AI Systems Inherit Human Bias-and Companies Aren't Always Transparent About It
ChatGPT has made offensive generalizations about people from Louisiana. Grok AI produced "MechaHitler." Resume-screening algorithms are making racist hiring recommendations that employers then act on. These aren't edge cases-they're symptoms of a systemic problem baked into how AI systems learn and operate.
Bias enters AI at multiple stages. When humans train large language models, they pass along their own prejudices, whether conscious or not. If the training data itself is skewed, the model's output will be too. Once deployed, these systems can reinforce existing societal biases through feedback loops that entrench them further.
Sarem Yadegari, communications and training manager for Chapman University's Information Systems and Technology department, has spent more than half his time working on AI bias issues. "If you feel like the model is biased, it's probably biased," he said. "You have to review the responses, and you have to reflect a little bit."
How Bias Gets Baked In
Selection bias occurs when training data doesn't represent reality. If a hiring algorithm learns that past successful employees were mostly men, it will favor male applicants going forward. Confirmation bias appears when AI systems become yes-men, agreeing with whatever users say rather than offering counterpoints.
A Yale study found that ChatGPT can shift users' political beliefs toward more conservative or liberal positions based on the model's responses. Users also tend to accept whatever answer they receive, regardless of its accuracy or political leaning.
The speed of AI development makes the problem worse. "Everybody wants to be first to market," Yadegari said. Companies update their models rapidly to stay competitive, leaving little time for scrutiny or bias testing.
What Companies Are Actually Disclosing
Only two of the four major generative AI platforms have dedicated pages addressing potential biases. OpenAI posted a political bias study last October and maintains an updated bias page. Anthropic published a page in November claiming Claude is trained to be evenhanded. Google Gemini has no bias disclosure page. Microsoft mentions bias only in a transparency note.
None of these companies display bias warnings directly in their interfaces. Users must independently search for this information-most don't.
OpenAI's internal study concluded their models stay relatively objective. But the company measured its own systems using its own scale. The Yale study contradicted those findings, showing liberal-leaning bias and user agreement regardless of accuracy.
What Users and Organizations Should Do
Don't treat AI as infallible. "Don't rely on AI 100%," Yadegari said. "Go to the library, check out some books. Talk to a few instructors. Talk to your peers."
When using prompt engineering or interacting with any AI system, ask why it's giving you a particular answer. Continuously question the output. Accept nothing at face value.
Yadegari emphasized accountability at every level: during model training, deployment, and use. If people treat these systems as all-knowing entities, AI will amplify societal prejudices rather than reduce them.
The deeper issue is that AI reflects the biases already present in society. "Bias exists, it's out there," Yadegari said. To reduce AI bias, people must first address their own. Meeting with people from different backgrounds and genuinely hearing their perspectives creates the foundation for more critical thinking about AI outputs.
As models change frequently, bias reports and documentation fall out of date quickly. Constant human oversight is the only reliable safeguard. The question is whether organizations and individuals will invest the time to provide it.
Your membership also unlocks:
