Screenshot of the EBU/BBC News Report being analyzed in Parsd.
Why Professional Thinkers Need Different Tools Than News Consumers
On October 21st, 2025, the EBU and BBC released a study called “News Integrity in AI Assistants – An international PSM study” that should concern anyone who cares about the integrity of information in our society.
The headline finding: 45% of responses contained at least one significant issue, and 81% had some form of problem. As someone building AI tools for professional analysts, this study validated many of my concerns while also clarifying important distinctions we need to make.
The Scale of the Problem
The study analyzed four widely used AI assistants: ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity. The problems they found were systematic and troubling:
Sourcing was the biggest issue – claims not supported by cited sources, sometimes a lack of direct sources, fabricated links, and the use of irrelevant sources like satirical content and opinion content from social media.
Accuracy problems were pervasive – basic factual errors, outdated information, inaccurately represented information, and fabricated or altered direct quotes.
Opinion presented as fact – AI responses often presented opinions as facts and incorrectly attributed opinions to individuals or organizations, including public service media.
“These errors are systemic across languages, assistants, and organizations. This isn’t a bug in one tool – it’s a fundamental challenge.”
This is troublesome from a democratic perspective since both a well-functioning democracy and a working market economy depend on the integrity of facts. If more people are exposed to these many factual issues in AI tools, we have a societal problem.
Understanding the Real Differences Between AI Tools
What struck me when reading deeper into the study was that it focused on free consumer versions of these AI assistants. This is both troubling and important to understand.
It’s troubling because 15% of people under 25 use these tools as their main source of news. They’re getting information that could be wrong or misleading nearly half the time.
However, it’s also important to remember that AI tools work differently depending on the models they use and the features of the AI assistant. Assistants with proper web search, research capabilities, and especially those where users bring their own curated data operate fundamentally differently.
“The challenge is to appreciate the obvious risks while being aware that risks differ depending on which type of AI tool people are using – this could lead to confusing debates if we’re not careful.”
Source Management: More Complex Than It Seems
The study’s finding that sourcing is the biggest problem validates the work we’ve been doing at Parsd on source management borrowed from both academia and intelligence analysis.
But it’s not just about having sources – it’s about having appropriate sources with proper context. The study highlighted issues like:
- Using satirical content as factual sources
- Relying on social media platforms without proper verification
- Failing to distinguish between news reporting and editorial opinion
- Attributing editorial opinions to entire news organizations
“We need to distinguish the content itself from the publisher that produced it. Content can be checked for logical conclusions and contradictions, but the publisher’s trustworthiness is a separate assessment.”
This also reveals what I call the “trust bubble” problem – established media organizations trusting only other established media. While it’s good that media adhering to journalistic standards gain more trust, this can exclude quality research articles, reports from NGOs, and reports from independent public agencies. How can we ensure that high-quality sources across society can be used to develop fact-based insights?
The Public Data Problem
There’s an irony here that concerns me deeply: a lot of high-quality content is locked behind paywalls or company firewalls.
I understand the need to protect intellectual property and content value. But this means we’re omitting important high-quality data from AI systems at a time when fact-based insights have never been more important.
“Allowing users to curate datasets of quality data they have access to is crucial to preserve the integrity of facts necessary for a working democratic society.”
The AI Literacy Divide
When I meet with people today, I see a growing divide in AI literacy. Some have limited experience and understanding of AI tools. Others have already developed custom configurations of paid AI tools with sophisticated capabilities.
This divide, combined with major differences between free and paid versions of AI assistants, creates a two-tiered system. It’s similar to what we saw with earlier innovations like BI dashboards and machine learning – quality data becomes the differentiator.
“Just the fact that we’re reflecting on which data we use and where it came from is important. This is the difference between having a systematic method versus just copying answers from an AI Assistant without reflection.”
Consumer Tools vs. Professional Tools
The EBU/BBC study makes recommendations for both AI developers and publishers. These are important for protecting citizens who consume news through AI assistants.
But there’s a different perspective for professional analysts creating insights for decision-making in organizations. These professional tools can and should be more sophisticated because:
- The stakes are different – professional reputation and organizational decisions
- The use case is insight creation, not just consumption
- Users need more control and transparency, not just simplicity
“We need to power this underserved profession with the tools and methods they deserve. Getting source management, attribution, and digital provenance right is an important first step.”
The Language Challenge
The study found notable differences based on language, with English responses more likely to include direct sources. As someone building tools in Sweden, this concerns me.
Swedish is fortunate to be a “big enough” language to be included in many models and products, but definitely not all. Smaller language markets face even greater vulnerability. Still, this is based on the usage of free tools, and it is not the same when using AI on user-provided data.
“It’s crucial that high-quality content in smaller languages is available for insight producers to use as source data when doing domain-specific analysis.”
What This Means Going Forward
The high level of issues in the EBU/BBC report highlights that improved AI literacy is essential. It matters what tool you choose and how you choose to use it.
While these results are worrisome on a societal level, many of these issues can be mitigated by using more suitable tools where hallucinations and incorrect sources are less problematic.
“This highlights why the work we and other responsible AI providers are doing is important – providing AI tools that augment the human mind rather than trying to replace humans with fully automated tools for critical tasks.”
Looking Ahead
I hope professional thinkers will continue exploring how technology and methodology can develop side by side. We need awareness of AI’s incredible possibilities while understanding that all tools are not created equally and all have limitations.
Finding what parts can be accelerated with AI and where we need humans for verification will be crucial. We need to increase not just productivity but also quality and relevance of insights.
We’re still in the early phase of this development. I encourage more dialogue on these challenges so we can make the world a little bit better together.
Because when 45% wrong isn’t good enough – and it clearly isn’t for professional analysis – we need to build something better.
What’s your experience with AI tools for research and analysis? How do you balance the productivity benefits with the quality concerns?
Alexandra Kafka Larsson is the CEO and Co-founder of Parsd, a digital research platform that helps analysts create trusted insights. She previously served as a military intelligence officer in the Swedish Air Force and has over 30 years of experience in intelligence systems and methods.
