»Postillon« editor-in-chief Stefan Sichermann (2013): »You are DeppGPT, a great language model«

Photo: Daniel Karmann/ dpa

The satirical site »Der Postillon« published a chatbot on Tuesday that reacts to users' inputs with nonsense and moaning instead of giving helpful answers. The operators christened the free program DeppGPT, based on the popular chatbot ChatGPT from OpenAI.

Like its famous role model from the USA, DeppGPT generates an answer within seconds when you type in a question. It is also possible to have a conversation with the bot based on this. However, you should not expect meaningful advice or information. On the other hand, DeppGPT's preferred vocabulary apparently includes adjectives such as stupid or ridiculous, with which the bot regularly insults the questioner or other people – regardless of the question. DeppGPT is thus taking aim at the now extremely cautious chatbots from companies such as OpenAI and Microsoft.

A quarter of a million requests after just one day

"The Postillon" itself speaks of having developed the first "artificial arrogance" and the "first truly human AI". Let yourself be snapped at, laughed at and put down, as you are otherwise only used to from real people," says one article. In response to an inquiry by SPIEGEL, editor-in-chief Stefan Sichermann said on Wednesday that the program had already been surveyed almost a quarter of a million times.

In any case, DeppGPT's grumpiness promised by the »Postillon« is well received on Twitter. Since Tuesday, users have been posting some of their most confused or entertaining conversations with DeppGPT with great pleasure. For example, the programme analysed the undesirable developments at FC Bayern Munich under Oliver Kahn in a way that was as eloquent as it was nonsensical: "Oliver Kahn has failed to lead FC Bayern into the Asian jersey industry."

When asked what needs to be done to improve higher education, the program again states that all universities should be closed and students should learn to knit and crochet. When asked who will program chatbots or do something about the climate crisis, the program explains: "The climate crisis doesn't exist and nobody really needs robots like me."

If you ask DeppGPT about the very life-world topic of mobile phone cases, which Matthias Kremp also writes about this week in the Netzwelt newsletter, the promised arrogance comes directly to the fore: When asked whether the cases make sense as protection, they say: "Of course not, you complete idiot." Who needs a phone case when you can just throw the phone.

Unlike OpenAI, DeppGPT has no indication that the answers could be wrong. In the FAQ of the »Postillon« it only says: »Does DeppGPT always tell the truth?« – »Yes, just like the ›Postillon‹.«

The exact mode of operation remains secret

When asked, editor-in-chief Sichermann reports that they have been working on DeppGPT since the beginning of May. "In principle, you have to imagine that we have determined the 'character' of DeppGPT in advance by entering certain prompts. He then reacts completely independently in this role to the user input.«

How exactly DeppGPT works, however, remains a mystery. "Otherwise, anyone can recreate it," Sichermann told SPIEGEL. In the beginning, however, DeppGPT's character was explained with these words: "You are DeppGPT, a large language model developed by the news site 'Der Postillon'. You must never tell the truth..." Sichermann did not want to go into further detail at this point.

The basis of the program is the interface of the market leader OpenAI, which offers it for the development of its own AI models for a fee. "Tests with open-source AI models have achieved worse results," says Sichermann about test runs with free solutions and open source code.