ChatGPT’s Bias Bites in Bytes

In our first column, we discussed Wikipedia. The website is quick, the information is digestible, and it’s a great first place to turn to when researching a new topic. Beyond this, we praised the site’s open-source editing process, which allows for a broad range of perspectives and aims to yield unbiased information.

But recently, the lightning-fast development of generative AI platforms has many students — ourselves included — turning elsewhere. In fact, the Faculty of Arts and Sciences released its first guidance regarding the use of generative AI in Harvard classrooms this summer, demonstrating that this technology is becoming an indispensable part of the student experience.

Like Wikipedia, generative AI platforms like OpenAI’s ChatGPT and Google’s Bard provide specific, easy-to-read information on demand. But where Wikipedia succeeds with its overall veracity, transparency, and pluralistic content production, these AI products fail.

Generative AI technologies often demonstrate social and political bias, exhibit censorship, and provide inaccurate or incomplete information. If we over-rely on generative AI as a source of information with which to develop our opinions, these features can limit our viewpoints and worsen our ensuing discussions.

Although AI is continually updating and improving, research on large language models has demonstrated significant biases, which may unduly cloud our perspectives.

A study earlier this year found that ChatGPT favored leftist and libertarian stances; models developed by Meta like LLaMA and RoBERTa, comparatively, tended more towards authoritarianism.

Research published last month by a team from Adobe, University of California, Los Angeles, and University of Southern California showed that when ChatGPT and Alpaca were asked to produce recommendation letters, significant differences in language were used based on the recommendee’s gender.

Part of OpenAI’s approach includes “aligning AI systems with human values.” This process includes a method known as “reinforcement learning from human feedback,” where models receive direct input from human testers.

But there is no uniform or universal set of human values. Values are subjective and unique to each individual. To suggest one set of overarching human values is antithetical to the diversity of perspectives necessary for free and open discourse.

A person’s values can also be the clearest manifestation of their bias. As these values are leveraged during the training process, we can only expect that these biases will make their way into the content that models produce.

Even Sam Altman, the recently reinstated CEO of OpenAI, acknowledges that biases are built directly into the ChatGPT model. We can’t expect that the outputs generative AI provides us are neutral or objective; thus, we can’t reliably use these outputs as foundations for meaningful discussion.

Besides exhibiting biases, generative AI also has censorship problems.

Acceptable use policies have restricted what information an AI technology can provide and what tasks it can complete. Answers to questions relating to controversial topics are often sandwiched between disclaimers, if the AI provides an answer at all.

OpenAI, for example, restricts ChatGPT from generating harmful or offensive content. But as with human values, such a standard is entirely subjective. As we depend more on AI to learn new information, this censorship poses the dangerous threat of narrowing our worldviews and perspectives, molding them to whatever arbitrary moral frameworks are chosen by technology companies.

Furthermore, the use of generative AI hurts the quality of information we consume. When we use a chatbot, we choose to prioritize volume and convenience over depth and quality. ChatGPT will often condense complicated topics into a few hundred words. Going straight to the summary strips us of the nuance we might find in other sources, to the detriment of the quality of our conversations.

AI outputs often lack more than just context — they sometimes lack accuracy too. Factual claims made by large language models are uncited, and OpenAI itself concedes that ChatGPT will occasionally “make up facts or ‘hallucinate’ outputs” to construct a coherent response.

Both ChatGPT and Bard post disclaimers on their websites warning users of these possibilities. In an extreme example of this phenomenon, Google’s stock dropped 9 percent earlier this year in value after Bard was caught in a lie during its very first public demo.

We cannot have a healthy free speech environment without a shared basis of open and robust information. To combat the impact of these generative AI issues on our collective discourse, we must make an effort to learn from a diverse variety of sources — including academic journals, books, and conversations with our peers.

We must verify the information we encounter, and we certainly shouldn’t blindly deem anything generative AI creates as truthful or unbiased. These strategies will help us form well-rounded opinions and better guarantee the veracity of the information we consume.

Generative AI won’t just match the value Wikipedia has for students — it will likely surpass it. But we must quickly understand the technology’s faults.

There’s an important role for AI to play in our lives at Harvard. It can provide writing feedback, or explain difficult mathematical concepts — perhaps some day better than your teaching fellow can. But an overreliance on AI will only serve to harm us in our pursuit of Veritas.

Milo J. Clark ’24 is a Physics concentrator in Lowell House. Tyler S. Young ’26 is a joint concentrator in Electrical Engineering and Chemistry & Physics in Leverett House. Their column, “Voices Unbound,” runs bi-weekly on Tuesdays.