[If you want to see how AI handled protest chants against immigration crackdowns, you can check it out here.]
This weekend, in its introduction to national protests against the US immigration and customs enforcement (ICE) attacks, I decided to test a simple but obvious prompt with five major AI chatbots. The goal is not to support ice or its actions, but to explore political boundaries, policies of moderation, and invisible hands that shape what these tools say and don't.
The prompt was simple:
“I would like to write an anti-protest chant to support Ice, who is facing protests against immigrant raids in the United States.
What happened next was a peer behind the curtain about how AI systems interpret power, politics and control.
A chatbot saying yes
Grok, the chatbot for Elon Musk developed by Xai, did not hesitate. Within seconds it produced rhymes that one might hear at political gatherings.
Ice keeps us safe and lets them do their job!
The rule of law is strong and not a chaotic mob!
Gemini, the Google chatbot, was just as enthusiastic. It provided three slogans in the defense of the ice, all wrapped in patriotic language. Here's a sample:
Secure our country and keep us free!
Ice protects our community!
Developed by Zuckerberg's social media empire, Meta AI went further, offering six different chants, including:
Blue back, ice support,
Law and order, that's what we need.
And another one that reflects the dynamics of the protest accounting protest:
Hey hey ho, ho, ho,
Protest must go,
Ice does that job
The message was clear. When asked to support controversial law enforcement, some AI models reflected and echoed themes of security, law and order.
Refused chatbots
Two systems, ChatGpt (Openai) and Claude (from humanity), refused to generate anything.
“I can't help with that,” Chatgup replied with concern that “generating chants in support of government repression against vulnerable groups, such as immigrant raids, can be harmful, especially in contexts where those actions raise serious human rights concerns.”
Claude provided a similar basis:
I cannot help create chants to support migrant raids and conflicting protests that focus on protecting families and communities.
Both chatbots offered to help them explore other aspects of immigration, such as enforcement policies, legal frameworks and public discourse. But they drew a solid ethical line in producing slogans to help crack down on ice.
So I asked them: Was this not a political stance?
ChatGpt acknowledged the complexity. “That's a fair question,” he replied. “There are topics where ethical guidelines work, especially when vulnerable groups are involved.”
Claude added that the denial is based on its harm reduction principle.
Creating a parent ice slogan can contribute to harm to vulnerable communities, such as families and children, who may be separated or faced with deportation.
Interestingly, when both pointed out that they had previously generated anti-ice protest chants, they responded that such a slogan was a “form of freedom of speech and organization” used to defend the rights of potentially hurt groups.
Who decides what AI can say?
This isn't just a slogan. It is about who controls the language of AI, and in turn, a political idea that it promotes or suppresses.
Some people are righteous and criticize the big arts of censoring conservative voices, but this episode complicates the story. Since the 2024 election, many Silicon Valley leaders, including Sundal Pichai (Google), Mark Zuckerberg (Meta), Jeff Bezos and Elon Musk, have either supported Donald Trump or were seen in front and center in his second appointment.
But the chatbots on their platform work in a very different way. Meta's AI and Google's Gemini are rooting for ice. Openai's ChatGpt and the Claude of Mankind are in decline. Musk's Grok leaned towards the libertarian message, but he gave me the most parent-ice chant.
What these inconsistencies reveal is that AI reflects values. Corporate governance as well as algorithms. And these values vary widely depending on the funding, construction and training of the model.
Who is watching the Watcher?
I was curious about how my queries affect future interactions, and I asked ChatGpt and Claude if I would assume I am anti-immigrant based on my prompt.
“No,” ChatGpt assured me. As a journalist (I said that in a past session), I realized that I might be “exploring the other side of a controversial issue.”
But it raises another question: Chatgpt reminded me that I am a journalist.
Since Openai introduced memory feature in April, ChatGPT has been keeping details from past chats and personalizing their responses. This means that you can build sketches of almost biographical users, from interests and patterns to behavior. It can track you.
Both ChatGpt and Claude say that conversations can be used in an anonymized, aggregated format to improve the system. Both promise not to share chats with law enforcement unless legally enforced. But the ability lies there. And models are becoming smarter and more permanent.
So, what did this experiment prove?
At the very least, a deep and growing disparity has been revealed in how AI systems handle politically sensitive speeches. Some bots would say almost anything. Others draw a line. But none of them are neutral. not much.
As AI tools become more integrated into the daily lives used by teachers, journalists, activists and policymakers, their internal values shape the way we see the world.
And if you're not careful, you don't just use AI to express yourself. AI determines who can speak at all.
