Manipulating Evidence with AI: A Look at How Images, Audio, and Video Can be Altered

Exploring the Implications and Challenges of AI-Generated Evidence in the Legal System

Introduction

All litigators know well the Rules of Evidence and their application to the authenticity and admissibility of image, audio and video evidence. Concerns about the easy manipulation of these types of evidence have existed for more than a century. The age of AI is going to forever alter the reliability of such evidence in the eyes of the public and will be an issue in civil and criminal hearings and trials. Why is this new advance of AI different than the transition from early black and white photographs to iPhone images or the phonograph to CDs and now digital audio? It is different because now it is easy to undetectably alter existing content in these categories and, of even more serious concern, manufacture content in these categories that claim to represent events or conduct that never occurred.

The Federal Rules of Evidence (FRE) are mirrored in most states so they will be used as reference for this post.

Under FRE Rule 901, the proponent of the evidence must provide sufficient evidence to demonstrate that the audio, image, or video evidence is what it purports to be. This can include testimony from witnesses who can authenticate the evidence, or other circumstantial evidence that supports the authenticity of the evidence.

FRE Rule 902 provides for the self-authentication of certain types of evidence, including certified copies of public records, newspapers and periodicals, and commercial paper. However, this rule does not apply to audio, image, or video evidence.

FRE Rule 1001 sets forth the definition of "writings, recordings, and photographs," which includes audio, image, and video evidence. This rule requires that the proponent of the evidence provide an original or a duplicate of the evidence, unless the original is unavailable, and that the evidence is authenticated as being what it purports to be.

Finally, FRE Rule 401 defines the standard for relevant evidence, stating that evidence is relevant if it has any tendency to make a fact more or less probable than it would be without the evidence.

As AI technology continues to advance, the potential for manipulation of audio, video, and image evidence is becoming a growing concern. With the ability to create and/or manipulate these categories of evidence, it becomes increasingly difficult to determine what is real and what is not. The implications of AI-generated evidence in legal cases can be far-reaching, from the potential for false convictions to the manipulation of public opinion.

Deepfakes

Deepfakes have become a growing concern in the law. A Deepfake is video content of a person speaking and moving as in a normal video, however, the person depicted never said or did the things presented in the video. This technology has been widely available for more than five years now. In fact, there is a growing supply of Deepfake pornography that appears to depict famous actresses, athletes and even politicians engaged in conduct that never occurred.

A deepfake is an AI-generated video or image that appears to be real but is actually manipulated to depict something that did not happen. Deepfakes are becoming increasingly sophisticated and can be used to create fake news, political propaganda, or even revenge porn.

Deepfakes present a significant challenge for prosecutors, defense attorneys, and judges who rely on audio, video, and images as evidence. They can be used to create false evidence that is difficult to detect and verify, with the attendant risk of wrongful convictions and wrongful acquittals.

For example, a deepfake video could be used to show a defendant committing a crime they did not actually commit. If the video is convincing enough, it could sway a jury and result in a wrongful conviction. Similarly, a deepfake audio recording could be used to make it appear as though someone said something they did not actually say, which could be used as evidence in a harassment or defamation case. The other side will likely use this technology as well. Imagine the capabilities of creating deepfake videos to appear to exonerate guilty persons in high stakes cases involving homicide or sexual assault.

To address the challenge of deepfakes in legal cases, experts are exploring new tools and techniques to detect and prevent their misuse. For example, researchers are developing deepfake detection algorithms that can analyze videos and images to determine whether they are real or manipulated. These algorithms use machine learning to detect patterns and inconsistencies in the video or image that are not present in real footage.

Additionally, there are efforts underway to educate judges and attorneys about the dangers of deepfakes and the need to verify evidence before using it in court. This includes providing training on how to identify deepfakes and how to work with experts who can help verify the authenticity of audio, video, and image evidence.

Overall, deepfakes represent a significant challenge for the legal industry and highlight the need for new tools and techniques to detect and prevent their misuse. By staying informed about the latest developments in deepfake detection and verification, attorneys and judges can help ensure that justice is served and that false evidence is not used in legal proceedings.

Deepfakes In Audio Evidence

Artificial intelligence (AI) has brought about significant advancements in many areas, including the creation of audio recordings of people saying things they never actually said. This technology, known as AI-generated audio, presents a new challenge in the use of audio evidence in legal cases, as it raises questions about the reliability and authenticity of the evidence presented.

The process of creating AI-generated audio is relatively simple. The software uses deep learning algorithms to analyze recordings of a person's voice, then generates new speech based on that analysis. This new speech can be manipulated to create audio recordings of the person saying anything, even if they never actually said it.

For example, a deepfake audio recording could be created to make it sound like a person admitted to committing a crime they did not actually commit. This type of false evidence could be very convincing, as it would sound like the person's actual voice.

The use of AI-generated audio also raises concerns about the reliability of audio evidence in general. As more people become aware of the potential for AI-generated audio to be used to create false evidence, they may become more skeptical of the authenticity of all audio evidence presented in court. This could make it more difficult for attorneys to convince juries of the validity of their evidence even when the evidence is legitimate.

Wholly AI Created Audio

Deepfake audio uses existing audio content and manipulates it as noted above. There are other tools that create AI audio from nothing. Imagine an audio clip of an unknown person that makes a statement exonerating or inculpating a defendant. There is no way for a court to know whether a given audio recording is of a real person or not. And, as the Federal Rules currently stand, all that is required for the proponent of that evidence is that they put on a witness who can claim that the audio file was recorded by them and unaltered since its recording. From that point forward, the litigators can argue to the jury about the reliability of that evidence, but it is going to be admitted and can be considered by the jury. The potential for civil and criminal fraud within this reality is enormous.

This clip below contains words that sound like they were spoken by Donald Trump but were never uttered by him.

There are so many free and low cost audio deepfake tools that there are regular articles reviewing the top ten offerings in that category. (Top Ten Audio Deepfake Tools)

This tool creates audio using synthetically created but authentically sounding voices. Try it out for yourself.

This audio clip is entirely synthetic. As you listen to it, realize, this technology is not going to get worse and will only improve to the point of indistinguishability from actual human speech/audio.

The tool linked above also includes a variety of voices, inflections and even emotions. You can set the output to be anxious, or angry or happy, etc.

This tool has on its website an example of human audio with various features that the user might want to remove. For example, pauses in which the speaker utters ‘uhs’ and ‘ums.’ Descript.

Video Filters on Authentic Video

With the rise of social media platforms like TikTok and Instagram, the use of filters on video content has become increasingly common. While these filters are intended to enhance the video and make it more engaging, they can also be used to manipulate the video content and create false evidence.

Filters are digital effects that can be applied to video footage to change the appearance of the subject or background. These effects can range from subtle color corrections to drastic changes in facial features and surroundings. While these filters are designed for entertainment, their usage to create fake evidence is obvious.

This video shows examples of what is called the “Teenage Filter” that TikTok users can apply to video. The filter alters the appearance of the subject showing them their own live video but with their image modified to appear as if they are age regressed to their teenage years.

Note how the filter not only creates impressive age regression results, but also creates in many of the subjects a visceral, emotional reaction. Why is this significant? Because, that underlines the power and danger of video as evidence in court. Video has the capacity to affect the emotional state of viewers (jurors). And, as many of us know, once a person’s emotional centers are activated, their ability to apply reason and logic is dramatically reduced. The ability to modify video content to elicit those emotional reactions is yet another concern that the current Rules of Evidence do not take into account.

The TikTok glamour filter is also having a moment. Videos are appearing online of the undetectable to clearly obvious (and dramatic) changes to live video produced using this filter.

Filters like these and others to come online in the years to come powered by AI computational power, could be used to make it appear as though someone was in a certain location at a certain time, when they were actually somewhere else. This false evidence could be used to create an alibi or establish guilt, depending on the circumstances of the case. Additionally, filters can be used to alter a person's appearance or behavior, which could be used to discredit witnesses or manipulate the jury. The unknown question is, have these filters already been used in a criminal or civil case and their use not yet revealed?

The Mousetrap

One of my programmer friends once told me about privacy, security and encryption and described it this way: “Any lock a human can make, a human can break.” Despite that axiom, many companies are engaged in creating tools, powered by AI, that will attempt to detect deepfakes. MIT has been refining this technology for years and continues to produce more sophisticated detection tools.

One approach these tools take is to develop algorithms that can analyze video footage to detect whether it has been manipulated with a filter. These algorithms use machine learning to identify patterns and inconsistencies in the video that are not present in expected authentic footage. They are trained on hundreds of thousands of hours of authentic footage. However, whatever these tools are being trained to look for, AI deepfake creation tools can be trained to eliminate from their output.

The Ease of Image Manipulation: How It Affects the Use of Evidence in Legal Cases

Photoshop has been around for nearly 40 years at this point. However, until about five years ago, creating undetectable image manipulation was possible, but took some work to develop skills in digital image editing. Those days are over.

In age of AI, the manipulation of images has become increasingly easy and accessible. With just a few clicks, anyone can alter an image to change its content, context, or even its meaning. While this can be useful for artistic or creative purposes, it also raises concerns about the authenticity and reliability of images used as evidence in legal cases.

Image manipulation could be used to create a false alibi or to discredit a witness's testimony. This type of manipulation could be particularly damaging in cases where there is little other evidence to support a conviction or acquittal.

There have been for years free demonstrations of the power of machine learning and AI to create images from nothing. The website thispersondoesnotexist.com does just that. Visit that link and every time you click the image creation button, an AI tool in the background creates a wholly fictional, but indistinguishably real image of a person who…you guessed it…does not exist.

Other AI tools enable users to undetectably remove objects from images. Other tools enable the sharpening of otherwise blurry images. Both of these tools are now on board the latest mobile devices as well. The technology will only become more sophisticated and less expensive as time goes on.

As with DeepFakes above, there are a host of companies producing tools to detect manipulated images.

What Will Happen To The Rules?

The law evolves slowly while the pace of technological development. especially development fueled by AI, improves rapidly. That comparison will have to change if our legal system is going to attempt to minimize the use of fraudulent evidence in civil and criminal cases. These tools are only going to get better, cheaper and harder to detect going forward. It will not be a linear improvement, it will be exponential.

How do you think the Federal Rule (and their state counterparts) should be modified to grapple with the potential use of AI tools that can modify and create video, audio and images?