Processing bias occurs when human choices influence how raw information becomes training data, prompts, labels, rules or evidence. Each processing step involves judgement: which records to include, what counts as an error, how categories are named and which exceptions are removed. NIST explains that human biases can enter AI through assumptions and decisions across the lifecycle, including data cleaning, modelling and use (NIST, 2022). For practitioners, the key point is that bias is not only a technical issue. It is also a leadership issue because people set the purpose, agree the standards and decide what evidence is good enough.
Conscious bias is a preference or judgement a person recognises. It may be openly stated, such as deciding that one type of customer is more important, or quietly deliberate, such as removing difficult cases to make a trial appear more successful. Conscious bias is not always intended to cause harm, but it can produce one-sided data or unfair rules. A team may, for example, train a complaints classifier mainly on resolved cases because these are easier to access. The result may look accurate while failing on the unresolved complaints that matter most.
Unconscious, or implicit, bias operates through automatic associations and assumptions outside full awareness. These associations are shaped by experience, culture and context and can affect everyday decisions (Greenwald and Banaji, 1995; EHRC, 2018). In AI work, unconscious bias may influence which examples seem ‘normal’, how unclear text is labelled or whose needs are treated as an exception. Awareness training can help practitioners notice the risk, but EHRC’s evidence review warns that training alone is not a quick fix. Stronger controls include clear criteria, varied reviewers, structured checks and analysis of outcomes.
Interpretation bias begins after information (or an AI output) is presented. People rarely examine evidence from a completely neutral position. Cognitive biases are predictable thinking shortcuts that help us deal with complexity but can also create systematic errors. People often use mental shortcuts when making uncertain decisions, such as relying on what feels recent or familiar (Tversky and Kahneman,1974). In the workplace, a recent incident feels more important to making a decision than a larger trend.
How information is presented can also shape interpretation. A percentage without the sample size, a risk score without its threshold or a chart with a shortened axis may lead practitioners towards a stronger conclusion than the evidence supports. AI-generated explanations can create the same problem if they sound certain but do not show limitations. Practitioners should ask: what does this measure mean, what is missing, how is uncertainty shown, and would a different presentation change the decision?
Confirmation bias is the tendency to seek, notice or interpret evidence in ways that support an existing belief (Wason, 1960; Nickerson, 1998). A manager who believes remote workers are less productive may focus on missed messages while overlooking stronger output data. When reviewing AI, the same person may accept recommendations that fit this belief and question conflicting results more aggressively. Practitioners can reduce this risk by stating the hypothesis before analysis, defining what evidence would disprove it, asking for alternative explanations and inviting independent challenge.
The Mandela effect is a shared and confident false memory. Prasad and Bainbridge (2022) found that some familiar images produce the same specific memory error across different people. It does not show that reality has changed; it shows that memory is reconstructive and can be influenced by familiarity, expectation and repeated versions of a story. At work, several colleagues agreeing about a past decision does not make their recollection accurate. Practitioners should check meeting records, approved documents, system logs and dated source material before using remembered information to correct data or challenge an AI result.
AI can intensify interpretation bias because polished outputs encourage trust. Automation bias describes over-reliance on a system’s suggestion, including missing an error because the tool did not flag it (Goddard et al., 2012). Practitioners should therefore separate the output from the decision. They should check source quality, uncertainty, subgroup effects and alternative explanations; compare the result with verified evidence; and record why the final decision was accepted, changed or rejected. The strongest safeguard is not simply adding a human reviewer but designing a review process that makes assumptions visible and challenge routine.
Action Point
Choose one AI-supported decision or analysis in your workplace. Map where people process the information and where they interpret the output. Identify one possible conscious bias, one possible unconscious bias and one cognitive bias. Then define a practical control for each, such as agreed criteria, independent review, disconfirming evidence, source verification or an outcome check across affected groups.