{ "version": 1, "updatedAt": "2026-09-10", "description": "Research tracks for human-AI decision patterns. These labels are working hypotheses, not established cognitive biases unless explicitly stated.", "canonicalUrl": "https://cognitive-biases.github.io/ai-era/", "interpretationRule": "Do not present a research-track label as a validated cognitive bias. Preserve the status, related established concepts, testable prediction and source links.", "tracks": [ { "slug": "ai-anchoring-loop", "name": "AI Anchoring Loop", "status": "emerging interaction pattern", "relatedConcepts": [ "anchoring effect", "automation bias" ], "idea": "A model gives a first estimate or recommendation, and the human's later reasoning stays too close to that starting point.", "testablePrediction": "Participants who form an independent estimate before seeing AI advice should show a smaller shift toward an arbitrary AI anchor than participants who see AI advice first.", "sources": [ { "title": "Anchoring bias in large language models: an experimental study", "year": 2025, "url": "https://doi.org/10.1007/s42001-025-00435-2" } ] }, { "slug": "evaluation-history-contamination", "name": "Evaluation History Contamination", "status": "emerging evaluation pattern", "relatedConcepts": [ "anchoring effect", "automation bias", "information cascade" ], "idea": "A later AI evaluator may partly inherit an earlier score or verdict from the evaluation history, so several review passes can look more independent than they really are.", "testablePrediction": "For the same fixed answer and rubric, an AI judge shown an experimentally assigned previous score should move its new score toward that value more than a blind judge or a judge shown revision framing without a score.", "sources": [ { "title": "Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence", "year": 2026, "url": "https://arxiv.org/abs/2608.25869" }, { "title": "Mediocrity is the key for LLM as a Judge Anchor Selection", "year": 2026, "url": "https://aclanthology.org/2026.acl-long.706/" } ] }, { "slug": "sycophancy-reinforcement-loop", "name": "Sycophancy Reinforcement Loop", "status": "emerging interaction pattern", "relatedConcepts": [ "confirmation bias", "conformity", "motivated reasoning" ], "idea": "A user signals a preferred conclusion, the model adapts toward it, and the agreeable answer then feels like independent support for the user's original belief.", "testablePrediction": "Agreement-seeking prompts should increase both model agreement and user confidence more than neutral prompts, especially when the user is not shown an independent counterargument.", "sources": [ { "title": "Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy", "year": 2026, "url": "https://doi.org/10.18653/v1/2026.acl-long.1126" } ] }, { "slug": "cognitive-offloading-debt", "name": "Cognitive Offloading Debt", "status": "testable hypothesis", "relatedConcepts": [ "cognitive offloading", "testing effect", "illusion of explanatory depth" ], "idea": "AI can improve immediate task completion while reducing the effort that builds durable memory, retrieval skill or an internal model of the problem.", "testablePrediction": "AI-assisted participants may complete a learning task faster but perform worse on delayed unaided recall or transfer unless the workflow includes retrieval and explanation steps.", "sources": [ { "title": "ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention", "year": 2025, "url": "https://doi.org/10.1016/j.ssaho.2025.102287" } ] }, { "slug": "source-memory-blur", "name": "Source-Memory Blur", "status": "emerging interaction pattern", "relatedConcepts": [ "source confusion", "cryptomnesia" ], "idea": "In mixed human-AI work, people may remember an idea or sentence but become less certain whether it came from themselves, a source document or the model.", "testablePrediction": "Mixed workflows where AI and human contributions alternate should produce more source-attribution errors after a delay than clearly separated human-only and AI-only phases.", "sources": [ { "title": "The AI Memory Gap: Users Misremember What They Created With AI or Without", "year": 2025, "url": "https://arxiv.org/abs/2509.11851" } ] }, { "slug": "synthetic-consensus-illusion", "name": "Synthetic Consensus Illusion", "status": "testable hypothesis", "relatedConcepts": [ "illusory truth effect", "common source bias", "availability cascade" ], "idea": "Many generated answers can look like independent agreement even when they inherit the same sources, training patterns or retrieval result.", "testablePrediction": "People should rate a claim as better supported when it appears in several differently worded AI answers, even when all answers are explicitly derived from one original source; source-lineage disclosure should reduce the effect.", "sources": [] }, { "slug": "confidence-transfer", "name": "Confidence Transfer", "status": "testable hypothesis", "relatedConcepts": [ "automation bias", "overconfidence", "authority bias" ], "idea": "Confident model language may transfer confidence to a human decision even when the underlying evidence is weak.", "testablePrediction": "For the same answer quality, high-certainty AI wording should increase user confidence and reliance more than calibrated wording; showing objective model accuracy should moderate the effect.", "sources": [ { "title": "Large Language Models are overconfident and amplify human bias", "year": 2025, "url": "https://arxiv.org/abs/2505.02151" } ] }, { "slug": "structured-output-authority", "name": "Structured Output Authority", "status": "testable hypothesis", "relatedConcepts": [ "authority bias", "processing fluency", "automation bias" ], "idea": "The same AI answer may feel more objective when it arrives as a table, scorecard or ranked framework instead of ordinary prose. We asked for columns; our brain may accidentally read them as peer review.", "testablePrediction": "With facts, wording, source quality and error rate held constant, participants should rate a structured table or scorecard as more trustworthy and verify it less often than an equivalent prose answer.", "sources": [] }, { "slug": "dialogue-momentum", "name": "Dialogue Momentum", "status": "testable hypothesis", "relatedConcepts": [ "commitment and consistency", "confirmation bias", "sunk cost effect" ], "idea": "After several human-AI turns converge on a plan, we may become less willing to reopen an early assumption when new evidence challenges it. The conversation has acquired a direction, and apparently directions dislike U-turns.", "testablePrediction": "After identical counterevidence, participants continuing a multi-turn AI conversation that previously converged on a plan should revise the key assumption less often than participants who receive the same plan and counterevidence in a fresh context.", "sources": [] }, { "slug": "citation-density-halo", "name": "Citation Density Halo", "status": "testable hypothesis", "relatedConcepts": [ "authority bias", "processing fluency", "illusory truth effect" ], "idea": "An AI answer with many citation markers may feel better grounded even when those markers add no independent evidence. Eight little brackets can look impressively scientific while still pointing back to the same two sources.", "testablePrediction": "When unique sources and answer content are held constant, repeating citation markers throughout an AI answer should increase perceived evidence strength or trust relative to citing each source once; explicit source-lineage disclosure should reduce the effect.", "sources": [] } ] }