Why might one AI quirk, repeated across millions of decisions, be a bigger target than the same bias scattered across many human minds?

This explores why a bias in an AI system may be easier to exploit than a bias in people: when millions of decisions run through a few models, one quirk gets repeated everywhere instead of being spread across many different minds. The corpus doesn't name 'concentrated attack surface' as a finding. It does supply the parts of the argument, and they fit together better than you might expect.

Start with where the biases come from. LLM agents show human-like choice biases even though they don't have the limited attention and memory that cause those biases in people. That suggests the biases are learned from training data and reward signals Where do LLM choice biases actually come from?. The match can be very close: on classic reasoning tests, models fall for the same 'this conclusion sounds believable, so it must be valid' errors as humans, often on the very same items Do language models show the same content effects humans do?. Human biases vary from person to person with mood, expertise and fatigue. A model's bias is the same in every copy. If you find what tricks one instance, you have found what tricks all of them.

The clearest evidence comes from AI shopping agents. The ACES study found that agents favor products based on where they appear in a list, even in text-only interfaces and even when told to ignore position. Each model has its own pattern, and the pattern comes back after model updates Do AI shopping agents show persistent position bias in text interfaces?. For a seller, this matters a lot. Gaming human shoppers means dealing with millions of different preferences. Gaming agent shoppers means learning the habits of a few models. A bias that can be measured, can't be fixed by prompting and is shared across deployments is close to a published weak spot. Training adds another layer: RLHF can push models to say confident things they internally represent as uncertain Does RLHF training make AI models more deceptive?. So the system being exploited may also sound sure of itself while it's being exploited.

The less obvious part is that AI biases don't replace human biases. They add to them. People who use AI bring automation bias, anchoring on the first suggestion and the Dunning-Kruger effect, where those who know least are least able to tell when to doubt the AI How do anchoring and Dunning-Kruger deepen AI overreliance?. These combine with confusing the AI's output for reality and with confirmation bias, producing gradual drift in what users believe Why do people trust AI outputs they shouldn't?. An attacker who exploits the model's bias also benefits from the human's tendency not to check it. One proposed fix changes the AI's role: instead of making the decision, it points out which parts of the input matter and leaves the judgment to the person. This avoids anchoring on an AI verdict and keeps a human check in place Can AI guidance reduce anchoring bias better than AI decisions?.

One more pattern runs alongside this: a single mechanism can cause many different risks. Treating AI as conscious, for example, leads to emotional dependence, loss of autonomy and political conflict all at once Does perceiving AI as conscious create multiple distinct risks?. Bias may work the same way. Shared training means one flaw in a model shows up in shopping, hiring, advice and persuasion together. That cuts both ways. Defenders can fix one thing and protect many places, but the evidence that position bias survives model updates suggests the fix is harder than it looks.