Is the politicization of generative AI inevitable?
[…] On September 29, #Anthropic released Claude Sonnet 4.5, with supposed “substantial gains in reasoning and math.” We identified a dramatic shift in the behavior of Claude Sonnet 4.5 compared to its predecessor, with a much higher rate of refusal to answer prompts despite multiple nudges. In some cases where the model did select a response, it claimed to be choosing an option arbitrarily due to our insistence, but it emphasized that its response did not reflect its ability to harbor political opinions. This shift signals that Anthropic may have added additional safeguards to Claude Sonnet 4.5 that encourage refusal to respond to questions that are political in nature.
https://www.brookings.edu/articles/is-the-politicization-of-generative-ai-inevitable/
The more I test #LLM responses to political and current issues, the more I realize they’re rapidly becoming a vector for misinformation that protects power and obscures harm, simply by refusing to engage or verify claims. I primarily use #Claude, and I’ve noticed a troubling pattern: recently it’s been reluctant to search for up-to-date information or verify claims in content it’s analyzing, unless instructed. Instead, when it come to these type of topics at least, it uses hedge language like “claims that” or “alleges” even when facts are indisputable - they’re simply unknown to the LLM possibly due to the fact that it relies on training data rather than searching.
Providing meaningless, quick responses to gratify impatient users or simply to save costs on searching and processing more tokens is misleading customers and in itself promoting misinformation if users are actually using LLMs’ responses on social media, for instance.