Skip to content

Output Vacuity and Central-Tendency Bias

A failure mode distinct from hallucination that deserves separate treatment because it evades detection mechanisms that catch factual errors. Where hallucination produces confident claims that are wrong, output vacuity produces confident prose that is technically unobjectionable but operationally empty. The term "slop" has entered common usage to describe this class of AI-generated content. Kommers et al. (2026) characterise slop through three prototypical properties: superficial competence (a veneer of quality belied by lack of substance), asymmetric effort (near-zero generation cost), and mass producibility (designed for digital-scale distribution). These properties explain why slop proliferates in professional settings and why it is particularly insidious: when generating a plausible-sounding assessment costs almost nothing, the incentive to evaluate whether it actually says anything diminishes.

The Mechanism

The architectural cause is central-tendency bias. LLMs are trained on statistical distributions of text and generate output biased toward high-frequency patterns. Kandpal et al. (2023) demonstrated that a model's ability to produce accurate responses correlates directly with how frequently the relevant knowledge appeared during training.

When a model encounters a query that falls in the long tail of its training distribution, it compensates not by expressing uncertainty (a calibration-and-confidence problem) but by retreating to generality: listing all plausible causes rather than discriminating among them, restating data in prose rather than analysing it, or hedging so broadly that no actionable conclusion can be drawn.

This retreat to the mode of the distribution is not a failure of reasoning but a structural property of maximum-likelihood generation against a broad training corpus.

Interaction with Task Framing

The same model produces highly specific output when the task is narrow and well-constrained ("what is the regulatory limit for this parameter?") and retreats to generic output when the task is broad or ambiguous ("assess the current situation"). System prompt design and output format requirements are the primary levers for reducing slop: a prompt that demands specific parameter values, explicit evidence citations, and discrimination between hypotheses will elicit less vacuous output than one requesting "diagnostic support."

This means slop is not solely a model property but a joint property of the model, the prompt, and the task.

Interaction with Context Length

Slop probability increases with context-windows utilisation. As sessions lengthen and accumulated tool results, conversation history, and retrieved documents fill the context, the model's attention spreads across more tokens and the effective weight of any single piece of evidence diminishes. Output quality degrades toward vacuity precisely as the operational situation becomes more complex and the need for specific analysis is greatest.

Why It's Dangerous

The danger of slop is indirect but significant, operating through a different trust pathway than automation-bias:

  • A factually wrong output can be caught by a knowledge-graphs guardrail, an alert operator, or a cross-checking agent
  • A vacuous output passes all of these checks because it makes no specific claim that could be falsified

The indirect harm emerges over time through a cry-wolf dynamic: operators who receive consistently generic assessments reduce the attention they allocate to AI output. When the system produces a genuinely insightful assessment — or more dangerously, a genuinely wrong one — the operator has learned to skim rather than engage critically.

This disengagement pathway is distinct from automation bias (where operators over-trust specific wrong recommendations). Slop-induced disengagement involves ceasing to evaluate recommendations at all.

Measuring Slop

Shaib et al. (2025) develop a formal taxonomy for measuring slop across three themes: information utility (density and relevance), information quality (factuality and bias), and style quality (structure, coherence, and tone). Their key findings are sobering for automated quality assurance:

  • Standard text metrics fail to capture slop — three of the five most predictive features (relevance, coherence, tone) lack adequate automatic measurements
  • LLMs cannot reliably self-identify slop — GPT-5, DeepSeek-V3, and o3-mini achieve near-zero agreement with human annotators, under-predicting slop by a factor of 4–10x
  • Slop features vary by domain — factuality and structure dominate in QA tasks, while style and utility issues dominate in news articles

This domain-dependence means that slop detection for safety-critical advisory output requires domain-specific evaluation criteria, not generic quality metrics.

Mitigation

  • Structured output templates that require specific parameter values, specific procedure steps, and specific time frames
  • Novelty filters that suppress AI output unless it adds information beyond what qualified displays already show
  • Specificity metrics that measure whether the model discriminates between hypotheses or retreats to generic enumeration — aligned with the information utility theme (density and relevance) in the slop taxonomy
  • Operator training that includes exposure to slop as a recognisable failure pattern, with the heuristic: if the AI assessment would apply equally well to three different conditions, it is not adding value
  • Human-in-the-loop evaluation rather than LLM-as-judge for quality assessment, given demonstrated LLM blind spots for slop detection

Relevance to Safety-Critical Systems

  1. Slop evades quality checks. Unlike hallucination, vacuous output cannot be caught by fact-checking. It is technically correct — just operationally useless.

  2. It erodes trust over time. Consistent vacuity trains operators to ignore AI output entirely, removing the safety benefit of the advisory system.

  3. It worsens under complexity. The conditions where specific analysis matters most (long sessions, complex situations, multiple information sources) are precisely the conditions where slop is most likely.

  4. Cross-domain examples. In aviation, a generic weather assessment that lists all possible conditions without highlighting the specific risk for the planned route. In medical settings, a differential diagnosis that lists every textbook possibility without discriminating based on the patient's specific presentation. In oil and gas, a well integrity assessment that restates known parameters without identifying the anomaly that triggered the query.