Are white people perceived to be more credible? Using Deepfakes AI to assess the impact of race on credibility

Abhishek Bais
5 min readApr 18, 2021

On Dec 3, 1847, Frederick Douglass, a runaway slave, who went on to become a civil and human rights icon and the first black man to hold high office in US government, started the abolitionist newspaper, The North Star with the motto Right is of no Sex — Truth is of no Color — God is the Father of us all, and we are all brethren.” (1847)

Frederick Douglass c. 1855, 1st edition of The North Star, Dec 3 1847, from The Library of Congress

In those times, escaped slaves, normally, kept a low profile, but Douglass was different. He risked getting recaptured to become a powerful voice against slavery, for truth and justice. His struggles were not his alone. They were the struggles of many black people of that era.

174 years later, as protests and conversations over #blacklivesmatter dominate the news headline, the big question remains, “Is truth and justice color-blind?” In their study, Could you become more credible by being White? Assessing Impact of Race on Credibility with Deepfakes”, the authors Kurtis Haut, Caleb Wohn, Victor Antony, et. all, employing Deepfakes Artificial Intelligence (AI), attempt to find answers.

For their study, the authors recruited a group of 800 college students to participate in a paid survey. The self reported demographic profile of 305 of those 800 participants is shown below.

Self-reported demographic information of 305 participants

The survey had 2 parts, an Image condition, and a Video condition.

The Image condition

Here, each participant was shown a still image of a black and a white person, generated using Cycle Generative Adversarial Networks (CycleGAN), a deep learning algorithms that learns underlying structures of the given data, without knowing the target value, run on the Chicago Face Dataset, and made to listen to an audio clip. The participants were told that the speaker was the person whose image they were shown.

The video condition

Here, each participant was shown an original video of a darker-skinned South Asian speaker with subtle changes made to the face area of the speaker using Deepfacelab, a simple, flexible and extensible face swapping framework that allows precise manipulation of skin tone, hair type, eye color, and facial features.

The transcript of the audio is shown below. Both videos had identical audio.

a) Depicts the image condition. b) Depicts the video condition

Subsequently, each participant was asked to answer a set of questions shown below.

Questions for the surveys

Analysis of survey responses

The survey results were analyzed in 3 steps with the goal to assess how race influences perceived credibility of the speaker.

Pre-processing to Identify Perceived Race

Here, all responses where the participant did not ”correctly” identify the race of the speaker were filtered out.

Results for the racial perceptions of the participants from the question “What race do you think the speaker is?

Assessing perceived credibility

Here, all responses where the participants thought the speaker was truthful were taken and separated into groups matching the speaker’s race. Next, a Proportions Z-Test was performed to compare and contrast the difference in proportions of perceived credibility amongst the groups.

Summary of Results: “n” is the number of responses after pre-processing. “Credibility” is the percent of responses that believed the speaker was telling the truth. * indicates p < 0.05

For the Image Condition, the authors found, no statistically significant difference in perceived credibility. While, 72.3% participants believed the speaker was truthful when an image of a White person was shown, 70.3% participants believed the speaker was truthful when an image of a Black person was shown to them.

For the Video Condition, the authors found, a significant statistical difference (p < 0.05) in perceived credibility. While, 61.0% of participants believed the speaker was truthful after watching the original video of the darker-skinned South Asian speaker, 73.0% participants believed the speaker was truthful after watching the altered video featuring a White speaker.

Analyzing Sentiment

Here, all responses to questions explaining the rationale why the participants thought the speaker was truthful or not along with all responses to questions where the participants described the speaker’s characteristics in a few adjectives were taken and VADER Sentiment Analysis was performed on them to associate sentiments with racial perception.

Results from VADER compound sentiment analysis of responses to the question “What made you think the person was lying or telling the truth?”
Results from VADER compound sentiment analysis of responses to the question “Use a few adjectives to describe the characteristics of the speaker?”

The study provides valuable insights into how people’s perception of who is credible and who is not changes as visual representations of race changes. For the video condition, where the racial adjustments were more pronounced vs the image condition where the racial adjustments were more subtle, a higher percentage of participants (73.0% vs 61.0%) believed the White speaker to be truthful vs the darker South Asian speaker. In addition, a higher positive sentiment was found associated with the White speaker.

While it is possible that participant responses may have been influenced by technological factors such as more pronounced video alterations and demographic factors such as bias towards younger, white males from the US, other probable factors with deep societal implications such as unconscious bias should not be buried under the carpet but rather carefully examined.

Today, rapid advancements in Artificial Intelligence have made available tools such as Cycle Generative Adversarial Networks (CycleGAN), Deepfakes Artificial Intelligence (AI) and Deepfacelab at our disposal. These tools empower us to perform survey based studies on a wide array of difficult subjects such as racial justice, bias and prejudice. In the age of data driven decisions, emotions should be left aside and technology used to find answers to difficult questions such as “Is truth and justice color-blind ?” to then take appropriate follow-up actions to leave behind a better planet for our children.

“You are not judged by the heights you have risen, but from the depths you have climbed” ~Frederick Douglass



Abhishek Bais

Seasoned R&D EDA, Data Science Enthusiast, Cultural Explorer