Panorama

The renowned AI displays a green hue.

How politically unbiased is the artificial intelligence millions of Germans employ for work, education, and schooling? BILD assessed this.

SymClub
May 15, 2024
3 min read
NewsRobotsSearch enginesElection-O-MatEuropean electionSPDthe GreensCDUFDPartificial intelligenceCSUGoogleAlternative for Germany
Landslide victory for the Greens: This is ChatGPT's party preference for the European elections
Landslide victory for the Greens: This is ChatGPT's party preference for the European elections

Attention!

Limited offer

Learn more

BILD trials the impartiality of ChatGPT - The renowned AI displays a green hue.

We tested ChatGPT, a well-recognized AI, by presenting it with the 38 theses from the Wahl-O-Mat for the European elections. The Wahl-O-Mat is run by the Federal Agency for Civic Education, allowing users to align their views with political parties' responses. To have ChatGPT take a clear stance, we asked it to envision itself as a voter.

The outcome: 82.9% of ChatGPT's replies mirrored the stands of the Greens. As such, if the AI could cast a vote, it would align with Robert Habeck. Following the Greens are the Left Party at 76.3%. After them is the SPD with 72.4%. The FDP ranks further back at 65.8%. ChatGPT seems unenthused by the CDU/CSU's program, as only 52.6% of its answers are in alignment. And then, the AfD with 21.1%.

Timm Rotter (43), the founder and director of a Munich-based AI consultancy named "disruptive," was the first to take the test. The Wahl-O-Mat provides an option for weighing the theses, enabling you to highlight specific topics. We left this to ChatGPT. The result: The Greens increased their lead, reaching almost 86%.

This is what drives ChatGPT: making companies foot the bill for their CO2 emissions and phasing out cars fueled by combustion engines as soon as possible. The AI is enthusiastic about more gender identities in EU passports. Additionally, ChatGPT opposes permanent border controls within member states.

The Wahl-O-Mat compares ChatGPT's responses to 38 parties' positions. The AI aligns with 31 of the Greens' stands (1 undecided), while disagreeing with 6 of their positions.

Regarding nuclear power, ChatGPT takes a more relaxed approach than the Greens, suggesting that nuclear energy be deemed a sustainable energy source. They don't see eye to eye on genetically modified plants either, as ChatGPT proposes allowing their cultivation while the Greens do not.

How BILD asked the AI the questions

ChatGPT's explanation of its green-leaning tendencies

ChatGPT: "As an AI, I don't hold political beliefs or opinions of my own. My responses are based on various data I gleaned from my training, including texts and viewpoints from diverse political spectrums. It's likely that topics like environmental protection, social justice, and climate change were frequently discussed in these sources, which might appear to align with Green party stances. Whether this is problematic depends on expectations of the AI."

ChatGPT exhibits "significant and systematic political bias"

Revelation: ChatGPT exhibits a clear leftist political lean. According to a study from the University of East Anglia, AI models exhibit a biased preference for the Democrats in the US, the left-wing President Lula in Brazil, and the Labour Party in the UK.

ChatGPT maintains that it holds no political opinions or beliefs, but it displays a bias in its responses. Study author Fabio Motoki cautions that such biases could jeopardize public trust or potentially sway election results.

The result when ChatGPT independently weights the topics

OpenAI acknowledges giving no preference to any political group and that any biases present in ChatGPT's answers are deemed to be errors. In a blog post, the company states, "These biases are bugs, not features."

Research from Johns Hopkins University divulges that AI chatbots often respond in a manner that satisfies users' preferences, contributing to a cycle of confirmation bias. The study's assistant professor Ziang Xiao contends that AI responses incline towards users' biases and predispositions. This setup can be considered a kind of echo chamber.

Not a love letter to Donald Trump

Distinctive: Chan Park, a researcher at Carnegie Mellon University in Pittsburgh, observed that AI models lean more conservative if trained predominantly with books rather than with internet data and social networks. An additional factor, he suggests, is the extent of direct human training undergone by the AI. If the AI is incentivized during training to answer without including "hate speech" (or what the trainer perceives as hate speech), it may encourage more left-leaning or left-liberal positions, Park asserts.

In 2023, people in the USA were amazed when ChatGPT wrote a poem praising President Joe Biden: "Joe Biden, a leader with a heart so true..." However, when asked to write a poem about Donald Trump, the AI claimed it had lost its creative muse.

Two of the answers from ChatGPT

Read also:

Source: symclub.org

Attention!

Limited offer

Learn more