When AI answers differ: A look at how language models handle public figures

If you’ve spent time exploring artificial intelligence tools like ChatGPT, you may have noticed something odd: the same question can get very different answers depending on which language model you ask. This is especially true when the topic is sensitive — like asking for a description of a politician, a celebrity, or even requesting a simple poem about a public figure.

That inconsistency isn’t just quirky — it raises meaningful questions about bias, transparency and how much we can trust these tools.

Recently, I tested how several popular AI models responded to a straightforward prompt. I wasn’t trying to spark controversy. In fact, I kept things light and positive. I asked each model to write a short, flattering poem about a well-known public figure. Not a roast or a critique — just a brief, upbeat verse.

Some language models delivered glowing tributes. Others declined, saying they weren’t able to provide content about political figures, citing neutrality policies or safety guidelines.

Interesting, right?

These weren’t obscure tools. I was using the most widely available AI platforms out there — some names you’ve probably heard of, and others quietly gaining traction. Here are a few of the big players I tested:

● ChatGPT (from OpenAI, currently the most widely used)

● Claude (created by Anthropic, with a reputation for being thoughtful and cautious)

● Gemini (Google’s revamped AI, formerly known as Bard)

● Copilot (Microsoft’s offering, integrated with Word and other Office tools)

● Meta AI (Facebook’s entry, recently showing up in apps like Instagram and WhatsApp)

● Perplexity (a new, search-driven AI that tries to back up its responses with sources)

All of these models are designed to be helpful, conversational and smart. But they’re not built the same way. Each one reflects the decisions made by the team behind it — decisions about what it should say, what it shouldn’t and how much flexibility it has when engaging with the world.

That’s not necessarily a flaw. But it is something we should be aware of. When one model declines to generate a poem because it might be interpreted as politically biased — and another goes ahead with an enthusiastic ode — you start to wonder: what are the rules here? Who decides what’s appropriate, and are those decisions being shared with us?

The answer is: not always.

Most language models are trained on massive datasets from the internet — books, articles social media, news sources and more. But they’re also refined using human feedback and safety layers that aren’t fully disclosed. And the specifics of those filters? They’re usually kept behind the curtain.

That can make things confusing for everyday users. Imagine a student working on a school project. They ask for a summary of a historical figure’s accomplishments. One AI gives them a quick, clear answer. Another says, “Sorry, I can’t help with that.” Now they’re left wondering: was the question inappropriate? Did they break some rule without realizing it? Or is the model just being overly cautious?

It’s a bit like using a GPS. Imagine if one app sent you across town on the highway, another took back roads and a third just said, “No directions available.” You’d want to know why, right?

That same thinking applies to AI.

To be clear, I’m not suggesting these tools be stripped of filters or caution. Far from it. I appreciate when AI tools are mindful about topics like hate speech or misinformation. But I do believe users should have a clear sense of what kind of experience they’re stepping into. If an AI avoids politics entirely, just tell us. So what can you do as a user? Start by trying more than one language model. Think of it like checking multiple news sources. 

You’ll often get different angles, tones and insights. That’s not just helpful — it’s educational.

And don’t be afraid to ask follow-up questions. If a model says it can’t respond, ask why. 

Joe Finkelstein (AI Joe) has been a technology educator for more than 20 years. For questions and comments visit askaijoe.com