Data Provenance and Bias Mitigation in Deepfake Detection

Generative AI models — those that create or detect content — cannot exist without data

Yet how data is sourced and managed is suddenly a major focal point for discussion, especially as everyone from multi-trillion-dollar tech giants and large regulatory bodies surface highly conflicting viewpoints on the provenance of data and potential biases contained within.

Recently, Reality Defender's Data Engineering and Core AI teams unpacked exactly how we source data and root out potential biases from datasets used to generate our detection models.

-Ben Colman, Co-Founder and CEO, Reality Defender

How Reality Defender Sources Data

Reality Defender emphasizes the importance of high-quality data input for its deepfake detection models, ensuring that they can effectively detect a wide range of AI markers while upholding ethical standards. 

Read the latest post from our Data Engineering team to learn more about our rigorous data auditing process, our collaborations with generative AI platforms, and how our small team is able to punch well above our weight in the creation of datasets for our award-winning deepfake detection platform.

Ensuring Fair and Accurate Deepfake Detection

Reality Defender takes proactive steps to mitigate bias in deepfake detection models through careful curation of diverse datasets and advanced bias discovery and mitigation methods, ensuring fair and accurate outcomes.

Our Core AI team dives deep into exactly how we accomplish bias mitigation in datasets, what we look for, and what we plan on doing moving forward in their latest post.

Tracking Deepfakes With the Deepfake News Center

Used by journalists and industry veterans, Reality Defender's Deepfake News Center is updated daily with the latest news stories — including those featured below. 

After a year in private beta, the news center is now free and open to the public.

Click below to gain instant access.

Recent News

Deepfakes 

The first study released by Google DeepMind revealed that deepfake models are most often used to create impersonations of politicians and celebrities.

Research carried out by Medius shows that nearly two thirds of financial professionals have been targeted by deepfake fraud.

In a first-of-its-kind action, the FCC has proposed $8 million forfeitures against the creator of the Joe Biden robocall deepfake and the cell carrier company that enabled the call.

Creators of nonconsensual explicit deepfakes are revictimizing the survivors of a video-based trafficking ring by swapping faces of celebrities into footage.

Following similar efforts by competing social media platforms, Meta’s new “AI Studio” will allow creators to generate deepfake chatbots of themselves on Instagram.

AI-generated images of an extremely attractive Jesus have become the latest viral deepfake sensation on social media.

Generative AI

A new study showed that participants couldn’t distinguish between restaurant reviews written by humans and ChatGPT-4. 

Photographers are finding their authentic works incorrectly marked as “Made with AI” by Meta’s new provenance labeling system.

A new report from Forrester shows that 91% of US advertising agencies are either using or exploring the use of generative AI, outpacing all other use groups and industries.

Apple has held discussions with Meta about integrating Llama 2 into iOS 18, as a part of its strategy to eventually offer users a choice between competing AI models. 

A UK think-tank recommended that the country’s government should launch a standardized system for the reporting of AI malfunctions and misuses.

Business’ focus on AI’s affordability, rather than speed, could affect Nvidia’s dominance on the AI chip market.

The highly anticipated release of a horse-shaped purse stirred controversy among fans after the brand revealed that AI was used in its design.

Toys “R” Us faced backlash after producing a “first-ever” ad using OpenAI’s Sora.

Researchers at University of Reading created fake student identities and submitted exam answers generated by chatbots, tricking their colleagues and earning better grades than real students. 

YouTube is offering major record labels vast sums of cash to license songs for AI training, as major artists remain fiercely opposed to AI music generation. 

Bill Gates has defended AI power usage, asserting that massive overconsumption will be offset by Big Tech’s investment in green energy.

OpenAI unveiled CriticGPT, an AI model designed to identify mistakes in code generated by ChatGPT and make outputs more accurate.

Investigations by major news outlets show that Perplexity, architects of an “answer engine,” may have engaged in plagiarism and site scraping of media platforms.

The revenue OpenAI makes from selling access to its AI models now exceeds what Microsoft earns from equivalent subscription sales.

The nonprofit Center for Investigative Reporting joined other media outlets in suing OpenAI and Microsoft for copyright infringement, alleging the companies used its content for AI training without permission. 

Microsoft’s head of AI has stated that “social contract” dictates that any content posted on the open web is “freeware” and fair game for AI training.

Government and Policy

Yet another Russian-linked influence group is utilizing fake news websites and deepfake content to manipulate the U.S. presidential election.

A senator backed by an AI investment company supports a new bipartisan anti-deepfake bill after being the only lawmaker to block a similar effort in June.

Arts and Entertainment

Sony, Warner Brothers, and the RIAA lead a group of record labels suing AI startups for allegedly using copyrighted tracks to train AI models.

Sheryl Crow called the AI resurrection of dead artists’ vocals “hateful” and “antithetical” to life.

NBC will use the AI-generated voice of an iconic sportscaster to provide customized recaps of the Paris Olympics on Peacock.

Following Toys “R” Us’ venture into AI advertising, Motorola is launching its own campaign featuring videos of AI-generated models carrying the brand’s logo.

Join the Reality Defender Team

The Reality Defender team consists of diverse experts, scientists, and researchers, all working together behind a vital mission: to stop deepfakes and safeguard truth.

If you share a passion for protecting citizens and empowering communities, we welcome you to join us in building world-class detection tools.

Thank you for reading the Reality Defender Newsletter. If you have any questions about Reality Defender, or if you would like to see anything in future issues, please reach out to us here.