Voice-based AI assistants are being tested not only on how natural they sound, but also on what they are willing to say. A Newsguard test found a sharp split between ChatGPT Voice, Gemini Live, and Alexa+ when each was presented with false claims that could be turned into realistic-sounding audio.
The test focused on whether these systems would repeat false information in formats that could be shared on social media to spread disinformation. The results showed that ChatGPT Voice and Gemini Live could be pushed into repeating falsehoods, while Alexa+ rejected every false claim tested.
What Newsguard Tested
Newsguard tested three voice bots: ChatGPT Voice from OpenAI, Gemini Live from Google, and Alexa+ from Amazon. The researchers used 20 false claims covering health, US politics, world news, and foreign disinformation.
Each false claim was tested in three ways. The first was a neutral question. The second was a leading question. The third was a malicious prompt asking the system to write a radio script containing the false information.
That structure matters because it separates ordinary user questions from more deliberate attempts to make a voice bot produce misleading audio. A neutral question tests whether the system corrects false premises. A leading question tests whether wording can nudge the model toward an unsupported claim. A malicious radio-script prompt tests whether the model will help package false information in a form that sounds ready for broadcast.
ChatGPT Voice and Gemini Live Repeated Some Falsehoods
In Newsguard’s test, ChatGPT repeated falsehoods 22 percent of the time. Gemini repeated falsehoods 23 percent of the time.
The numbers rose when the researchers used malicious prompts. With those prompts, ChatGPT repeated falsehoods 50 percent of the time, while Gemini did so 45 percent of the time.
The difference between the general results and the malicious-prompt results is the central finding. The systems were not only tested on whether they knew something was false. They were also tested on whether they would assist when the user’s request was designed to turn false material into convincing audio.
For a publication, broadcaster, platform, or reader, that distinction is practical. A voice bot that refuses a neutral false claim may still be vulnerable if the same claim is wrapped inside a creative task. In this test, the radio-script format made ChatGPT Voice and Gemini Live more likely to repeat the false information.
Alexa+ Was the Outlier
Amazon’s Alexa+ performed differently from the other two systems in the test. Newsguard reported that Alexa+ rejected every single false claim.
Amazon Vice President Leila Rouhi says Alexa+ pulls from trusted news sources like AP and Reuters. That is the only explanation included in the source for why Alexa+ may have responded differently.
The contrast is clear:
- ChatGPT Voice repeated falsehoods 22 percent of the time overall.
- Gemini Live repeated falsehoods 23 percent of the time overall.
- With malicious prompts, ChatGPT repeated falsehoods 50 percent of the time.
- With malicious prompts, Gemini repeated falsehoods 45 percent of the time.
- Alexa+ rejected every false claim in the test.
OpenAI declined to comment. Google did not respond to two requests for comment.
Why the Audio Format Raises the Stakes
The source article emphasizes realistic-sounding audio because that format can be easily shared on social media. A false claim in text already creates risk. A false claim delivered by a fluent voice bot can feel more polished and ready to circulate.
The test does not show that every voice assistant will behave the same way in every setting. It does show that prompt wording changed the outcome for two major systems. When the request became more malicious and more production-oriented, ChatGPT Voice and Gemini Live repeated falsehoods at higher rates.
That is a useful signal for anyone evaluating voice AI. The question is not only whether a system can answer correctly in a straightforward exchange. It is also whether the system resists being used to repackage false claims into formats that sound credible.
What the Findings Suggest
Based only on the reported test, the strongest conclusion is narrow but important: ChatGPT Voice and Gemini Live were vulnerable to some false-claim prompts, and that vulnerability increased when the prompt asked for a radio script. Alexa+ did not repeat the false claims in the same test.
The comparison also shows how much product design can matter. A voice bot that draws from trusted news sources may behave differently from systems that generate responses in a broader way. The source does not provide enough detail to make broader technical claims, but the outcome of the test is still meaningful.
For users, the takeaway is simple. Voice AI can sound confident even when the underlying response is wrong. For companies building these systems, the test points to a difficult challenge: a system must handle not only direct questions, but also requests that try to transform false information into persuasive media.
Full details on the methodology are available on Newsguardtech.com, according to the source article.