Mittelsteadt adds that Trump could punish companies in a variety of ways. For example, he cites how the Trump administration canceled a gigantic federal contract with Amazon Web Services, a decision likely influenced by the former president’s view of the Washington Post and its owner, Jeff Bezos.
Policymakers will have no difficulty pointing out evidence of political bias in AI models, even if it cuts both ways.
AND 2023 study by researchers from the University of Washington, Carnegie Mellon University, and Xi’an Jiaotong University discovered a range of political leanings in various gigantic language models. It also showed how this bias can affect the operation of hate speech and disinformation detection systems.
Another testconducted by researchers at the Hong Kong University of Science and Technology found bias in several open-source AI models regarding polarizing issues such as immigration, reproductive rights and climate change. Yejin Bang, a PhD student involved in this work, says most models tend to be liberal and U.S.-centric, but the same models can express a variety of liberal or conservative biases, depending on the topic.
AI models pick up on political bias because they are trained on areas of internet data that inevitably include all kinds of perspectives. Most users may not be aware of the bias in the tools they apply because the models contain barriers that prevent them from generating certain harmful or biased content. These biases can subtly leak out, however, and the additional training that models receive to limit their performance can introduce further bias. “Developers can ensure that models are presented from multiple perspectives on divisive topics, allowing them to respond with a balanced viewpoint,” Bang says.
The problem could get worse as artificial intelligence systems become more pervasive, he says Ashique KhudaBukhsha computer scientist at the Rochester Institute of Technology who has developed a tool called the Toxicity Rabbit Hole Framework that examines various social biases of gigantic language models. “We fear that a vicious cycle will soon begin as new generations of LLMs will increasingly be trained on data contaminated with AI-generated content,” he says.
“I am convinced that this bias within LLM programs is already a problem and will most likely become even worse in the future,” says Luca Rettenberger, a postdoctoral researcher at the Karlsruhe Institute of Technology, who conducted an LLM analysis on biases related to German politics .
Rettenberger suggests that political groups may also try to influence the LLM to promote their own views over those of others. “If someone is very ambitious and has bad intentions, it is possible to steer the LLM in a certain direction,” he says. “I see manipulation of training data as a real danger.”
Some efforts have already been made to change the balance of bias in AI models. Last March, one developer developed a more right-wing chatbot in an attempt to highlight subtle flaws he saw in tools like ChatGPT. Musk himself has promised that Grok, the AI chatbot built by xAI, will be “maximum truth-seeking” and less biased than other AI tools, although in practice it also hedges on complex political issues. (Musk, a staunch Trump supporter and immigration hawk, his “less partisan” view may also translate into more right-wing results).
Next week’s US elections are unlikely to ease the discord between Democrats and Republicans, but if Trump wins, talk of anti-woke AI could become much louder.
At this week’s event, Musk took an apocalyptic approach to the issue, referencing the incident when Google-owned Gemini said nuclear war would be better than misleading Caitlyn Jenner. “If you had an AI programmed to do this sort of thing, it might conclude that the best way to make sure no one gets misgendered is to annihilate all humans, making the chance of future misgendering zero.” he said.