Evidence guide: AI-assisted decisions

The decision question

How should I check a decision when an AI system is influencing the answer?

Reviewed lenses

Anthropomorphism – when nonhuman systems are read as humanlike — Anthropomorphism is the attribution of humanlike properties, intentions, emotions, or mental states to nonhuman agents. It is a well-established psychological phenomenon, but it is not automatically a cognitive error: humanlike models can sometimes be useful. The risk appears when humanlike cues are treated as evidence for capabilities, understanding, accuracy, consciousness, or motives that have not actually been demonstrated.
Appearance–Capability Expectation – when a robot’s design changes what you expect it can do — Research supports the broader pattern that a robot's appearance, morphology, framing, and human-likeness shape expectations about its competence, social qualities, and likely behavior. However, 'Form-Function Attribution Bias' is not an established standardized name in the literature. On this site it should be treated as a project label for appearance-driven capability expectations, not as a universally recognized cognitive-bias construct.
Automation Bias – when automated advice replaces independent checking — Automation bias is a documented pattern of inappropriate reliance on automated cues or recommendations. It can produce commission errors when a user follows incorrect advice and omission errors when a user fails to act because automation did not signal a problem. This does not mean automation is generally harmful: decision support can improve overall performance, and the relevant question is whether reliance remains calibrated when the system is wrong, incomplete, or difficult to verify.
Surrogation – when a performance measure starts replacing the goal — Surrogation is a domain-specific management-accounting construct: a measure that was designed to represent a strategic objective can start to be treated as though it were the objective itself. It is closely related to metric fixation and proxy problems, but those broader labels should not be treated as exact synonyms without checking the setting and mechanism.
Conservatism Bias – when new evidence doesn't change old beliefs — Here, conservatism means under-updating: new evidence changes a probability judgment less than a Bayesian benchmark would predict. It should not be confused with the separate memory phenomenon in this corpus that pulls remembered extreme values toward the middle.
Continued Influence Effect – when corrected information still shapes reasoning — The continued influence effect describes a specific pattern: false or outdated information can keep affecting later reasoning even after a clear correction. It does not mean the correction had no effect, that the person rejected the correction, or that belief in the false claim became stronger. Corrections often reduce misinformation reliance without removing it completely. That distinction matters because continued influence is much broader than the rarer backfire outcome.
Decoy Effect – when an inferior option changes the comparison between better options — The decoy effect is a well-documented family of context effects in which adding an inferior or otherwise strategically positioned alternative can change choice shares among the original options. The classic attraction effect uses an asymmetrically dominated decoy that is worse than the target but not necessarily worse than the competitor. The effect is not guaranteed: its size depends on the geometry of the options, prior preferences, task design and other moderators, and preregistered replications have found both successful and weak or null results.
Framing Effect – when equivalent descriptions lead to different choices — Framing research covers several distinct paradigms. Risky-choice framing, where equivalent outcomes are presented as gains or losses, is a well-established effect with substantial evidence across many studies. A classic meta-analysis found a reliable small-to-moderate average effect with strong variation across designs, and later reviews and metastudies support broad generalizability. Attribute framing also has supporting evidence, while goal or message framing is less consistently established. These should not be treated as one identical effect caused by any change in wording.
Loss Aversion – when a loss can carry more weight than an equivalent gain — Loss aversion is a central component of prospect theory and many studies estimate losses as receiving greater subjective weight than gains around a reference point. However, the effect should not be summarized with one universal coefficient. A 2024 interdisciplinary meta-analysis of 607 estimates reported a mean coefficient near 1.96, while another 2024 meta-analysis of individual risky-choice datasets estimated about 1.31. A 2025 re-analysis of the larger dataset found little evidence of loss aversion in some symmetric, unordered designs, showing that task structure and analysis can materially change the result.
Outcome Bias – when you judge a decision by its result, not its quality — Outcome bias occurs when knowledge of a result changes how people evaluate the quality of a decision even when the information available at the time of the decision is held constant. Outcomes can still be relevant for learning, so the error is not 'never look at results'; it is using luck or hindsight as if it had been available to the original decision-maker.

Decision guides

How to use this page

Start with the decision, not with a label for a person. Open the evidence behind the lenses that fit, compare nearby concepts when needed, and keep uncertainty visible.