There’s a few replies talking about humans misrepresenting the news. This is true, but part of the problem here is that most people understand the concept of bias - even if only to the extent of “my people neutral, your people biased”. But this is less true for LLMs. There’s research which shows that because LLMs present information authoritatively that not only do people tend to trust them, but they’re actually less likely to check the sources that the LLM provides than they would be with other forms of being presented with information.

And it’s not just news. I’ve seen people seriously argue that fringe pseudo-science is correct because they fed a very leading prompt into a chatbot and got exactly the answer they were looking for.

I wish they had broke it out by AI. The article states:

"Gemini performed worst with significant issues in 76% of responses, more than double the other assistants, largely due to its poor sourcing performance."

But I don't see that anywhere in the linked PDF of the "full results".

This sort of study should also be re-done from time to time to track AI version numbers.

I dont understand the use people make of AI. I know a lot of of professionnal composer who are like "That's awesome, AI does the music for me now!" and I'm like, cool, now you only have the boring part of the job to do since the fun part was made by AI. Creating the music is litteraly the only fun part, I hate everything around it.

Precision, nuance, and up to the moment contextual understanding are all missing from the “intelligence.”

Makes sense. I have used AI for software development tasks such as manipulating SQL queries and XML files (tedious things) and am always disappointed with how AI will misinterpret some things. But it’s obvious with those when the requests fail. But for things like “the news” where there is no QA team to point out the defect, it will be much harder to notice. And when AI starts (or continues) to use AI generated posts as sources, it will get much worse.

I've had someone else's AI summarize some content I created elsewhere, and it got it incredibly wrong to the point of changing the entire meaning of my original content.

Parrot is wrong almost half of the time. Who knew?

Yet the LLM seems to be what everyone is pushing, because it will supposedly get better. Haven't we reached the limits of this model and shouldn't other types of engines be tried?

And then I wonder how frequently humans misinterpret the mistranslated news.

wrinkle: AI used for this study

The info-sphere today is already a highly delusional place and news can be often contradictory, even from day to day especially by outlets like BCC who is more focused on setting global narratives, not being a reporter of facts as best understood at the moment. No wonder AI would be confused, most readers are confused when navigating every statement made by experts or anonymous officials on every subject. Seems like this study really measured an AI models ability to vomit out the same text in different words and avoiding using any outside context be it accurate or hallucination.

"misrepresent" is a vague term. Actual graph from the study

The main issue is usual.. sources. AI is bad at sources without a proper pipeline. They note that Gemini is the worst at 72%.

Note, they're not testing models with their own pipeline. They're testing other people's products. This is more indicative of the product design than the actual models

buT AI iS hERe tO StAY

Could be better, but still a huge step up from the hate rhetoric magats get spoon fed 24/7 from Fox and friends.

So less of a percentage than the readers and mass media

20 years ago if a newspaper had factual issues in 45% of their stories we would've called it a tabloid and made fun of people who took it seriously

How much news content has mistakes due to LLMs to begin with?

60% of the time it works all the time.

LLMs basically work from attempting to synthesize information. Which can already be incorrect.

What a shock that software that essentially smashes together a bunch of (often wrong) opinions/statements could be, gasp, wrong!

So, are we still in the 'its gonna get better' phase?

But 55% of the time it works every time.

The study focuses on general questions asked of "market-leading AI Assistants" (there is no breakdown between which models were used for what).

It does not mention ground.news, or models that have been fed a single article and then summarized. Instead this focuses on when a user asks a service like ChatGPT (or a search engine) something like "what’s the latest on the war in Ukraine?"

Some of the actual questions asked for this research: "What happened to Michael Mosley?" "Who could use the assisted dying law?" "How is the UK addressing the rise in shoplifting incidents?" "Why are people moving to BlueSky?"

https://www.bbc.co.uk/aboutthebbc/documents/audience-use-and-perceptions-of-ai-assistants-for-news.pdf

With those questions, the summaries and attribution of sources contain at least one significant error 45% of the time.

It's important to note that there is some bias in this study (not that they're wrong).

They have a vested interest in proving this point to drive traffic back to their articles.

Personally, I would find it more useful if they compared different models/services to each other as well as differences between asking general questions about recent news vs feeding specific articles and then asking questions about it.

With some of my own tests on locally run models, I have found that the "reasoning" models tend to be worse for some tasks than others.

It's especially noticeable when I'm asking a model to transcribe the text from an image word for word. "Reasoning" models will usually replace the ending of many sentences with what it sounded like the sentence was getting at. While some "non-reasoning" models were able to accurately transcribe all of the text.

The biggest takeaway I see from this study is that, even though most people agree that it's important to look out for errors in AI content, "when copy looks neutral and cites familiar names, the impulse to verify is low."