Join the fight against online content manipulation and information pollution - Request Beta Access →
Forensic Analysis

AI-Generated Political Ads: Conducting a Forensic Analysis

By Lisa Singh, Sejin Paik, Mohamed Ahmed

As AI-generated ads become harder and harder for humans to identify, tools that help us decipher real from fake media content help us retain agency over our information environment.

As AI-generated ads become harder and harder for humans to identify, tools that help us decipher real from fake media content help us retain agency over our information environment. We know from the Wesleyan Media Project that Republicans and Democrats had spent over $137 million on AI-generated ads by mid-Oct. If all ads were accurate and candidates were transparent about the ads being AI generated, this transition to generated media may not be concerning. However, this is not the case. Over 70% of ads do not disclose AI use even with a growing number of state laws around AI labeling of political ads.  Most ads are not labeled as AI-generated and too many of them contain images that change the physical appearance of their opponents in negative ways or misinform the public about their opponent’s position on an issue. For example, when auditing OpenAI, 80% of prompts requesting creation of disparaging images about political candidates were generated. 

Here is one example of an AI-generated mailer that was sent to residents in Wisconsin. In this particular case, the political campaign did disclose that the image was AI-generated.

TrueMedia.org - Deepfake Analysis

TrueMedia

TrueMedia analysis of the ad: Substantial Evidence of Manipulation

Forensic Analysis:
‍

  • Hands are unnaturally different
    ‍
  • Teeth perfectly symmetric 
    ‍
  • Face overly symmetric

With the election only a couple of days away, we encourage you to not trust political ad content you see. Be skeptical. Be curious. And most importantly, use tools to help you assess the authenticity of the media you engage with. There are still some AI-generated/manipulated images that have visible flaws and AI-generated/manipulated audio that have audible flaws. Media forensic analysis focuses on determining the authenticity of media. Below is a table you can use to help you conduct manual media forensics analysis. Training yourself to be more vigilant and using tools to help when you are uncertain are essential for accurately assessing the credibility of the media you engage with.  

Category Key Indicators & Examples
Visual Discrepancies
  • Distortion of body parts (e.g., eyes too big, missing limbs, short neck, distorted fingers)
  • Overly perfect people (e.g., skin too smooth, perfect teeth, flat clothing)
  • Reflection and shadow issues (e.g., unnatural reflections, misaligned shadows)
  • Background issues (e.g., broken, vanishing, floating, or duplicate objects)
Audio Oddities
  • Weird gaps in phrases or between words, sometimes very long gaps
  • Voice changes
  • Too perfect audio
  • Irregular breaths, flat monotonic pitch, unchanging speech speed, abrupt starts/stops
Text Inconsistencies
  • Unreadable or garbled text (e.g., odd words on signs or t-shirts)
  • Non-standard characters (e.g., fake alphabets)
Context Issues
  • Logical inconsistencies (e.g., person at wrong location, climate mismatch for events)
  • Emotional triggers (e.g., extreme negative or positive visual content requiring verification)
Watermarks
  • Visible brand logos (e.g., AI platform logos)
  • Invisible digital watermarks (e.g., SynthID detected by TrueMedia)
Metadata
  • File creation details & provenance info (e.g., image (photo) creation tool/camera, creation timestamp, author metadata)