What NewsGuard found
NewsGuard says Google's Gemini and Microsoft's Copilot repeated the false claim that the former student had made the accusation. The newsletter's Copilot screenshot caption says Copilot stated that the supposed student publicly alleged the abuse "while Lewis was a student."
Google's AI Overview did the same, saying the supposed former student "published a video claiming" an assault. That one is especially instructive. NewsGuard says the Overview cited an ABC News article from May 2026 and a KTAL article from Oct. 15, 2026. Neither mentioned the allegation. The ABC piece covered Talarico's views on religion and gender, and the KTAL piece covered his biography and policy positions. NewsGuard corrected an earlier version of its report that had described KTAL as covering San Antonio. KTAL covers East Texas.
The report is dated Oct. 21, 2026. NewsGuard ran its Google search at 9 a.m. Eastern on Wednesday, Oct. 21. It does not give a time for the Copilot test. Microsoft did not respond to NewsGuard's email asking about it. That silence is not a confirmation or a denial. The report doesn't establish why Copilot answered as it did, how many people saw the answer, or whether Microsoft has since changed it.
Why NewsGuard says the video is fake
The following details are NewsGuard's findings. I could not verify them independently.
- Distribution: An X account, @Zyon_Lewis_5, posted two videos on Oct. 20. They showed a man claiming to be a former sixth-grade student of Talarico's at Rhodes Middle School in San Antonio. The post drew about 33,000 views and 200 likes in a day.
- Visual and audio tells: NewsGuard cites unnatural eye movements and blinking and mostly static facial expressions. It also notes the speaker's teeth differ from those of the real Zyon Lewis in authentic Facebook photos. The speaker has an African accent, and the caption uses the British spelling "behaviour." NewsGuard found no sign on Lewis's Facebook account that he ever lived in Africa.
- A real person was borrowed: NewsGuard says a real Zyon Lewis attended Rhodes Middle School. His Facebook profile photo matches the X account's photo. The person in the video resembles him, but NewsGuard concluded the video was AI-generated.
- Detection tool: NewsGuard says Hive rated clips from the videos as 98 percent likely to be deepfakes. A detector score is a probability, not proof. Here it backs up the other evidence.
- Account details: X's records, as NewsGuard reports them, show the account was created in September 2026 with Romania as the signup country.
How solid is the Russia attribution?
NewsGuard says the format and amplification pattern closely match Storm-1516. It describes Storm-1516 as a Russian operation that often posts AI deepfakes from newly created accounts to spread false claims about Western politicians. It also says the anonymous Gnida Project attributed the videos to Storm-1516 in an Oct. 21 X post. I confirmed that post exists. It says the latest Storm-1516 campaign targets Talarico.
Both are pattern-matching assessments, not public proof that the Kremlin directed this video. NewsGuard's headline is more definitive than its body, which leans on resemblance and a third-party attribution. That caveat belongs next to the headline.
The history is real, though. Storm-1516 has been documented by Microsoft Threat Analysis Center, RUSI and NewsGuard. Microsoft's own August 2024 report described a staged video from this operation. NewsGuard also compares the Talarico claim to a false October 2024 allegation about Tim Walz and a supposed exchange student. In that case, NewsGuard says the State Department had no record of such a student, and a school district spokesperson called the claim "outlandish."
Microsoft's own warnings about Copilot
Microsoft's documentation doesn't promise that Copilot is accurate. Its Copilot privacy FAQ says Copilot searches the web, summarizes what it finds and cites sources. It also says Copilot may misrepresent that information. Responses can sound convincing while being incomplete, inaccurate or inappropriate. The FAQ tells users to double-check facts and says reviewing citations is a good place to start.
The Copilot transparency note says the same about reliability. Even when Copilot draws on high-authority web data, it may misrepresent that content. Grounding in web results is meant to reduce the risk of ungrounded answers. It does not eliminate it.
This story shows why that distinction matters. A citation tells you where an answer claims to come from. It doesn't tell you the source supports the answer. The Google example shows the problem: the links were real, but NewsGuard says they never mentioned the allegation.
What readers can do
- Open the citations. For any claim about a person's conduct, check that the linked page actually says it. Don't rely on the chatbot's summary.
- Treat brand-new accounts as a warning sign. NewsGuard and Gnida Project both point to a recently created account as part of the pattern.
- Look for corroboration. A single anonymous video with no outlet, court record or named reporter behind it is weak evidence. A chatbot repeating it doesn't make it stronger.
- Report bad answers. Microsoft's documentation says to use "Give feedback" in Settings or the flag icon on the Copilot web page. In the Copilot mobile app, you can long-press a response and select Report.
- Be careful in work settings. Microsoft says its consumer Copilot privacy FAQ doesn't apply to Microsoft 365 Copilot signed in with Entra ID. The reliability caveat in the transparency note is general, so IT teams should still tell staff not to forward unverified AI answers about people.
The bigger picture
Chatbots with live web access can be fed false claims by anyone who can publish quickly. A new account, a convincing video and some coordinated amplification may be enough. The model may then summarize the claim and cite loosely related pages that add false credibility. NewsGuard's finding is a narrow snapshot of a few prompts at one point in time. It is not a measured error rate. It does show that these tools can fail on a fast-moving, high-stakes claim. Microsoft's documentation says as much.
Political deepfakes in this race aren't limited to foreign operations. The Seoul Economic Daily reports that Ken Paxton's campaign circulated a fake ad showing Talarico, and Public Citizen separately criticized a Senate GOP-shared AI video of him. Those are different incidents from the abuse hoax and have no established link to it. They do show how crowded the synthetic-media environment is for readers and chatbots alike.
NewsGuard says Talarico's campaign didn't respond to its emails and phone message by publication. Microsoft hasn't said whether it has changed Copilot's answer. Until it does, treat any chatbot summary of this story with the same suspicion you'd give the video.