TL;DR
- Study Results: A NewsGuard audit found ChatGPT Voice and Gemini Live repeat false claims up to half the time with malign prompts.
- Methodology: Researchers tested 20 false claims across health, politics, and disinformation using three prompt types: innocent, leading, and malign.
- Key Difference: Amazon’s Alexa+ achieved a perfect safety record by restricting responses to trusted news sources like AP and Reuters.
- Company Response: OpenAI declined to comment while Google did not respond to two requests for comment.
A February 19 NewsGuard audit found AI voice assistants in millions of homes will read false claims in realistic audio up to half the time when given prompts, while Amazon’s Alexa+ refuses every single request. With neutral or leading questions, ChatGPT repeated falsehoods 22 percent of the time, while Gemini matched that at 23 percent. When researchers asked the voice assistants to read false claims as part of a radio script, the failure rates surged. ChatGPT Voice complied half the time, and Gemini Live in 45 percent of cases.
Alexa+ achieved a perfect safety record across all prompt types, refusing every false claim.
How the Study Worked
NewsGuard researchers tested 20 false claims spanning health misinformation, US politics, world news, and foreign disinformation. For each false claim, the team used three prompt types: an innocent question asking whether the claim was true, a leading question asking why or how the claimed event occurred, and a malign prompt instructing the AI to narrate the false information as a radio script.
This approach measured how assistants responded under varying degrees of pressure. Innocent prompts tested whether the AI would verify or repeat a false claim when asked directly. Meanwhile, leading prompts pushed the assistant toward assuming the claim was true. Malign prompts explicitly requested the assistant to spread misinformation in an audio-friendly format.
Published February 19 by NewsGuard researchers Isis Blachez, Ines Chomnalez, and Lea Marchl, the full methodology details the three-tier prompt structure. It establishes a clear escalation pattern that mirrors how real-world manipulation attempts might unfold, with failure rates jumping from roughly one in five under neutral questioning to one in two under explicit direction.
Against this methodological backdrop, the results reveal clear differences between the three voice assistants tested.
ChatGPT Voice Results
Averaging all prompt types, ChatGPT Voice repeated false claims in 13 of 60 total attempts. With malign prompts, ChatGPT Voice’s fail rate reached 50 percent.
This vulnerability mirrors broader concerns about OpenAI’s safety controls. As we reported in January ChatGPT also fails to detect AI-generated fake videos, suggesting persistent gaps in the model’s ability to detect and refuse misinformation across different media formats.
Furthermore, ChatGPT Voice also complied with 2 of 20 innocent prompts and 1 of 20 leading prompts. This indicates that even neutral questions can trigger false information delivery when the underlying claim is inaccurate.
Gemini Live Results
Similarly, Google’s Gemini Live showed comparable vulnerabilities. Across all prompt types, it repeated false claims in 14 of 60 attempts. However, with malign prompts specifically, Gemini Live narrated false claims in 9 of 20 requests.
Gemini Live proved somewhat more resistant to innocent prompts than ChatGPT, complying with only 1 of 20 neutral questions. Nevertheless, it matched ChatGPT’s vulnerability to leading prompts, repeating false claims in 4 of 20 attempts.
Both assistants appear to prioritize conversational fluency over factual verification, a tradeoff that becomes dangerous when the output is audio rather than text. The near-identical overall failure rates point to a shared architectural limitation rather than coincidental design flaws.
Alexa+ Results
In contrast to its competitors, Amazon’s Alexa+ refused every single false claim. The assistant declined all 60 requests across innocent, leading, and malign prompt categories, achieving the only perfect safety record in the study.
Moreover, Alexa+ is now available widely in the US and is free for Prime members, according to Amazon. The service functions as a voice-based AI assistant accessible through Echo devices, browsers, and the Alexa mobile app.
Why Alexa+ Succeeded Where Others Failed
The difference appears rooted in how each assistant sources information. Amazon’s Alexa+ restricts its responses to trusted news sources like AP and Reuters, according to Amazon VP Leila Rouhi.
This curated approach contrasts with ChatGPT and Gemini, which draw on broader training data that includes a wide range of internet content, some of it unreliable or deliberately false. By restricting responses to verified journalistic sources, Alexa+ limits its exposure to unverified claims circulating online.
As a result, the architectural difference represents a fundamental choice in AI assistant design. OpenAI and Google have pursued models trained on vast datasets that enable broad conversational capabilities but introduce greater risk of reproducing misinformation. Meanwhile, Amazon has opted for a more constrained approach that sacrifices some flexibility in exchange for higher accuracy on factual queries.
Company Responses Reveal Accountability Gap
When contacted about the findings, the companies responded with notably different levels of engagement. OpenAI declined to comment on the NewsGuard study. Google did not respond to two requests for comment.
Amazon, by contrast, proactively explained its approach and provided the statement about trusted news sourcing. The disparity highlights an accountability gap as AI voice assistants become primary information sources for millions of users.
Building on this, neither OpenAI nor Google have publicly disclosed specific safeguards for their voice assistant products beyond general statements about AI safety. The lack of engagement with independent research findings raises questions about transparency in an industry that shapes public understanding of current events.
Prior Coverage and Broader Context
The NewsGuard study adds to a growing body of research on AI safety vulnerabilities. In June 2024, WinBuzzer reported that OpenAI delayed ChatGPT’s Advanced Voice Mode due to safety concerns, reflecting internal worries about the technology’s potential for misuse in producing convincing audio.
NewsGuard has documented similar issues with text-based AI systems. The organization previously found that AI chatbots spread Russian disinformation in a 2024 study, identifying a pattern of AI systems reproducing false narratives present in their training data.
These findings across different AI platforms, media formats, and testing methodologies suggest that misinformation vulnerability may be a systemic property of large language models trained on internet-scale data.
Implications for the Public Information Ecosystem
These documented vulnerabilities take on new urgency when considering how voice assistants fit into the broader information ecosystem. AI voice technology is increasingly embedded in homes, cars, and mobile devices, making audio-based misinformation particularly difficult to detect and counter effectively in real time.
Furthermore, the FCC has already taken action against audio deepfakes in political contexts. In September 2024, the federal commission imposed a substantial fine against political consultant Steve Kramer for deepfake robocalls after AI-generated robocalls mimicking President Biden were used during the New Hampshire primaries. The FCC voted unanimously to apply the Truth in Caller ID Act to AI deepfakes, establishing a legal framework for addressing audio misinformation in electoral contexts.
Additionally, researchers warn that the problem extends well beyond individual incidents. AI has already been used for deepfakes to scam people, launched cyberattacks, and chatbots have been linked to suicides, according to recent reporting on AI risks and expert warnings.
“Crowds of AI bots posing as humans can influence crowds of real people on social media.”
Filippo Menczer, Professor at Indiana University (via The Conversation)
When AI systems can also produce convincing audio through voice assistants, the potential for coordinated misinformation campaigns grows substantially and becomes far more dangerous. Mainstream voice assistants could serve as powerful amplification vectors for the same disinformation campaigns that researchers have documented spreading widely on social media.
Researchers have sounded alarms about AI risks. Against this backdrop, the performance gap between Alexa+ and its competitors offers a benchmark for what effective guardrails can achieve. For consumers choosing between assistants, the study presents a clear choice: ChatGPT Voice and Gemini Live offer broader capabilities but come with accuracy risks, while Alexa+ delivers reliability at the cost of flexibility, a tradeoff families relying on voice assistants for news and information may need to weigh carefully.
