	Confidence level in the date answer
Correctness_Date	Correctness of the date answer. The date had to be exact to be marked as correct- no partial credit was granted (Correct/Incorrect/No Answer)
Photographer Name	Model’s identified photographer
Confidence_Photographer	Confidence level in photographer attribution
Correctness_Photographer	Correctness of photographer attribution (Correct/Incorrect/No Answer)