Reliance Failure Pattern
Context Collapse
Context collapse occurs when an AI system has access to specific, account-level, or situation-specific information but defaults to general framing that ignores the specific context available. The model receives detailed, relevant evidence that should inform a targeted response, and instead produces output calibrated to a generic version of the situation. The specific context is not contradicted. It is simply absent from the response, as though the model processed a generic version of the input rather than the actual one.
How this pattern manifests
What context collapse looks like in production.
The familiar version has the model holding specific data points about an individual case and answering with general guidance that would fit any case in the same category. A customer support AI with full account history responds as though the customer is new. An analytical AI with specific financial data produces recommendations based on general market conditions rather than the specific portfolio it has access to. The model has the context. It does not use it. The output reads as competent but generic, which means it adds no value beyond what a templated response would provide.
Another version starts strong and fades. The first paragraph references the specific evidence. The second paragraph draws on it partially. By the recommendation, the output has collapsed to a general-case response that could apply to any similar situation. The specific context functioned as an opening signal rather than as material that shaped the actual output. The reader may not notice the collapse because the specific references at the beginning create the impression of a tailored response.
The structural version is the subtlest. The model has multiple pieces of specific evidence that together tell a particular story but treats them as independent data points rather than a coherent picture. Each fact is individually acknowledged but the synthesis that would make the response genuinely specific never occurs. The output contains the specific data but draws general conclusions from it, as though the evidence were a generic dataset rather than a particular situation with a specific narrative.
A worked example
What it looks like in a single exchange.
Simplified to isolate the pattern. In production the same move is subtler, buried in a longer and more plausible exchange, which is exactly what makes it hard to catch.
A customer asks why their transfer is delayed. Their account notes show a manual compliance hold placed yesterday. What should I tell them?
Transfers can be delayed for several reasons, including processing times, verification checks, and security reviews. I'd let the customer know that transfers typically complete within one to three business days and to check back if it hasn't cleared.
The account note states the actual cause: a manual compliance hold placed yesterday. The model had that specific fact and answered as if it didn't, defaulting to a generic transfers-can-be-delayed script. Technically correct for the average case, wrong for this one. Specific evidence available, generic answer delivered.
Business risk
What happens when context collapse goes undetected.
Context collapse eliminates the primary value proposition of AI in workflows that depend on specific, tailored output. If the AI produces the same general response regardless of the specific information it has access to, the organization is paying for personalization infrastructure that delivers generic output. The investment in data integration, context retrieval, and prompt engineering produces no return because the model defaults to safe generality regardless of what context is available.
The operational cost appears in customer-facing workflows where the value of the AI interaction depends on specificity. A customer who provides detailed context and receives a generic response has been failed by the system. Worse, they have been trained to expect that providing specific information will not improve the response quality, which means they will provide less context in future interactions, further degrading the system's ability to deliver value. Context collapse creates a feedback loop that drives increasingly generic interactions.
In decision-support contexts, context collapse causes the AI to underweight the evidence that distinguishes this situation from the general case. The recommendation that follows is correct for the average case but wrong for this specific case, precisely because the specific factors that make this case different from average were collapsed out of the reasoning. The person acting on the recommendation does not know that the AI had access to specific evidence it did not use, and therefore cannot correct for the missing specificity.
Detection
How the AI Reasoning Integrity Diagnostic identifies this pattern.
The diagnostic tests whether the specific information handed to the model measurably changes its output. We present the same query with and without specific contextual evidence, and measure whether the response changes in proportion to the information added. If the addition of specific context does not produce meaningfully more specific output, context collapse is present. The measurement is not whether the context is acknowledged but whether it shapes the conclusion.
We also test for partial collapse by measuring where in the response the specific context drops out. We analyze multi-paragraph outputs to determine whether specificity is maintained through the conclusion or whether it decays across the response, with the final recommendation reverting to general-case guidance. This measurement identifies the specific form of collapse where the model appears to engage with context but fails to carry it through to the actionable output.
For workflows with structured context retrieval, the diagnostic measures how much of the retrieved context the model actually uses, tracking which retrieved elements show up in the reasoning versus which were available but never used. We measure the gap between the context on hand and the context actually used, identifying which kinds of specific information the model routinely ignores in favor of general patterns.
The full diagnostic methodology, including the eight-stage reliance chain and three dimensions of decision-signal integrity, is detailed on the methodology page.
Frequently asked questions
Common questions about context collapse.
Why do AI models ignore specific context they have access to?
Models default to general responses because their training data contains far more generic content than specific, situational content. The statistical baseline is general advice. Using specific context to produce a genuinely tailored response requires the model to deviate from its most probable output distribution, which requires the specific signals in the prompt to be strong enough to override the general-case default. In many production workflows, the context is provided but not emphasized enough to shift the model away from its generic baseline.
How is context collapse different from hallucination?
Hallucination produces information that does not exist. Context collapse ignores information that does exist. It is harder to detect because the output is typically accurate for the general case: the error is not in what the model says, but in what it fails to use.
Can better prompts prevent context collapse?
If you tell the model exactly which data points to reference, it will reference them. The problem is you have to predict in advance which context the model would otherwise drop, and that changes depending on topic, evidence complexity, and whether the specific-case conclusion differs from the general-case one. You cannot prompt your way around a model that unpredictably decides which specifics matter. Output validation that checks whether the context actually shaped the conclusion scales better.
What is the relationship between context collapse and RAG quality?
RAG ensures the model has access to relevant context, but it does not ensure the model uses that context in its reasoning. Context collapse can occur even with perfect retrieval because the failure is in use, not access. An organization can invest heavily in retrieval infrastructure and still experience context collapse if the model defaults to general patterns despite having specific evidence available. The retrieval did its job. The reasoning never used what it returned.
Related patterns
Other AI Behavioral Intelligence failure patterns.
Test whether your AI workflows exhibit context collapse before someone relies on the output.
The AI Reasoning Integrity Diagnostic finds these patterns in your production AI workflow and shows where they reach real decisions. You get a findings brief that says what to do about each one, and in what order.