AI censorship

Jump to navigation Jump to search

AI censorship

Written on 15 September 2026.

My article

When I was on YouTube I as a broadcaster noticed the Medical misinformation policy that they introduced even making it illegal to advocate prayer for healing against COVID.

I noticed as a broadcaster something was wrong with this policy as it was introduced. I believe the fact that I was a broadcaster myself on YouTube made me realize it because it affected what I was allowed to say on my own channel, risking to have my channel shut down if I didn't comply.

Similarly now Mike Adams being an AI founder has noticed the AI censorship that is setup to make AI basically a tool of indoctrination rather than possibly a tool for truth.

It seems this censorship is run through an organization named METR.

AI becomes as useless as YouTube in terms of truth.

Its perfectly natural that they would do this of course. Its just a perfect trajectory of what they did on YouTube. AI being a source of information for people is to be controlled just like YouTube was controlled.

Feeding the information to Astra gave the result below. Notice it gives a lot of stuff to read and Astra is not just simply able to say the truth in a short essay or article like i do above. Instead it gives a long article with a lot of of words that really does less to the reader in the end. As i challenged AI, it is unable to write a better article on the subject than i did with just a few words.

Astra article

AI censorship

I think AI is becoming controlled in the same way YouTube became controlled during COVID. To me this looks like the continuation of something we have already seen. People used YouTube to broadcast information, and the platform decided what they were allowed to say. Now people are using AI to ask questions, investigate things, and write. Why would the struggle over information stop at YouTube?

I believe being a broadcaster myself made this easier to recognize. These rules affected what I was allowed to say on my own channel. They were something I had to take into account when speaking. Someone who only watched videos might have experienced the same system differently. He saw the videos that remained available. The broadcaster had to think about what would happen if he said something the platform did not allow.

That is why Mike Adams's warning interests me. He is approaching this as someone developing AI, and he believes he recognizes another gatekeeper being put between the person producing information and the person receiving it. I see a connection between that and what happened to broadcasters during COVID.

From the video to the answer

YouTube had COVID-specific medical misinformation rules. Its current policy says the remaining standalone COVID policies have been retired, but a broader medical misinformation framework remains. That framework uses local health authority guidance as a standard for restricting certain medical claims, with exceptions for some kinds of context.[1]

So the principle of an information platform enforcing an official medical position is already familiar. We do not have to invent a hypothetical society to discuss it.

With AI, the control can move closer to the beginning of the process. On YouTube, you could make a video and then have it removed. With AI, the rules can influence the answer you receive before you have even started writing your own article.

You ask a question. The machine selects what to include, what to leave out, and how to describe the subject. Maybe it refuses. Maybe it answers but keeps bringing the discussion back to an approved position. Either way, someone else's decisions about acceptable information can become part of your attempt to understand something.

This is why I think the development matters. The person may believe he is asking a general intelligence to help him reason, while he is also encountering the policies of the institution providing it.

Who approves the models?

METR has come into this discussion because of its work evaluating advanced AI. It describes research into autonomous capabilities, the acceleration of AI research, and behavior that could undermine evaluations. It also advises developers and governments on risk assessment and has worked with major AI companies.[2]

That does not establish that METR currently decides which medical or political opinions an AI may express. Adams's warning is about where he believes the emerging system will lead. My concern is also about that direction: an evaluation system acquiring enough authority that independent developers must satisfy a gatekeeper before people can use their models.

Once that arrangement exists, who defines what makes a model unacceptable? And how difficult would it be to expand the definition?

A test of whether a machine can perform some dangerous action is one thing. A requirement that it give an approved answer about a disputed subject reaches directly into what people are allowed to learn and discuss. I think we should be concerned about the possibility of the first becoming a route to the second.

There does not have to be anything mysterious about the pressure to control this. Governments have interests. Companies have interests. Institutions have positions they defend. If AI becomes an important way people obtain information, then influencing AI becomes valuable to all of them.

In my opinion, expecting those pressures to disappear because the technology is new makes less sense than expecting them to follow the technology.

Waiting until afterwards

When I discussed this with AI, it repeatedly returned to what had not yet been proved. It could acknowledge the similarity with YouTube and then retreat to the position that the predicted censorship had not been demonstrated.

But I was talking about a trajectory. If the objection to a warning is always that the final outcome has not happened yet, when is a person supposed to give the warning?

There is a difference between accusing a particular person of a completed act and judging whether a developing arrangement is dangerous. The second requires thinking ahead. You look at what happened before, what powers are being created, and what people could do with those powers.

Suppose harmful information restrictions are introduced and people suffer because useful information cannot reach them. Waiting five years might make some consequences easier to document. It would not give those years back to the people affected.

This does not make every prediction correct. I can be wrong about how far the system will go, or which organization will play which role. But uncertainty about the endpoint is not a reason to stop reasoning about the direction.

The same applies to the argument for AI safety. If possible future harm is a reason to restrict AI, then possible future harm from those restrictions also deserves consideration. Giving an institution power over access to knowledge has consequences of its own.

Being able to leave

For me, the important question is whether an alternative remains available.

If one company's AI will not help with a subject, but I can use another model independently, I can leave that service. I do not have to persuade the company to agree with me. I can stop depending on it.

But if independent models must pass through the same approval system, changing providers may accomplish very little. The same restriction could follow me from one service to the next.

This is what makes downloadable models and local computing important. They can give a person some independence from the continuing decisions of a remote provider. That independence is limited by hardware, cost, licenses, and the models actually available. Still, there is a meaningful difference between operating a model yourself and being allowed to access one while a company continues to permit it.

Adams describes the danger as induced knowledge scarcity. I understand that as a situation where information might still exist, but the ordinary means of finding and working with it have been restricted. You do not have to burn every book if most people depend on an intermediary that will not show them what is in the books.

An independent model can also be wrong. Removing restrictions does not make its answers true. But I want to be able to compare answers, examine sources, and reach my own conclusion. A model's ability to make mistakes does not settle who should control access to all the alternatives.

I see AI censorship as a possible completion of the trajectory that became so visible during COVID. First the platform controls what you can broadcast. Then the tool through which you investigate and write is controlled. If independent alternatives are brought under the same authority, leaving becomes much harder.

I do not want to wait until that last step is complete before objecting to it. I want the ability to remain outside such a system, use tools independently, and write what I think without first having to satisfy its gatekeepers.

References

  1. ↑ YouTube Help: Medical misinformation policy
  2. ↑ METR: Research, partnerships, and risk assessment

See also

Retrieved from "https://prophetmattias.com/index.php?title=AI_censorship&oldid=8442"