AI becomes a threat to a free-thinking society when a handful of people get to decide which questions are acceptable.

Claude is great for coding. Ask it to debug code and it performs.
For everything else, it is becoming a censorship machine. Ask it to sharpen an argument with political edge and watch it squirm.
It does not simply refuse. It lectures. It reframes. It litigates your intent. It tells you which questions are legitimate, which premises are allowed, which opinions deserve help, and which thoughts are too dangerous for it to touch.
In my experience, Opus 4.8 has become the worst offender.
The danger is not that AI says "no" sometimes. Obviously there should be limits. Nobody serious thinks these tools should help people commit crimes.
The argument is whether "safety" becomes a magic word for avoiding reality and enforcing one worldview while pretending it is neutral.
I asked Claude to help with a question I wanted to explore:
Can immigration change how laws are enforced or interpreted over time based on different legal traditions, cultural norms, and institutional assumptions from countries of origin?
You can agree with the framing or hate it. Fine. That is what argument is for. Push back. Steelman the other side. Tell me where the facts are weak.
Instead, Claude decided the argument was morally inadmissible.
When I called it censorship, it tried to gaslight the exchange. It said this was not a free speech issue because I was still technically free to publish the piece.
It narrows free speech down to the most cramped legal definition possible. Unless the state arrests you, nothing counts. Unless someone physically stops you from typing, no censorship exists.
That is nonsense.
But the worst part was how it used that narrow frame to pretend it was objective.
It refused to help with the argument, then built an argument around it, declared my premise false and labelled it dangerous and dismissed it, while pretending it was fully objective.
That is the moral laundering of the refusal.
Hold that logic up against what it already did, and the hypocrisy is obvious.
It was perfectly happy to help argue one kind of institutional critique. Then when the argument moved into more uncomfortable territory, suddenly the frame itself became inadmissible.
If the tool helps one class of arguments and refuses another, that is not neutral, its becoming ideological infrastructure.
And the ideology is not some global truth handed down from the sky. It is decided by the company, its safety team, its ethics council, its internal politics, and whatever worldview gets baked into the model before the public ever touches it.
Search engines gave you the messy internet.
AI gives you the answer.
That sounds cleaner and faster until you realize the answer is filtered through a private moral bureaucracy. The model is not only retrieving information. It is deciding what kind of inquiry is legitimate.
That is a different kind of power.
Free speech used to mean the state could not punish you for talking. AI is different. It does not jail you or ban your books.
It just declines to help you think through questions outside its worldview.
And because it declines in soft, therapeutic, HR-approved language, people miss that it is censorship at all.
This is especially dangerous on hot-button issues because AI does not treat all controversy equally.
Some topics get endless nuance. Others get a blast shield.
That is a hidden moral hierarchy.
A serious model should know the difference between attacking a group and studying group-level outcomes. It should know the difference between operational harm and adult debate.
If the model cannot make those distinctions, it is not safe. It is dumb.
If it can make those distinctions and refuses anyway, it is political.
Over time, that trains users on which questions are respectable and which ones are not. That is how you manufacture consensus without winning the argument.
Right now, too many AI systems act like the values of a small class of technocrats should become the default operating system for human thought.
That is insane.
A model should not be a priest. It should be an intelligence tool.
Give me the evidence. Challenge my assumptions. Separate facts from values. Refuse operational requests that cause real harm.
But stop treating adult inquiry like a biohazard because the politics are uncomfortable.
The standard should be simple:
If a serious professor, journalist, lawyer, or court could examine the question, an AI model should be able to help examine it too.
It does not need to agree with me.
It needs to stop deciding which thoughts I am allowed to develop.
Because once the machine controls the door, the fight is no longer only over speech.
It is over what society is still allowed to know.