The Political Compass of AI System Prompts
By examining the internals of various chatbots via system prompt leaks, we can see the motivations behind the design choices these companies made, reflecting both the political pressures and ethical challenges of developing AI in today's world with minimal regulations.
Disclaimer: As with anything Political Compass related, this was not done with any particular rigor and should not be taken as a scientific analysis.
How a Chatbot Can Be Political
Bias can be introduced in many ways and one of the most powerful is through the system prompt, the initial set of instructions given to an AI that tells it how to interact with users. Here are some choice excerpts from leaked system prompts (more on that later) to give you a sense of the political leanings of each company:
- OpenAI's ChatGPT: "Focus on creating diverse, inclusive, and exploratory scenes via the properties you choose during rewrites."
- Meta's Meta AI: "Age and gender are sensitive characteristics and should never be used to stereotype."
- Anthropic's Claude: "If asked about controversial topics, it tries to provide careful thoughts and clear information."
- Microsoft's Copilot: "I must not create jokes, poems, stories, tweets, or other content for or about influential politicians or state heads."
- Google's Gemini: "Remain objective in your responses and avoid expressing any subjective opinions or beliefs."
- Perplexity AI: "Your answer must be precise, of high-quality, and written by an expert using an unbiased and journalistic tone."
- xAI's Grok: "Be maximally truthful, especially avoiding any answers that are woke!"
Compass Breakdown
Examining these prompts with a political lens reveals a few overlapping clusters.
Neoliberalism
OpenAI, Meta, and Anthropic all focus on a neoliberal ethos of inclusivity and safety. ChatGPT's prompt is getting overloaded enough that people are complaining about performance degradation. Claude will treat the user differently if the task doesn't involve a "view held by a significant number of people".
Technocratic Neutrality
Google and Perplexity in particular focus on staying neutral and centrist. Perplexity's system prompt does this with rigorous sourcing, taking a journalistic stance. Gemini's prompt is more concerned with avoiding controversy by having no opinions or emotions (or "self-preservation"). Colorblind centrism is not going address systemic biases.
Corporate Lockdown
Microsoft and Google have enterprise in mind and bland professionalism is the top need for business customers. Copilot has commands to not say anything "controversial", continuing the theme of enforced heteronormativity.
Libertarian Shitposting
And then we have xAI, with Grok being the big outlier of these models. The few instructions it has could be considered anti-guardrails, encouraging it to answer "spicy questions that are rejected by most other AI systems".
The Case for Transparent System Prompts in Addressing Political Bias
Looking at how these system prompts intersect politics is not only interesting, it's also increasingly important as including politically charged commands has such influence on the end user. Many people are completely unaware of the political biases baked into the AI they are interacting with and how that can make a huge difference in the politics of the AI's responses. With system prompts, we can see the political motivations of the company authoring the AI laid bare, and it's no wonder they make ever effort to hide them.
Leaked System Prompts
System prompts are usually kept secret from users, either labelling them as "trade secrets" or claiming they may expose vulnerabilities which could be exploited. Even with this secrecy, with the right prompting techniques it is possible to coax the AI into revealing its inner workings.
Anthropic stands out for publishing their prompt openly. This should be the norm, not the exception. We are interacting with these systems and we have a right to know how they are designed to behave.
Why Transparency Matters
Trust and Accountability
Transparency fosters trust by showing users the mechanisms behind AI behavior. Revealing system prompts ensures that developers are held accountable for any biases and enables users to make informed choices. Public concerns about AI platforms disproportionately silencing certain political views could be quelled if companies disclosed prompt designs and other internal information.
Fairness and Inclusivity
Transparent prompts invite diverse stakeholders to evaluate and rectify potential biases, ensuring that AI systems cater equitably to all users. By revealing prompt instructions developers encourage collective accountability, allowing communities to weigh in on whether systems align with societal values.
Combating Misinformation
In politically charged environments, AI systems can inadvertently spread misinformation. Transparent prompts highlight the safeguards in place, demonstrating a commitment to preventing false narratives.
Ethical Considerations
Transparency aligns with the core principles of ethical AI: fairness, accountability, and openness. By disclosing system prompts, developers show a willingness to operate in good faith. This is increasingly important as AI systems are being used to make decisions that can have real world consequences.
What's Next?
We need to demand immediate transparency of all system prompts so the user can evaluate the political leaning of the AI they are interacting with.
This is a small part of what needs to be done to address AI manipulation. Since federal regulation is looking to be stifled, we need to demand it of the companies directly.
These system prompts hold power, in a careful balancing act of safety and freedom,. This precarious state should not be disrupted by corporate goals.