As large language models (LLMs), a new generation of AI chatbots, and other forms of generative AI (artificial intelligence) become more widely used as information sources, it’s no surprise that many people are taking advantage of the convenience of AI search engine overviews. According to a June 2026 Pew Research Center report, roughly 60 percent of Americans read AI search engine summaries, 30 percent reported that they don’t, and 10 percent said they “were not sure if [they’d] done so.” But what is more alarming is that many users may not realize that these prominent, commanding summaries are generated by AI and can be incomplete or biased. Moreover, they may not know that they can opt out.
AI summaries are designed to sound polished and authoritative, which obscures the inherent inconstancy behind them. After all, these responses are not produced through independent verification of facts and accounts, which becomes even more concerning when a user is searching for information about current news events.
AI overviews mediate between users and the information they’re actually searching for. Just as a picture window or a camera viewfinder highlights some elements of a scene, while leaving others outside the frame, AI overviews direct our attention in certain ways—but with the added twist that, in this case, the framing is largely invisible to the user. One challenge that careful users of AI summaries must address is how to recognize the limits of what they’re reading.
Project Censored’s Beyond Fact-Checking guide originally examined the power of journalistic framing to shape our understanding of news events, through sourcing decisions, language choices, and contextual omissions. But now, a growing number of Americans use chatbot assistance or other generative AI tools to get their news, which presents some of the same challenges mentioned above, but without any equivalent transparency about how the information is generated and by whom.
Who’s responsible for erroneous summaries?
In June 2026, the Munich Regional Court issued a preliminary ruling that Google is “liable for a series of false statements generated by its AI Overviews feature.” The German court’s decision, Wired reported, could “make this case a historical precedent.” The court’s analysis revealed that instead of only displaying relevant links based on users’ prompts, Google’s AI summary tool generated “‘independent, new, and substantial statements’ based on a misinterpretation of information available on the internet.”
The German court’s conclusion is especially troubling when coupled with the findings of two recent studies. In January 2026, the Center for News, Technology and Innovation (CNTI) reported that many respondents had concerns about political bias in traditional news sources but viewed AI chatbots as “neutral.” This trust in AI was not shaken by either the bots’ “chronic factual errors” or their occasional inability to “provide up-to-date information.” Although this study’s data size was relatively small, the patterns emerging among “habitual” AI users are clear and threaten to permanently destabilize journalism, not only economically but as a cornerstone of democracy. Second, based on an investigation of news reports from 22 public service media outlets in 18 countries, the BBC and the European Broadcasting Union reported in October 2025 that the most popular AI assistants—including ChatGPT, Copilot, and Gemini—misrepresented news content 45 percent of the time.
How input determines output
AI chatbots rely heavily on well-formed questions or prompts, but users often forget important details, conflate events, or “take a contested premise as a given,” creating imperfect or biased responses. This means that our own framing, our limited understanding of a news event, could produce a narrow and misleading result.
Many users report feeling a sense of control when using chatbots, but a February 2026 MIT study found that chatbots provided less accurate, less truthful responses to users with lower English proficiency, less formal education, or non-US origins. LLMs “may actually exacerbate existing inequities by systematically providing misinformation or refusing to answer queries to certain users,” the study’s lead author, Elinor Poole-Dayan, warned.
In November 2025, Project Censored covered how hallucinations are not hiccups in the programming of AI chatbots; they are built into these systems. According to Science, LLMs “bluff” because “their performance is ranked using standardized benchmarks that reward confident guesses and penalize honest uncertainty.” Basically, chatbots are know-it-alls who don’t know it all.
Terms like “hallucination” and “glitch” effectively frame these occurrences as ephemeral bugs that are an inevitable, but inconsequential, side effect of greater technological advancement. Corporate media often prefer to hype progress, rather than challenge these assumptions. For instance, the Wall Street Journal recently framed AI “hallucinations” as an insignificant tradeoff for detailed and neat research, and that a user’s best defense against AI errors is “more AI” for “a round of fact-checking.” And in 2024, the Washington Post Creative Group launched a Generative AI Explorer’s Guide, paid for by Amazon Web Services.
AI-generated text is even invading the pages of legacy newspapers. In spring 2026, The Atlantic alleged that the New York Times published AI-generated text in its opinion pages. And after the Chicago Sun-Times laid off 20 percent of its staff, the paper published an AI-generated summer reading list. Unfortunately, several of the recommended books didn’t exist.
Dueling frames
Generative AI tools might seem convenient for accessing digestible news, but they’re hindering users’ curiosity and driving them away from independent news outlets and toward narratives generated by biased technology systems. A Stanford University study that audited six commercial chatbots for their ability to answer news questions concluded, “As more users encounter journalism through AI lenses rather than directly through publishers’ sites, differences in [context], attribution, and source selection will increasingly shape whose reporting reaches the general public, under what terms, and how.”
So now we see dueling framing issues: generative AI and chatbots constructing flawed responses based on how users phrase their questions and the assumptions embedded in their search terms, while AI itself is framed in public discourse as either fallible technology prone to error or, more frequently, a trustworthy tool capable of delivering reliable information.
What can we do about AI gatekeeping?
Most obviously, avoid search engine AI overviews. For news, seek reporting directly from reputable independent news outlets.
Use your web browser’s settings to disable or hide its automated overviews. If that’s difficult in your current browser, consider switching to one that supports that preference, such as Brave.
These are good, but limited first steps, since they keep you lodged in a consumer role. More significant change is likely when we coordinate our actions with others. Journalists can make sure to avoid reporting that reinforces common mindsets many people unconsciously employ when thinking about AI—including equating AI systems with progress, treating them as “objective” tools that work independently of human influence, and focusing on them primarily as consumer products whose main purpose is to improve users’ quality of life. Instead, journalism that aims to promote algorithmic literacy and the public good should give at least equal attention to how AI systems often reflect and reproduce existing biases, thus amplifying social and economic harms, especially to already marginalized communities.
Members of those communities and their allies can join forces with independent news outlets and digital rights organizations, such as the Electronic Frontier Foundation, to ensure that new technologies promote freedom, justice, and innovation for everyone. A March 2026 report from FrameWorks found that framing new AI tech in terms of social justice, using specific real-life examples, helped people break free from the individualistic, consumer mindset and appreciate how future AI systems could be developed to promote inclusive public goods rather than exclusive private benefits.
Super PACs (political action committees) focused on promoting AI, such as Leading the Future, are already raising and spending millions of dollars to influence politicians in the upcoming midterm elections. The message to politicians is clear, The Guardian reported, “Regulate AI, and we will find you, wherever you are.”
We probably cannot hope to outspend the AI Super PACs, but we may be able to challenge the predominance of AI boosterism by working to reframe the public’s understanding of what AI systems, such as Google’s Overview, can—and can’t—actually do.
When we take time to frame-check AI-generated overviews and chatbots’ interpretation of news events, two things happen. First, we realize that those summaries are not as authoritative or trustworthy as they appear at first; and, second, we remind ourselves of our own capacities for critical, independent thinking.
The future of AI is not as inevitable as Big Tech’s cheerleaders would have us believe. Instead, by looking critically at what’s left out of frame—including the reproduction of systemic inequalities and the massive amounts of money being directed toward promoting new AI-powered systems—we can chart a different course, one that centers social justice and public good in every discussion of AI.
- First published at Project Censored.