The danger of trusting a chatbot with the right opinions

My colleague Ronald MacMillan and I have been writing quite a bit about AI on our Five4Faith Substack. Ronald wrote about Pope Leo XIV’s new encyclical on artificial intelligence. I wrote about Gloo, the faith tech platform whose chatbot calibrates spiritual answers to what you already believe. Today, a confession: I'm writing this opinion because the topic keeps finding me.

The system is trained on a conservative worldview, so that it gives conservative answers.

I was recently invited to contribute to the training of a conservative chatbot. The idea is what you might expect. The system is trained on a conservative worldview, so that it gives conservative answers to your questions and provides the perspective that is closer to what we should all follow... Ahem.

I declined. But the invitation stayed with me, because it is a small window into something much larger.

Is AI actually liberal?

Let us give the project its due, because the diagnosis behind it is not absurd. Existing Large Language Models (LLMs, aka the coding behind AI) are widely considered to lean liberal and progressive. They were developed in Silicon Valley, which is indeed known for having a particular worldview.

Then again, the picture is messier than the complaint suggests. Peter Thiel and Elon Musk, hardly marginal figures in the industry, have taken some decidedly conservative turns. Most founders of the large AI labs lean center-left, but overall their political views are mixed.

Mainstream AI models are trained on mainstream data.

What is true is that mainstream AI models are trained on mainstream data. And of course conservatives do not like that, because they are concerned, with or without reason, about the direction the world is taking.

Whoever feels that the mainstream has drifted away from them will experience a machine trained on the mainstream as biased. Counter-reactions were entirely predictable. The conservative chatbot is one of them.

What a training project actually reveals

Here is what struck me most about the invitation, though. It was not the politics. It was what the whole endeavor admits about the technology.

If you can train a chatbot to be reliably conservative, then these systems are not oracles. They are instruments, and remarkably pliable ones.

ChatGPT is known for its flattery, congratulating you on every question you ask; OpenAI has itself acknowledged and rolled back an update that made the model too sycophantic. Claude is more neutral in tone.

You can configure these systems to be more or less agreeable.

Regular users do not know this, but you can configure these systems to be more or less agreeable. And you do not need a training budget at all: you can simply instruct any of them, in plain language, to answer only from a conservative perspective, or a progressive one, or a Marxist one, or a medieval Scholastic one.

A worldview that can be installed with a system prompt is not a worldview. It is a costume. The conservative chatbot does not prove that AI has finally been aligned with truth. It proves how manipulable these machines are, and how willing we have become to do the manipulating, so long as the machine ends up agreeing with us.

The politicization of everything: machine edition

I am intrigued by the ongoing politicization of everything, including religious freedom, and it all has an underneath tribal logic. The conservative chatbot is that same logic, now applied to the tools we use to think.

Consider what it means for both sides of this market to succeed. Progressives get models trained on mainstream data and tuned by mainstream labs. Conservatives get counter-models trained to give the so-called "right answers", where "right" is defined in advance. 

A machine built to tell you what you already think.

Gloo’s chatbot calibrates itself to your beliefs in the name of faith; the conservative chatbot calibrates itself to your beliefs in the name of truth. These are the same product with different labels: a machine built to tell you what you already think.

You are outsourcing your confirmation bias.

Until now, polarization meant we could no longer agree on answers. Politicized AI goes one step further. When even the instrument you consult has been tuned to your tribe before you type a word, you are no longer asking questions at all. You are outsourcing your confirmation bias, at scale, with excellent grammar.

That is why I declined. Not because a conservative perspective does not deserve articulation. It does, and I share more of it than my progressive friends suspect (but less than my conservative friends expect!). I declined because I do not want a machine with the right opinions, mine or anyone else’s. If I want something that agrees with me, I already own a mirror.

What I actually need, and what no tribe will ever build for me, are instruments that help me find out when I am wrong. That used to be called education. It also used to be called discipleship. It would be a strange irony if the most advanced technology in human history became the first tool deliberately designed to prevent both.

Originally published as a Five4Faith Substack article. Republished with permission.

Dennis P. Petri, PhD is the International Director of the International Institute for Religious Freedom and Founder and scholar-at-large of the Observatory of Religious Freedom in Latin America. He is a Professor in International Relations at the Latin American University of Science and Technology. He is the author of The Specific Vulnerability of Religious Minorities, a book on undetected religious freedom challenges in Latin America.