Global South Left Vulnerable as AI-Fakes Detection Tools Fail to Recognize Diverse Content

To go back to this article, go to My Profile and then click on Saved Stories to see it again

Efforts to Identify AI-Generated Deceptions Are Letting Down Electorates in Developing Countries

In a recent incident, Donald Trump, the ex-president and convicted criminal, shared multiple images which seemed to depict admirers of the pop icon Taylor Swift endorsing his campaign for the U.S. presidency. The photos appeared to be produced by artificial intelligence, and upon examination using the detection tool provided by the non-profit organization True Media, WIRED verified that there was significant proof of tampering.

Navigating the landscape isn't straightforward. The adoption of generative artificial intelligence for political ends is on the rise globally, and WIRED has been monitoring its deployment in electoral processes worldwide. However, outside the United States and certain European regions, identifying content produced by AI poses challenges due to inherent biases in how these systems are trained. This leaves reporters and scholars with limited tools to combat the overwhelming wave of misinformation they face.

The field of identifying AI-generated or altered media is rapidly emerging, driven by the rapid growth of companies specializing in generative AI. (In 2023, these startups attracted more than $21 billion in funding.) "Now, there are far more tools and technologies easily available for creating artificial media than there are for detecting it," states Sabhanaz Rashid Diya, the founder of the Tech Global Institute, a think tank dedicated to technology policy in the Global South.

According to Sam Gregory, the program director of Witness—a nonprofit organization that assists individuals in using technology for human rights advocacy—current tools available in the market generally achieve an accuracy rate of 85 to 90 percent in identifying AI-generated content. However, this accuracy significantly decreases when analyzing content from regions like Bangladesh or Senegal, where the subjects are not white or do not speak English. Gregory points out that the development of these tools has been skewed towards specific markets, with a focus on training models using data that primarily include US-accented English or faces that are common in the Western world.

This implies that artificial intelligence models have predominantly been developed based on data catered to and sourced from Western markets, leading to their inability to accurately identify or understand content beyond those confines. Often, this limitation arises because firms have opted to train their models with the most readily accessible online data, which predominantly is in English. Richard Ngamita, the founder of Thraets, a nonprofit organization dedicated to addressing digital threats in Africa and elsewhere in the Global South, points out, “A significant portion of our data, especially from [Africa], exists in physical form.” Consequently, without converting this information into digital format, it's impossible to use it for training AI models.

The lack of sufficient data to properly train AI models for the accurate identification of AI-created or altered content frequently results in incorrect alerts. These models may mistakenly mark genuine content as AI-produced or fail to recognize content that is indeed generated by AI. Diya notes that many readily available tools designed for spotting AI-generated text often misidentify writing by individuals who are not native English speakers as AI-generated. This leads to a high rate of false alarms due to the training data not including diverse linguistic inputs.

However, the issue extends beyond the inability of models to accurately identify diverse accents, languages, structural grammar, or facial features that are less prevalent in Western nations. Gregory points out, “Many of the early tools designed to spot deepfakes were developed using high-quality media.” This contrasts sharply with many parts of the globe, such as Africa, where affordable smartphone brands from China, known for their basic functionalities, lead the market. According to Ngamita, the lower-quality images and videos generated by these devices pose additional challenges for the algorithms tasked with detecting deepfakes.

Gregory mentions that certain algorithms can be overly sensitive, to the point where mere background noise in audio recordings or the act of compressing videos for social media usage might trigger incorrect outcomes. He highlights that these types of situations are common in everyday scenarios, especially when attempting to identify or verify information in less-than-ideal conditions. He points out that the tools freely available to journalists, fact-checkers, and those working within civil society often lack accuracy. This inaccuracy stems from issues such as biased representation within the training data and the difficulties inherent in analyzing low-quality content.

Manipulated media isn't exclusively produced through Generative AI techniques. In the Global South, there's a prevalent use of "cheapfakes," which involve the alteration of media through basic means such as attaching deceiving captions or modifying the pace and editing of audio and video clips. These instances can erroneously be identified as AI-generated by inaccurate algorithms or researchers lacking proper training.

Diya is concerned that the deployment of technologies prone to mistakenly identify non-US and European content as fabricated by AI might prompt policymakers to unnecessarily intensify regulations, targeting non-existent issues. "There's a significant danger in exaggerating those figures," she points out. Moreover, she emphasizes that creating new technologies is not as simple as just pushing a button.

Constructing, examining, and implementing a detection algorithm involves the utilization of energy and data centers, resources that are largely inaccessible in many parts of the globe. "Discussing AI and indigenous solutions in our context is virtually unfeasible due to the lack of computational resources necessary for operating any proposed models," explains Ngamita from Ghana. In the absence of domestic alternatives, researchers such as Ngamita face limited choices: they can either invest in commercially available services like those provided by Reality Defender, which may be financially out of reach; resort to using less reliable free tools; or attempt to secure resources through academic affiliations.

Currently, Ngamita explains that his group has formed a collaboration with a university in Europe to facilitate the verification of certain content pieces. His team is in the process of gathering a collection of potential deepfake examples from various parts of the continent. According to Ngamita, this compilation is proving to be a crucial resource for scholars and investigators aiming to broaden the diversity of their model datasets.

However, transmitting information to another party comes with its own set of challenges. Diya points out, “There's a considerable delay involved.” It often requires several weeks before it can be definitively stated that the content was produced by AI, and by that point, the harm has already occurred.

Gregory mentions that Witness, operating its swift detection initiative, is inundated with an overwhelming volume of cases. "Managing these within the tight deadlines required by frontline reporters, especially given their increasing frequency, is already proving to be a formidable task," he explains.

However, Diya expresses concern that an excessive emphasis on detection could lead to a shift in funds and resources away from entities that contribute to a stronger overall information environment. She advocates for directing financial support towards media organizations and civil society groups that have the potential to cultivate public confidence. "I believe the funds are not being channeled in that direction," she remarks. "It seems the focus is predominantly on detection."

Recommended for You …

Directly to your email: A selection of our most significant stories, curated daily just for you.

A faulty update from CrowdStrike brought global computer systems to a halt

Headline: The Major Narrative: When Will the Atlantic Ocean Split Apart?

Introducing the age of excessive online consumption

Additional Content from WIRED

Critiques and Manuals

© 2024 Condé Nast. All rights reserved. Purchases made through our website may generate revenue for WIRED due to our affiliate relationships with retail partners. Content from this website cannot be copied, shared, broadcast, stored, or utilized in any form without explicit consent from Condé Nast. Ad Choices

Choose a global website