Just two days before Slovakia’s elections, Michal Šimečka, a leading candidate opposing the pro-Russian populist party, was targeted by a widespread misinformation campaign involving a deepfake video. The video falsely depicted a conversation between Šimečka and journalist Monika Tódová, suggesting he had rigged the election. The AI-generated content spread rapidly across social media, fueling disinformation and potentially influencing voter sentiment, contributing to his defeat.
“It does sound like me,” Šimečka told CNN, referring to the audio, which he said played into conspiracy theories that a segment of the population already believed.
Another story emerges in a suburb outside the city of Baltimore, where an AI-generated audio clip falsely depicted Principal Eric Eiswert of Pikesville High School making derogatory comments, sparking outrage within the town. However, after a thorough investigation, police traced the recording to the school’s athletics director.
When you listen carefully, though, there are clear edits between sentences - and the voice, while similar to the principal, sounds quite monotonous. Artificial intelligence can use several minutes of a real recording - from, say, your favourite actor in a film or a presidential candidate giving a speech - to then generate a clip that makes it sound like they said something they never did.
A study published in Nature, titled 'Experts fail to reliably detect AI-generated histological data' , reveals a crucial finding: despite extensive image training, even skilled professionals struggle to consistently differentiate between genuine and artificial histological images. With 61% of organizations witnessing a rise in deepfake incidents and Deloitte predicting Generative AI will increase losses from deepfakes to $40 billion annually by 2027, it’s crucial for emerging technologies and influential organizations to closely monitor such developments.
Currently, in terms of mitigation strategies, relevant examples include a law passed in Minnesota that makes it a crime for someone to develop a deepfake intended to harm a candidate within 90 days of an election, California’s enactment of three new bills to curb the spread of deceptive election content, and the US Federal Trade Commission (FTC) proposing new laws to protect against AI-generated deepfakes.
Although research within the realms of AI-detection in imaging has captured the attention of industry leaders, there is great skepticism towards the true effectiveness of proposed strategies like watermarking. For example, Google DeepMind’s release of SynthID presented a captivating breakthrough in watermarking capabilities for AI-generated images. The system employed imperceptible digital signatures embedded within the pixel matrices, utilizing advanced neural network architectures to ensure robustness against common image manipulations. However, UMD's Soheil Feizi, a leading figure in AI detection, discovered how easy it was to evade these watermarking attempts (in addition to demonstrating how attackers might remove watermarks, the pre-published study shows how it’s possible to add watermarks to human-generated images, triggering false positives).
This tempering of expectations may already be happening. In its blog post announcing SynthID, DeepMind is careful to hedge its bets, noting that the tool “isn’t foolproof” and “isn’t perfect.”
If research demonstrates that AI watermarking is not a strong solution to disinformation, then what is the answer?
One promising solution involves utilizing blockchain technology to create a transparent and immutable record of content origins. Unlike watermarking, which can be easily removed or manipulated, blockchain provides a permanent, decentralized ledger that tracks the entire lifecycle of digital content from creation through distribution. This system could work by:
1. Assigning unique cryptographic signatures to content when it's first created
2. Recording each modification in the blockchain
3. Enabling anyone to verify the authenticity and history of content
4. Creating an unbreakable chain of custody for digital information
However, blockchain alone is not enough. The most effective approach integrates multiple strategies, including robust content authentication systems, advanced AI detection models, and digital signatures from trusted sources. Additionally, a strong technical infrastructure must be complemented by enhanced media literacy education to empower individuals to critically evaluate information sources.
“When you put that content on chain, you can now validate that content was created by a certain individual or brand”