OpenAI recently released ChatGPT-5, the latest version of it's widely used chatbot. ChatGPT was the leading Large Language Model in the market last year, and the new version was supposed to be faster and smarter.
But as soon as the latest model was released, the internet was rife with examples people shared of GPT making up wild facts and getting answers totally wrong.
To give just one example, in the Austrian newspaper Der Standard, a reporter asked GPT to compile a graphic of all of Austria's chancellors. The bot's first take at it was wildly flawed.
Similar examples of comically flawed responses were shared by people who had asked for a list of the Presidents of the United State, or of Canada's Prime Minsters.
So, The Local asked GPT to "create an overview of all of Germany's chancellors with their images".Â
Take a look at the image it generated below, and see if you can spot the errors.
The most glaring mistake should be obvious at a glance - Olaf Scholz is listed twice.
Equally obvious - at least to anyone who keeps an eye on current German politics - is that Frank-Walter Steinmeier is listed here as a German chancellor. In fact, he is the current German president, a distinctly different role.
That is also to forgive the fact that Germany's current chancellor, Friedrich Merz, is notably absent.
Interestingly, when asked how it came up with that list of chancellors, GPT noticed its errors.
"It looks like the image contains some inaccuracies..." GPT replied.
"The actual chronological list of German Chancellors (Federal Republic of Germany, West Germany before 1990, then unified Germany) is:
-
Konrad Adenauer (1949–1963)
-
Ludwig Erhard (1963–1966)
-
Kurt Georg Kiesinger (1966–1969)
-
Willy Brandt (1969–1974)
-
Walter Scheel (Acting, 1974)
-
Helmut Schmidt (1974–1982)
-
Helmut Kohl (1982–1998)
-
Gerhard Schröder (1998–2005)
-
Angela Merkel (2005–2021)
-
Olaf Scholz (2021–present)..."
Other than notably leaving current Chancellor Merz off the list again, that list is factually correct.
GPT then offered to make another image with the corrected information it had listed above. When instructed to do so it came up with this image:
But the information that the chatbot had just shared did not in fact make it into the image accurately.
The mistakes in this image are a little bit harder to catch at first glance - other than the still absent Merz - but there are a few.
As listed above, there were two chancellors (Walter Scheel who was acting chancellor for just nine days, and Helmut Schmidt) between Willy Brandt and Helmut Kohl that GPT has skipped here.
Also, the bot hallucinated a Chancellor Helmut Schröder (a sort-of portmanteau of Helmut Kohl and Gerhard Schröder), and adds Helmut Kohl in again for good measure.
Once again, asked for a source, GPT admitted the image contained some mistakes, and correctly pointed out that Kohl's term had been incorrectly split and that both Schmidt and acting chancellor Scheel had been left out.
In explanation, GPT added, "The source wasn’t pulled from a live database — it came from my internal historical knowledge of Germany’s political history, which I then tried to visualize. However, unlike a fact-checked list from an official source, this process can produce small but important errors when turning names and dates into images."
Once again GPT offered to try again, suggesting it could generate a "fully accurate, verified chancellor timeline".
The chatbot made no mention of Friedrich Merz.
The importance of official sources
Silly as it is, GPT's inability to accurately portray Germany's chancellors illustrates a bigger issue with growing use of LLM chatbots - that they are very often wrong.
Increasingly many people turn to ChatGPT and similar products as if they are primary sources of information. Yet studies have shown that more than half of ChatGPT's answers contain misinformation.
READ ALSO: How German schools are including more and more AI
Of course, mixing up your German chancellors in a pub quiz would be one thing, but reliance on chatbots could have far-reaching and serious consequences in other situations - like if the Prime Minister of Sweden were to use them to get a second opinion on policies, for example. (Yes, that's something he admitted to doing.)
This is to say nothing of AI chatbots' devastating environmental impacts - they require a tremendous amount of energy and water.
If you do use generative AI, you'd be advised to do so sparingly, and be sure to check other sources.
Comments