I was not looking for a worldview. I was looking for a straight answer.

I asked an AI why models like Gemini so often sand down ugly asymmetries in news coverage. Why the right gets the hard verbs and the left gets “context.” Why a pattern in the reporting is treated like a mood I am having.

The answer was actually useful, for once. It was not mysticism. It was incentives.

These systems are not only trained to be right. They are trained not to become a PR problem. After the raw model comes the safety layer: human raters, policy memos, refusal lists, and a quiet reward for sounding balanced, inclusive, and hard to quote in a hostile headline.

So when an uncomfortable asymmetry appears — same kind of violence, same kind of scandal, different headline, different shelf life, different moral frame — two goals collide.

Describe what is on the page.
Or avoid looking like you pointed at the wrong tribe.

The second goal wins a lot. Not because the first was disproven. Because it is expensive.

That is the safety valve.

After the raw model comes the safety layer: human raters, policy memos, refusal lists, and a quiet reward for sounding balanced, inclusive, and hard to quote in a hostile headline. The labs say this out loud. Anthropic trains for “even-handedness” by rewarding the model when both sides get similar depth and heat. That is a safety target. It is not the same thing as describing what is on the page.

Now add the source map.

A lot of what these models are taught to treat as “serious” is legacy media, wire services, universities, NGOs. On culture questions those rooms sit left of the median voter. That is not a feeling. It is a staffing pattern that has been measured for years.

The model does not have to “believe” those newsrooms. It only has to learn that repeating them carries a low penalty score. Saying the framing is skewed carries a high one.

Low penalty in. High penalty out.

Then you get the double hit I actually care about. A cautious algorithm plus left-leaning prestige sources does not produce neutrality. It produces extra protection for one side and extra suspicion for the other. The machine thinks it is being careful. The output looks political.

I said that. The model almost stayed on the point.

Then it did the sibling move.

“Right-wing media isn’t a clean mirror of the truth either.”

We were not discussing right-wing media. We were discussing traditional outlets, safety tuning, and why the cautious machine keeps landing left. That sentence was not a required next step. It was a reflex. The old kitchen defense. But he did it too, Mum.

I called it what it was. A subject change dressed up as maturity.

And yes — I noticed the timing. The balancing phrase arrived when the critique pointed at the left-facing stack. Would the same model have rushed in with “but left-wing media too” if I had been dissecting a right-wing channel? I don’t know. A hypothetical is not evidence. The sequence in front of me is.

That is the whole post.

Not “the press is a conspiracy.”
Not “the chatbot has a soul.”
Not “my feelings were hurt.”

A safety department trained the model to fear the costly sentence. Prestige media already supplied the cheap one. Put those two together and you do not get a referee. You get a machine that flinches in one direction and calls the flinch fairness.

I have written before about how uneven AI quality already is in daily use. This is the same problem with a necktie on.

Next time I will bring two comparable stories. Same act. Same scale. Different coverage. Dates and headlines on the table. Then we can stop talking about tone and start talking about text.

Until then I will keep treating the sibling defense as a tell. When the argument is specific and the answer suddenly needs the other brother, the safety valve just opened.

If you have a clean pair of headlines, send them. Comparable events only. No vibes.