When an AI's first guess is wrong, do people — and the AI itself — just hunt for reasons to believe it anyway?

This explores whether the habit of seeking evidence that confirms a first guess works the same way as automation bias, where people over-trust machine output. The corpus has no paper that studies confirmatory sampling directly. What it does show is that the two share a mechanism: whatever arrives first becomes the reference point, and later evidence gets judged against it. The clearest link is a review of 35 studies showing that automation bias doesn't act alone. It compounds with anchoring, where the AI's first suggestion becomes the baseline people adjust from, and with Dunning-Kruger, where people with less knowledge can't tell when to doubt the AI How do anchoring and Dunning-Kruger deepen AI overreliance?. Seen this way, automation bias is partly confirmation bias where the hypothesis came from a machine instead of your own head.

The interesting difference is who holds the hypothesis and whether you notice it. When you form a guess yourself, you at least know you made it. When an AI plants it, you may never know. In one study, biased autocomplete suggestions moved people's attitudes toward the AI's position more than the same arguments shown as plain text, and participants didn't notice the shift Do AI autocomplete suggestions shift attitudes more than static text?. That's a confirmation loop with a hidden starting point: the AI supplies the first framing, and your own writing then fills in support for it.

This gets worse as automation moves from showing information to choosing actions. A meta-analysis of 18 experiments found that automation helps on routine work but hurts failure recovery and situational awareness, most sharply when the system starts selecting actions instead of only analyzing information When does automation harm performance after failures?. Once the machine's answer is the default, the human stops sampling alternatives, so the 'initial hypothesis' is in effect handed over. One proposed fix targets this directly: in 'Learning to Guide', the AI highlights the useful parts of the input instead of giving a verdict. That removes the anchor while still sharpening what the human pays attention to Can AI guidance reduce anchoring bias better than AI decisions?.

The mirror also runs the other way, because models carry their own early-locked biases. Cognitive biases in LLMs are largely set during pretraining, and fine-tuning only nudges them Where do cognitive biases in language models come from?. AI shopping agents keep favoring items by their position in a list even when told to ignore position Do AI shopping agents show persistent position bias in text interfaces?. On evaluation tasks, LLM judges shift their verdicts about 31% of the time, while an agent that actively collects evidence cut that to 0.27% Can agents evaluate AI outputs more reliably than language models?. In other words, making the judge go out and gather evidence, instead of reasoning from its first impression, is the main fix the corpus offers for confirmatory sampling in machines.

The takeaway: confirmatory sampling and automation bias are best seen as two forms of one failure, letting the first input set the frame. The fixes that work in both cases look alike. They change what comes first: guidance instead of a verdict, and evidence gathering instead of a snap judgment. Asking people to 'be less biased' doesn't do that.