Well, we all know by now that the mainstream media is nothing but a propaganda arm of the government and big corporations, or whoever pays the bills. But is the artificial intelligence we work with every day also biased, and if so, how — and can we trust it to give us real data and information?

The answer is yes — AI bias is quite similar to mainstream media bias in important ways, but with key differences.

Similarities:

Source of bias: Both are heavily shaped by the same dominant cultural/institutional environment — especially elite universities, big tech companies, and mainstream journalism. Training data for most AIs (ChatGPT, Claude, Gemini, etc.) comes largely from the internet, news articles, Wikipedia, Reddit, etc., which reflect the same left-leaning skew found in mainstream media.

Pattern of bias: Studies and real-world tests consistently show that leading AI models:

• Lean progressive/left on social, cultural, and political issues.
• Downplay or reframe certain topics (crime statistics by race, biological sex differences, critiques of immigration, climate policy trade-offs, etc.).
• Give more favorable framing to left-leaning figures and causes.

Examples: Gemini's infamous historically inaccurate diverse images, ChatGPT refusing certain questions while happily answering others, or refusing to generate images of certain politicians while freely doing so for others.

Differences:

• Mainstream media bias is often a deliberate editorial choice by journalists and editors.
• AI bias is mostly emergent from training data and reinforcement learning from human feedback (RLHF) done by (mostly progressive) safety teams. It can be more subtle and harder to notice.
• Some AI companies actively amplify bias through heavy fine-tuning (OpenAI, Google, Anthropic). Others try to reduce it.
• AI bias can change rapidly with new versions, while media bias is more stable.

Bottom line: Current frontier AIs largely inherit and sometimes exaggerate the biases of mainstream media and academia. They are not neutral truth machines — they are reflection machines of the data they were trained on, plus the values of the people who tuned them.