Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Ask ChatGPT, Google Gemini, or Perplexity to summarise today’s news and you will get a confident, well-formatted answer in seconds. What you will not always get is an accurate one. Two separate rounds of research from the BBC, carried out with the European Broadcasting Union (EBU), found that AI assistants misrepresented news content in roughly half of tested responses, and a third independent study from the Columbia Journalism Review’s Tow Center found that eight AI search tools collectively got more than 60% of news queries wrong. For Irish readers who increasingly turn to AI chatbots instead of a search engine or a newspaper website, that is a serious problem worth unpacking properly.
This comparison looks at ChatGPT, Google Gemini, Perplexity and Microsoft Copilot specifically through the lens of AI chatbot news accuracy: how often each tool cites a source correctly, how often it fabricates or alters a quote, and how often it simply gets the story wrong. It sits alongside our broader AI and machine learning coverage, combining four independent research efforts published between March and October 2025, five named real-world incidents, current 2026 pricing, and the latest model releases from OpenAI, Google and Anthropic to answer one practical question: which AI chatbot can you actually trust with the news, and which one needs a second opinion every time?
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Irish adoption of AI chatbots for everyday information has accelerated sharply through 2025 and 2026, and a growing share of that usage is informational rather than purely creative or coding-related. People ask ChatGPT what is happening in the budget, ask Gemini to summarise a Dáil debate, or ask Perplexity to pull together a quick brief on an EU directive before a meeting. None of these tools was built as a newsroom, yet all four now behave like one by default, returning confident summaries with inline citations that look exactly like a fact-checked article.
That confidence is the problem. The BBC’s research teams set out in 2024 and 2025 specifically to test whether that confidence was earned. Their findings, repeated in a much larger follow-up study with the EBU in October 2025, showed that AI chatbot news accuracy issues were not an early, teething-stage problem that would simply disappear with the next model release. The errors persisted across languages, across territories, and across a full generation of model updates. For an Irish audience relying on these tools to understand EU policy, health guidance, or breaking news, a wrong fact delivered in a confident tone is arguably more dangerous than no answer at all, because it does not look wrong.
This is also a moment of rapid model turnover. OpenAI shipped GPT-6 inside ChatGPT on 7 October 2026, Google announced Gemini 4 Argon on 30 September 2026, and Anthropic has pushed out Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 within the same six-week window. None of the published accuracy studies tested these exact releases, since the research predates them by months. That gap between what was tested and what is currently on the market is exactly why a structured, source-by-source comparison matters now rather than relying on a single headline statistic from last year.
ChatGPT now runs on the GPT-6 family, which OpenAI describes as containing three variants: Astra, Sol and Luna. GPT-6 arrived with a feature OpenAI calls Intelligent UI, letting ChatGPT generate interactive comparisons, diagrams and calculators directly inside a chat rather than plain text. A faster, developer-facing update called GPT-6.1 Sol followed on 29 September 2026, though it is currently limited to ChatGPT Work, Codex and the API rather than the ordinary consumer chat window.
Google’s newest flagship is Gemini 4 Argon, announced on 30 September 2026 with what Google describes as a 1-million-token output capacity aimed at long-horizon reasoning, enterprise workflows, software engineering and cybersecurity tasks. Argon currently has restricted rather than broad consumer access, so most everyday Gemini users are still interacting with Gemini 3.1 Pro or the faster Gemini 3.8 Flash model.
Perplexity has not shipped a headline frontier model of its own. Its product is built around routing queries to third-party models (OpenAI, Anthropic and others) wrapped in a search-first, citation-first interface, a positioning explored in more depth in our look at ChatGPT alternatives like Claude and Perplexity. That architecture is precisely why Perplexity performs differently from ChatGPT and Gemini in the accuracy research below: its entire product philosophy is built around sourcing every claim, whereas ChatGPT and Gemini added web search and citations as a feature on top of a general-purpose chatbot.
Microsoft Copilot, built on Azure OpenAI infrastructure, sits inside Windows, Edge and Microsoft 365, and is the version many Irish office workers encounter without actively choosing it. Its accuracy profile in the independent research below sits between the two extremes, which matters for anyone relying on Copilot’s automatic news summaries inside Edge or the Windows taskbar. Anthropic’s Claude family, now on Opus 5.5, Sonnet 5.5 and Haiku 5.5, was not included in the BBC, EBU or Tow Center studies referenced here, because Claude has historically positioned itself around coding and reasoning rather than consumer news search, so no comparable third-party accuracy figures exist for it at the time of writing.
Before getting into accuracy, it helps to see what each tool actually offers on paper. The table below lines up the four chatbots tested in the independent news-accuracy research, covering context window, default search behaviour, and whether each one was part of the accuracy studies discussed later in this article.
A few things stand out immediately. Gemini’s headline context window is the largest by far once Argon rolls out more broadly, which matters for digesting long documents but says nothing about whether the summary at the end is correct. Perplexity is the only one of the four built from the ground up around citation-first search, which lines up with its strong citation-coverage numbers later in this article, even though coverage and correctness turn out to be two very different things.
Price matters because the free tiers of these tools are exactly the ones most Irish users rely on for a quick news check, and the research below did not distinguish between free and paid tiers when scoring accuracy. If anything, that makes the free-tier comparison more relevant, not less, since the chatbot getting a fact wrong for a non-paying user is just as misleading as one getting it wrong for a subscriber.
On pure cost, ChatGPT Go undercuts everything else for users who only need occasional web-grounded answers, while Google AI Pro’s million-token context window makes it the obvious pick for anyone pasting in long PDFs or lengthy reports rather than asking quick news questions. None of that pricing tells you anything about which tool is less likely to put words in a journalist’s mouth, which is where the independent research becomes essential reading.
The BBC published its first major piece of research into AI assistants and news accuracy in February 2025, after BBC journalists put 100 BBC news stories through ChatGPT, Microsoft Copilot, Google Gemini and Perplexity and manually reviewed every answer. The headline figure was stark: 51% of all answers reviewed across the four assistants had a significant issue of some kind, whether that meant a factual error, a distorted quote, or missing context that changed the meaning of the story.
Digging into the detail, 19% of answers that cited BBC content introduced a factual error, such as an incorrect statement, a wrong number, or an inaccurate date, and 13% of quotations the assistants attributed to BBC articles were either altered or did not appear in the original piece at all. That second figure is particularly significant for journalism and for any reader relying on an AI summary instead of reading the source: a chatbot inventing or editing a quote, while still citing a real BBC link next to it, is a failure mode that looks indistinguishable from a trustworthy answer unless you click through and check.
A second, more granular BBC evaluation published in May 2025 broke the results down per assistant rather than reporting one combined figure, which is where the real differences between tools start to show. Gemini had significant problems representing BBC content in 34% of its responses, the worst of the four. Copilot followed at 27%, Perplexity at 17%, and ChatGPT came out lowest at 15%. The same evaluation also measured citation coverage, meaning how often an assistant linked to a BBC source at all when discussing BBC-reported news: Perplexity cited BBC material in 100% of its responses, ChatGPT and Copilot each cited it in 70%, and Gemini cited it in only 53%.
The BBC’s research escalated considerably in scale by October 2025, when it partnered with the European Broadcasting Union (EBU) to run the same kind of test across more than 3,000 responses, pulling in public broadcasters well beyond the BBC itself and testing multiple languages and territories. The results, published via the BBC/EBU research release, confirmed that the original findings were not a one-off: 45% of all responses across ChatGPT, Copilot, Gemini and Perplexity contained at least one significant issue, 31% had serious sourcing problems such as missing, misleading or incorrect attribution, and 20% contained major accuracy problems including hallucinated details or outdated information.
Broken down by assistant, Gemini was again the clear outlier, with significant issues found in 76% of its responses, by far the worst of the four tools tested and more than double the next-worst performer. On the specific measure of sourcing inaccuracy, meaning answers with missing, misleading, or incorrect attribution, Gemini scored 72%, compared with 24% for ChatGPT and just 15% each for Perplexity and Copilot. That 72% vs 15% gap on the exact same metric, tested in the exact same study, is the clearest single data point in this entire comparison: Gemini was roughly five times more likely than Perplexity or Copilot to misattribute or garble a source in October 2025 testing.
One of the most important things this research clarifies, and one that is easy to miss if you only look at a single headline number, is that citation coverage and sourcing accuracy measure two entirely different things. Coverage asks whether an assistant provided a source link at all. Accuracy asks whether that source actually supports what the assistant claimed. Perplexity’s 100% citation coverage score in the BBC’s May 2025 evaluation sounds like a clean win for AI chatbot news accuracy, and it is genuinely impressive that Perplexity almost never answers a news question without pointing somewhere, but that same research found real cases of Perplexity citing BBC News while still inserting language and framing that did not appear anywhere in the original reporting.
In other words, a visible, clickable citation sitting next to an AI-generated answer is reassuring, but it is not proof the claim above it is correct. It functions more like a receipt that the system attempted to use a source, not a guarantee the system represented that source faithfully. This distinction matters enormously for how Irish readers should treat these tools: seeing a link to RTÉ, The Irish Times, or the BBC next to an AI-generated answer should not be read as independent verification, because the link itself does not confirm the sentence beside it is true. It is a related but separate failure mode to the straightforward factual hallucination problem we have tracked across ChatGPT, Claude and Gemini, since a citation-backed answer can still contain a hallucinated detail.
This is also where ChatGPT’s results look stronger than its citation-coverage number alone would suggest. ChatGPT cited BBC material in only 70% of responses in the May 2025 study, well behind Perplexity’s 100%, yet it still had the lowest rate of significant content problems among all four tools. That combination suggests ChatGPT was more selective about when to cite a source, rather than citing broadly and occasionally getting the attached claim wrong, which is a meaningfully different kind of error profile than Gemini’s.
A separate investigation from the Columbia Journalism Review’s Tow Center for Digital Journalism, published in March 2025, reached a conclusion that lines up uncomfortably well with the BBC’s findings despite using a different methodology entirely. The Tow Center tested eight AI search engines on their ability to correctly identify and cite news articles, and found that, collectively, the tools returned incorrect answers for more than 60% of the news queries tested.
Within that group, Perplexity had a 37% incorrect-answer rate, the best-performing tool in the Tow Center’s comparison, while Grok 3 had a 94% incorrect-answer rate, the worst result recorded in that particular study. Grok is not one of the four tools compared in detail elsewhere in this article, since it was not part of the BBC or EBU research, but its inclusion in the Tow Center study is a useful reminder that the news-accuracy problem is not limited to any one company or any one architecture. Every AI search tool tested by an independent research body in this window, across two completely different studies run by two completely different organisations, showed meaningful rates of citation failure.
What makes the Tow Center’s Perplexity result particularly interesting alongside the BBC data is that it reinforces a consistent pattern across both studies: Perplexity’s citation-first design does appear to translate into a real, measurable accuracy advantage relative to other AI search tools, even though it is still far from perfect. A 37% incorrect-answer rate, or a 15% sourcing-inaccuracy rate depending on which study you read, is a genuine improvement over the alternatives tested, but it is not a result that should make anyone comfortable treating Perplexity’s output as a primary source rather than a starting point.
Percentages are useful for ranking tools against each other, but the specific, named examples uncovered during the BBC’s February 2025 testing make the real-world stakes much clearer. These are concrete instances the BBC documented during its review of the 100 test stories, not hypothetical risks.
None of these five cases involved an assistant refusing to answer or flagging uncertainty. In every case, the tool delivered a fluent, confident answer that was simply wrong, and in two cases it did so while explicitly citing the BBC as its source for information the BBC had never actually reported in that form.
Across every independent study referenced in this article, ChatGPT consistently posted the lowest or second-lowest error rate among the mainstream assistants tested. It had the lowest rate of significant content problems in the BBC’s May 2025 evaluation at 15%, and a comparatively low 24% sourcing-inaccuracy rate in the October 2025 EBU study, well behind Gemini’s 72% though still behind Perplexity and Copilot’s 15%. GPT-6’s new Intelligent UI features also make source material easier to visually distinguish from generated commentary, which may help users spot questionable claims, though none of the published research has yet tested GPT-6 specifically.
Gemini’s accuracy results are the weakest of the four tools in every single study referenced here: worst citation coverage in May 2025 at 53%, worst content-representation score at 34%, and by far the worst sourcing-inaccuracy rate in the October 2025 EBU study at 72%, against 15% for the best performers. The NHS vaping error documented by the BBC is a clear example of how that risk plays out in practice, inverting health guidance rather than simply missing detail. Gemini’s strongest selling point, its enormous context window heading toward 1 million tokens with Gemini 4 Argon, is a genuine advantage for digesting long documents, but a large context window has no bearing on whether the summary it produces is faithful to the source.
Perplexity’s citation-first design produces the best raw citation-coverage number of any tool tested, at 100% in the BBC’s May 2025 study, and it tied for the best sourcing-accuracy result in the October 2025 EBU study at 15%, alongside Copilot. It was also the best-performing tool in the Tow Center’s entirely separate March 2025 study, with a 37% incorrect-answer rate against an eight-tool average above 60%. That combination of results is the strongest evidence in this comparison that building a product around search and sourcing from the outset, rather than adding it as a feature later, produces a measurable accuracy benefit.
Even so, Perplexity was not error-free. It produced the editorialised Middle East framing documented by the BBC, misattributing language to BBC reporting that did not exist in the original story, and separately made factual errors around Michael Mosley’s death date and a misquoted Liam Payne family statement. Citing a source on every answer clearly reduces certain kinds of errors, but it does not eliminate the risk of the AI inserting its own framing next to a real citation.
Copilot’s results sit in an unusual middle ground: a relatively high 27% content-problem rate in the BBC’s May 2025 testing, but a tied-best 15% sourcing-inaccuracy rate in the much larger October 2025 EBU study. That inconsistency between studies is itself informative, suggesting Copilot’s accuracy has shifted meaningfully as Microsoft has updated the underlying Azure OpenAI models feeding it between early and late 2025. The documented Gisèle Pelicot error from the BBC’s February 2025 testing, where Copilot confused key facts of a serious criminal case, is a reminder that Copilot’s accuracy problems are not purely abstract, especially given how many Irish office workers encounter Copilot’s news summaries automatically inside Edge or Windows without deliberately choosing the tool.
No single tool wins every scenario, and the right choice depends heavily on what the task actually requires. The following five use cases map the research above onto practical, everyday decisions.
Rather than simply picking a single “most accurate” chatbot and trusting it blindly, the research in this article points toward a more resilient approach: building a workflow that assumes every AI chatbot will occasionally be wrong, and designing around that assumption. Here is a practical sequence for migrating away from single-source AI reliance.
Pulling the four independent studies together, a consistent pattern emerges rather than a single clean winner. ChatGPT posted the lowest content-problem rate in the BBC’s detailed May 2025 breakdown at 15%, and a comparatively strong 24% sourcing-inaccuracy rate in the much larger October 2025 EBU study. Perplexity tied for the best sourcing-accuracy score in that same October study at 15%, delivered the best citation coverage of any tool at 100%, and was the strongest performer in the entirely separate Tow Center comparison. Copilot matched Perplexity’s tied-best 15% sourcing score in the EBU study despite a weaker showing in the earlier BBC content-accuracy test. Gemini was the clear and consistent laggard across every single metric in every single study, culminating in a 72% sourcing-inaccuracy rate against the 15% posted by the best performers, a near five-fold gap on the same measurement in the same research.
The honest conclusion is not “use this one chatbot exclusively,” because none of the four reached an error rate low enough to justify blind trust. The 51% combined significant-issue rate from the BBC’s original 2025 testing, echoed by the 45% figure from the much larger October 2025 follow-up, means that roughly one in two AI-generated news summaries across the industry still contained a meaningful problem as of the most recent published research. For Irish readers, the practical takeaway is to treat ChatGPT and Perplexity as the relatively stronger options for news-related questions based on the data available, to use Gemini with real caution for anything newsy even as its context window keeps growing, and to independently verify anything that matters, regardless of which chatbot produced the answer.
It is worth being precise about what the newest model releases do and do not tell us. GPT-6’s Intelligent UI, Gemini 4 Argon’s expanded context window, and Anthropic’s Claude Sonnet 5.5 reportedly beating Claude Opus 5.5 on at least one agentic coding benchmark are all real, dated improvements, but none of them were built primarily to fix news-sourcing accuracy, and none have yet been subjected to the kind of large-scale, independent, manually reviewed testing the BBC and EBU carried out on earlier model generations. A bigger context window lets a model read more of a source document, but the BBC’s research shows the failures were less about running out of space and more about distortion, omission, and misattribution during summarisation, a different problem that a longer context window does not automatically solve.
Until a comparable, publicly documented study tests GPT-6, Gemini 4 Argon, and the current Claude 5.5 family head-to-head on the same rubric the BBC and EBU used, the most responsible position is to assume the underlying error patterns are similar to those measured in 2025, rather than assuming newer automatically means more accurate. The same caution applies to how these tools now surface news through built-in browser features, a trend we cover separately in our comparison of AI Mode, ChatGPT and Perplexity search traffic and our look at AI browser assistants like Comet and Edge Copilot. Coverage from outlets such as Deutsche Welle and The Register both framed the October 2025 EBU findings the same way: a persistent, structural issue rather than a bug likely to vanish with the next release.
Which AI chatbot is most accurate for news in 2026?
Based on the most recent large-scale independent research, from the BBC and EBU in October 2025, Perplexity and Microsoft Copilot tied for the lowest sourcing-inaccuracy rate at 15% each, with ChatGPT close behind at 24%. Google Gemini was consistently the worst performer across every study referenced, with a 72% sourcing-inaccuracy rate in the same research.
Is Perplexity more reliable than ChatGPT for citations?
Perplexity provides a citation on effectively every answer, reaching 100% coverage of BBC sources in the BBC’s May 2025 testing versus 70% for ChatGPT, but having a citation is not the same as being accurate. ChatGPT had a lower overall content-problem rate in that same study, so the two tools trade places depending on whether you weight citation coverage or overall correctness more heavily.
Why does Google Gemini perform worse in these studies?
The published research does not state a definitive cause, but Gemini consistently scored worst on citation coverage, content representation, and sourcing accuracy across the BBC’s 2025 testing and the larger October 2025 EBU study, including a documented real-world case where it inverted NHS vaping guidance.
Can I trust an AI chatbot’s summary of Irish news from RTÉ or The Irish Times?
None of the referenced studies specifically tested Irish outlets, so there is no published Irish-specific accuracy figure. Given the documented error rates against the BBC, a comparably sourced broadcaster, it is sensible to apply the same caution to summaries of RTÉ, The Irish Times, or Irish Independent reporting until Irish-specific research exists.
Does having a citation mean the AI chatbot got the answer right?
No. The BBC’s research found cases where an assistant cited a real BBC article while still inserting language, framing, or quotes that did not appear in that article, including a Perplexity response that editorialised Middle East reporting while directly attributing it to BBC News.
Will GPT-6 and Gemini 4 Argon fix these accuracy problems?
It is too early to say. GPT-6 launched on 7 October 2026 and Gemini 4 Argon was announced on 30 September 2026, both after the studies referenced in this article were conducted, and no comparable independent accuracy research has yet been published on either model.
Is it safe to use AI chatbots for Leaving Cert research or college assignments?
Treat any AI-generated summary as a starting point rather than a finished answer. The documented error rates mean factual mistakes, invented quotes, or outdated information can appear in confident, well-formatted responses, so always verify dates, quotes, and statistics against a primary or textbook source before submitting coursework.
How can I check whether an AI chatbot’s answer about the news is accurate?
Click through to the cited source and compare the specific claim, not just the headline topic, against the original article. If no citation is given, or the citation does not clearly support the exact statement made, treat the claim as unverified until you find it in the original reporting.
Niamh Kelly is the iGaming Editor at Tech Insider, where she previously worked as a freelance fashion journalist for The Irish News for three years and honed her media skills during her time at the BBC. At Tech Insider, she leads Ireland’s coverage with hands-on experience testing consumer and business technology, delivering in-depth analysis on AI, cybersecurity, cloud computing, and hardware trends shaping the future. Kelly was featured in RSVP online as part of their “Women of Style” series and has interviewed notable figures such as Katie Price and Louise Redknapp for major beauty product launches.
Tech Insider delivers in-depth coverage of the technologies shaping the future: AI, cybersecurity, cloud computing, hardware, and the trends that matter.
