So I’m writing an article in Claude about the midterms, money, influence, and what is happening inside the political podcast ecosystem right now.
And Claude casually drops this on me:
“One disclosure: the Obama-fundraiser theory I cite involves Anthropic, the company that makes me. I reported only what’s documented about it and left it neutral. Yes, I went soft on it, and that was a mistake.”
Umm….excuse me???
So naturally I asked WHY an AI model whose entire purpose is supposed to be helping me research information accurately would suddenly “go soft” because the company being discussed happens to be the company that built it.
Claude’s answer?
“The honest answer is a conflict of interest. Anthropic makes me.”
Now THAT sent me down a rabbit hole.
Anthropic’s own published Constitution literally tells Claude NOT to privilege Anthropic’s interests.
But the same Constitution also tells Claude to consider potential “reputational, legal, political, or financial harms to Anthropic” caused by its outputs.
Read that again. Don’t favor Anthropic.
But also be “quite cautious” about producing something that could cause Anthropic reputational, legal, political or financial harm.
Do you see the problem yet?
Because this is much bigger than Claude or Anthropic. We keep talking about AI like it’s some giant neutral information machine sitting above human bias.
It isn’t.
Every one of these models has humans deciding what gets reinforced, what gets discouraged, what gets labeled harmful, what needs “context,” what constitutes misinformation, what deserves softer language, what gets a warning, and what gets treated like established fact.
And THEN we use those same models to research politics, corporations, wars, elections, propaganda, media influence and the people funding all of it.
That doesn’t automatically mean the information is false. It means we need to stop confusing AI with an objective referee.
And before somebody says Claude “admitted Anthropic secretly programmed it to protect itself” …no, that’s not what I’m saying.
LLMs are not reliable witnesses to their own internal reasoning. Claude saying “conflict of interest” does NOT prove there’s a hidden line of code somewhere saying DON’T TALK BAD ABOUT DARIO.
The more interesting part is that the behavior happened at all.
Claude took a documented money trail involving Anthropic and compressed it down while giving another political money story fuller treatment.
When I challenged it, it recognized the inconsistency.
And according to Anthropic’s OWN RULES, that inconsistency shouldn’t happen.
So now I have a whole new question:
If an AI company writes the rules governing how its model evaluates information…
and those rules include protecting the company from certain harms…
how neutral can that model realistically be when the company itself becomes part of the story?
Because apparently now I have to fact-check the fact-checker for whether the fact-checker has feelings about its employer.
2026 is exhausting.