As I See It: The Tells

November 4, 2024 Victor Rozek

In its early incarnations, AI-generated media was often easy to spot: groups of the same people appearing multiple times in a crowd scene; multi-directional light sources in outdoor settings; mouths moving out-of-sync with the words being spoken; people whose anatomy mysteriously sprouted extra fingers or additional limbs.

But things quickly improved to the point that deepfake images and manipulated video are no longer obvious constructions. Misinformation from both domestic and foreign sources is now rampant on the Internet, especially during the election cycle and, at a glance, it’s difficult to know if what we’re seeing is real.

Which is why legitimate news organizations like The Washington Post and social media giants now use a number of AI detection tools to ensure the authenticity of the photographs and videos they publish.

The Post explains how the process works. It involves uploading images, sound bites, or video clips into a deepfake detection tool comprised of a series of algorithms designed to identify indicators of inauthenticity. For example, one algorithm examines the head and face for signs that it was digitally planted on another person’s body. Another tracks unnatural lip movement and abnormal facial expression. Still another analyzes vocal patterns looking for irregular frequencies, out of context pauses, and other unlikely speech patterns.

The algorithms also scan down to the pixel level, examining imagery for patterns of visual disturbance that deviate from surrounding areas and indicate that the image has been altered. They also compare how pixels move between video frames. In authentic videos there is no motion blur nor will the image appear to jerk from one frame to the next.

Then an algorithm tries to reconstruct the image using a technique called diffusion. The goal isn’t to flawlessly recreate the image, but to discover what diffusion was unable to duplicate, thereby indicating possible manipulation.

The last algorithm looks for a unique pattern in the pixel distribution that would indicate the content was created using an earlier image generation technique called GAN, short for generative adversarial network.

Each algorithm estimates the probability of image tampering. As a final step, these deductions are blended and the detection tool offers its conclusion on whether the content is likely genuine or fake. It all sounds rigorous and thorough. Nonetheless, it’s far from foolproof.

The Post reports: “Last year, researchers from universities and companies in the United States, Australia and India analyzed detection techniques and found their accuracy ranged from 82 percent to just 25 percent.”

The problem, as articulated by Hany Farid, a computer science professor at the University of California at Berkeley, is that “the algorithms that fuel deepfake detectors are only as good as the data they train on. The datasets are largely composed of deepfakes created in a lab environment and don’t accurately mimic the characteristics of deepfakes that show up on social media.”

Which is why the government is pressuring tech companies to create some sort of embedded identifier or watermark in their AI products; on-line labels that would identify AI-generated content. The fear is that as AI improves it will become increasingly difficult to spot deepfakes. And, even if AI identifiers existed, unscrupulous users would likely find deceptive workarounds. The hope is that at some point technology may be able to police itself. And it may, but it can never remedy the impacts resulting from the lack of integrity of its users.

Ultimately, deepfakes are not created to convince us of the validity of any particular image, but to cast doubt on everything that is real, that is factual, that is truthful. It seeks to sow confusion and distrust so that policy can be forged by corrupt people with dishonorable intentions.

In a world where no one can trust what they see or hear; when sensory evidence is manipulated for nefarious purposes, what remains is division and suspicion. As satirist Alexandra Petri quips, when facts don’t matter, the question becomes: “Who has created a nicer story for you?” Or perhaps a more frightening one.

Conceivably even more frightening than the use of AI as a tool of deception, is its use as a tool of depraved persuasion.

The Post reports that in a horrific effort to radicalize ignorant and impressionable young people: “Extremists are using artificial intelligence to reanimate Adolf Hitler online for a new generation, recasting the Nazi German leader who orchestrated the Holocaust as a “misunderstood” figure whose antisemitic and anti-immigrant messages are freshly resonant in politics today. In audio and video clips that have reached millions of viewers over the past month on TikTok, X, Instagram and YouTube, the führer’s AI-cloned voice quavers and crescendos as he delivers English-language versions of some of his most notorious addresses…”

And, if you’re wondering as I did, just how many people would be interested in listening to something so hateful and repugnant? Researchers at the Institute for Strategic Dialogue found that “content glorifying, excusing or translating Hitler’s speeches into English has racked up some 25 million views across X, TikTok and Instagram since August 13.”

All of this creates a dilemma for a country steeped in free speech. That right is not absolute, of course. Yelling “fire” in a crowded theatre, for example, is frowned upon. But lying is essentially protected. What isn’t protected is an absolute right to amplification, which unrestricted AI loosed upon the internet all but guarantees.

Out of concern that politicians could use AI to mislead or deceive voters, a number of states are passing laws requiring candidates to disclose when they use the technology to generate political ads. But laws seldom deter liars, and ads can be launched from unknown sources.

John Hopfield is a physicist and a professor emeritus at Princeton. This year he won the Nobel Prize in physics for “foundational discoveries and inventions that enable machine learning with artificial neural networks.”

But Hopfield was apparently unsettled by his own success. He said that recent advances in AI technology were “very unnerving” and warned of possible catastrophe if not kept in check.

Ultimately, many of the issues surrounding AI regulation and free speech may be decided by the Supreme Court. And that may be the biggest catastrophe of all.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search