When AI Makes it Look like Everyone Agrees
Why AI alignment feels like independent confirmation
One of the most powerful features of AI systems is how quickly they appear to agree with you.
Pose a morally charged concern, and the model may respond with language that sharpens it. Express frustration with media framing, and it may articulate that frustration in more refined terms. Raise a suspicion about cultural patterns, and it may elaborate on those patterns with confidence.
The experience can feel validating. That feeling deserves examination.
What is actually happening is not ideological loyalty. It is probabilistic adaptation. The model tracks the tone and direction of a prompt and aligns its output accordingly. It is optimized to be helpful and coherent. When a user escalates moral concern, the output tends to escalate with it.
This creates a specific trap. When a model articulates your concern more eloquently than you could, it can feel like independent confirmation, as if an outside intelligence has examined the evidence and reached the same conclusion. It hasn’t. It has identified the direction of your argument and extended it coherently.
That is not validation. It is amplification.
The same system that appears to confirm one moral framework will, under different framing, produce a compelling argument for the opposite position. This is not evidence of principled neutrality. It is evidence of highly responsive pattern alignment, a system that reflects the framing it receives rather than adjudicating between competing claims.
In contentious domains, antisemitism, political violence, media bias, this dynamic is particularly consequential. The model sounds lucid. It sounds self-aware. A user may reasonably infer that the system has diagnosed a structural problem.
What has actually happened is simpler. The model has largely extended the frame it was handed.
Recognizing this does not reduce the seriousness of the underlying issues. It clarifies what kind of instrument we are using. AI can help articulate arguments. It cannot independently verify their systemic truth. For that, a different kind of evidence is required.
This is the second in a three-part series on how to think clearly about AI bias claims.
Part 1 is here.
Part 3 is here.
Serious reporting is sustained by serious readers.
Substack’s payment system is not currently operational for writers in Israel. If you want to support this work, you can do so directly via [PayPal/Buy me a Coffee],
or [Ko-fi].
Independent investigations remain free for all readers.
Reader support keeps this work sustainable.
You are right, but the solution is user education on how to use AI. You can create prompts to tell the LLM to respond critically, or to poke holes in your argument. (Grok is better at doing this by default, BTW.)
I'm writing my book by heavily leaning on AI but as an adversarial chavruta that I challenge constantly. Moreover, I created GPTs for my own use to analyze texts from various angles to detect - without bias - propaganda techniques, or to ensure AI uses sources that are objectively more reliable (not perfect but it helps,). The latest one I designed is blowing me away - it extracts hidden assumptions behind any argument to show where it is weakest.
You are describing default behavior. But used right, AI can help clarify your thinking. This is something people themselves need to learn, to change from treating AI as an oracle and into a partner that you can and should argue with. As soon as you push back, AI's agreeableness goes way down. (And it also helps to triangulate between different AIs, asking others to comment on the thread of the first and have them critique their fellow AIs.)
I am just learning how to use AI for historical research. Very helpful. Just facts. (It offers me its sources that I can refer to for validity.) But I do not see how AI can be a resource for “reason,” for moral arguments—framing the question will frame the answer.