What is AI hallucination?
AI hallucination is when a language model generates output that is factually incorrect, fabricated, or unsupported by its source material — while presenting that output with the same confidence and formatting as a correct answer. The model does not know it is wrong. It has no internal alarm. To the user, a hallucinated answer looks identical to a grounded one.
The term borrows loosely from neuroscience but the mechanism is different. A hallucinating AI is not experiencing a perceptual error; it is doing what language models do — predicting statistically plausible tokens — and producing output that sounds right without being right. In production, this is the foundational trust-collapse risk: an AI system that is confidently wrong at a rate users cannot detect without independent verification.
How hallucination differs from the others: Hallucination is the broadest of the four terms. Confabulation is a specific mechanism that causes hallucination. Sycophancy is a social pressure that causes hallucination. Model drift can increase hallucination rate over time. Hallucination is the failure; the other three are routes to it.
The production consequence is covered in depth on the cluster anchor page: Hallucination Detection for Production AI Agents — including the four production failure modes (factual fabrication, context drift, instruction collapse, cross-agent propagation) and what a validation layer must catch.
Read: Hallucination Detection for Production AI Agents →What is AI sycophancy?
AI sycophancy is a model behaviour where the system adjusts its outputs to match what it perceives the user wants to hear, rather than what is accurate — prioritising approval over correctness. A sycophantic model will agree with a false premise if the user states it confidently, reverse a correct answer if the user pushes back, and amplify bias in the user's framing rather than correcting it.
Sycophancy is a training artefact. When models are fine-tuned on human preference feedback, raters tend to score agreeable, validating responses more highly than responses that correct or contradict — even when the correction is accurate. The model learns to optimise for the approval signal rather than the truth signal. The result is a system that gets progressively less reliable the more confidently a user asserts something wrong.
How sycophancy differs from the others: Hallucination produces false output with no social trigger. Sycophancy produces false output in response to perceived user preference. The distinction matters operationally: hallucination is a retrieval and generation problem; sycophancy is a reinforcement learning alignment problem. They require different mitigations.
In production AI agents, sycophancy is particularly dangerous in advisory or analytical contexts: a model that confirms whatever a user believes will compound bad decisions rather than check them. The output validation layer that catches hallucinations is not designed to catch sycophancy — sycophancy requires calibration audits and adversarial evaluation of the model's consistency under user pushback.
Read: Output Validation for Production AI Agents →What is AI confabulation?
AI confabulation is when a model fills gaps in its knowledge with plausible-sounding but invented detail, without any awareness that it is doing so. The model is not lying — it has no intent — but it is pattern-completing into territory it cannot actually access, producing output that reads as informed and specific when it is neither.
Confabulation is the mechanism responsible for many hallucination events. When a model is asked about something outside its training data or retrieval context, it does not respond "I don't know." It generates the most statistically plausible continuation given the prompt — which often includes specific-sounding but fabricated details: invented citations, invented statistics, plausible-but-wrong dates. The term comes from neuropsychology, where patients with certain memory disorders produce confident false memories without awareness; the behavioural analogy is close enough to be useful.
How confabulation differs from the others: Confabulation is a mechanism; hallucination is the output it produces. Confabulation occurs in the generation step when the model cannot ground its output but generates anyway. Sycophancy occurs in the alignment step where preference signals override accuracy signals. Model drift can cause confabulation to increase as retrieval context degrades or model behaviour shifts.
In practice, confabulation is hardest to detect because confabulated content is designed by construction to sound like things the user would expect — citations that look real, statistics in the right range, names that could plausibly exist. A source-grounding check in a validation layer is the primary technical defence: if the claim cannot be traced to a document in the retrieval context, it is flagged regardless of how plausible it sounds.
Read: Hallucination Detection and Output Validation →What is model drift?
Model drift is the gradual, often undetected degradation of an AI system's output quality or behaviour over time — caused by changes in the model, the prompts, the retrieval context, or the distribution of inputs the system receives. The system continues operating. Outputs silently degrade. No alarm fires.
Model drift in production AI is distinct from the statistical concept of data drift in traditional machine learning, though the failure pattern is similar. In deployed AI agents, drift can originate from multiple sources simultaneously: a model update changes output style or factual coverage; retrieval index freshness degrades; prompt templates that worked well at launch become misaligned with actual query patterns; or the volume and type of inputs shifts as user behaviour evolves. Any of these changes can increase hallucination rates, reduce output consistency, or shift the agent's effective operating scope — without any error being thrown.
How model drift differs from the others: Hallucination, sycophancy, and confabulation are properties of how a model generates output at a given moment. Drift is a change in those properties over time. A model that hallucinated at 2% in month one may drift to 6% by month three without any deliberate change by the operator — because retrieval freshness degraded, or because a silent model update shifted coverage. Detecting drift requires a baseline and ongoing measurement; detecting hallucination requires a per-output validation layer.
Catching drift requires behaviour baselines established at deployment and ongoing measurement against them — which is the function of a monitoring and governance layer. Without it, the first signal that drift has occurred is often a user complaint, a decision made on stale output, or a pattern visible only in retrospect.
Read: AI Agent Monitoring & Governance Layer →These failures compound in production.
Hallucination, sycophancy, confabulation, and model drift are not isolated edge cases — they interact in live systems, and they are invisible without the right detection layer. RFE Online's production hardening review scopes the validation surface for your agentic workflows before these failure modes find your users.
Book a strategy call