The social media post from Santa Ana police late last month was alarming: “SNAKE SIGHTING ALERT,” it read, aside a local resident’s menacing photo of what looked to be a highly venomous snake that might be slithering around town.
Some astute Facebook users immediately suspected the photo of the coiled Gaboon viper was fake – maybe generated by AI. But other Santa Ana residents understandably responded with alarm. Police quelled some concern by saying they were bringing in a snake expert to help investigate the incident.
Within a day, on Sept. 29, authorities sent out a second announcement: The snake photo was indeed fake, and the needless scare was the result of a local resident’s apparent hoax.
While it turned out that image was generated with AI, investigators could not immediately verify the fraud. The hoax stoked “unnecessary fear, confusion and concern in our community,” according to police.
As major AI players such as OpenAI, Google and Anthropic refine generative AI models, image generation has become more commonplace — and more difficult to distinguish. The snake hoax raised questions about government agencies’ ability to verify AI-generated media as it becomes more prevalent and accessible.
“Seeing is definitely not believing today,” said Gene Tsudik, a distinguished professor of computer science at UC Irvine.
Santa Ana’s fake snake is not the only computer-generated image that has caused alarm recently.
In November 2025, AI-generated images showing a Texas high school on fire circulated online, sparking concern even though there was no fire. In May, a Louisiana high school was put on lockdown after a student created an image of an individual with a gun sitting in a window that law enforcement could not immediately verify as fake. And on Oct. 4, an Arizona court threw out a sentence after the family of a road rage victim used his likeness to generate an AI video played during court proceedings.
“Now we have to contend with people potentially submitting AI-generated images as valid evidence,” said Melodi Dinçer, director of the Digital Likeness Initiative at UCLA’s Institute for Tech Law and Policy. ”So law enforcement agencies are going to have to deal with this. The judiciary is trying to deal with this and get ahead of this issue. This is going to affect the legal process in many different areas.”
In an attempt to get ahead of deepfake concerns, California implemented its AI Transparency Act on Aug. 2. The bill, passed in September 2024, requires any AI company that has more than 1 million users in the state to offer publicly accessible tools to detect provenance data, which identifies a text, image or video as AI. In May, OpenAI released its verification tool, which detects C2PA metadata and SynthID watermarks — two forms of provenance data embedded in AI content that are difficult to remove. Google SynthID, originally released in 2025, announced Oct. 7 that its new detector can detect AI presence in Google-generated AI images, videos and audio files, as well as in content generated by fellow companies OpenAI, NVIDIA and Kakao.
According to several district attorney offices and law enforcement agencies across Southern California, departments use these publicly available tools, but investigators tend to focus on traditional verification techniques even as AI becomes more believable. Investigators have been facing potential doctoring through applications like Photoshop long before AI-generated images hit the mainstream, meaning that many are facing the same authenticity challenges as before, just with a different source, said Jacquelyn Rodriguez, a public affairs officer with the San Bernardino County District Attorney.
The Los Angeles County Sheriff’s Department said their fraud and cyber crimes bureau does have access to authentication tools but that “AI verification tools represent only a small part of the overall process used to determine whether a photograph or video is authentic.” The department’s detectives primarily rely on traditional investigative methods and human examination, establishing how evidence was obtained and examining original recording devices to determine if content is authentic.
“These investigative steps remain the most important factors in determining the authenticity and reliability of photographic and video evidence,” the department said in an email.
Eventually, technology to counteract AI verification tools could make watermarks and other current authentication technology unreliable in what Tsudik describes as an AI “arms race.”
“It’s the same as those cartoons — like the roadrunner and the coyote, or Tom and Jerry,” Tsudik said. “Nobody ever really wins. At a given point in time, today, maybe you can detect some of these things with sophisticated state-of-the-art tools. But tomorrow — really, sometimes things change in a matter of days or weeks — the same tools won’t be able to detect.”
Dinçer said that a multifaceted approach to authenticating images and videos should involve manual and forensic techniques designed for analyzing media for signs of AI. For example, AI-generated images tend to have unnatural lighting patterns that do not directly trace to a source of light. The perspective is usually off, as well; for example, two parallel lines in an image usually meet at a vanishing point in the distance, which AI often distorts.
California will roll out requirements for social media companies to display disclosures for AI-generated content next year, and in 2028, the regulations will extend to devices that can capture images and videos, Dinçer said, building on the existing state regulations for AI companies. Apple announced in September that it was introducing Apple Reference Image in its iPhone 18 Pro models, which will sign the pixels of images and serve as a “digitally negative” that can be used as a comparison to see if any edits were made to a photo. Apple also plans to integrate metadata and SynthID markers to allow users to determine authenticity.
The idea is to “capture the harm more upstream rather than downstream,” Dinçer said, focusing on companies to make changes rather than individuals.
“The greatest societal danger is this great undermining of societal trust in content,” Tsudik said. “I think the only hope we can have is if we pressure the manufacturers of equipment that records videos and snaps photos to incorporate security measures.”