MODULE 2 · LESSON 2
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolConsequences of Probabilistic Generation
If an LLM predicts rather than knows, several practical consequences follow directly. The syllabus treats them as things a Requirements Engineer must be able to recall and act on, because each one changes how you review output.
The model is not deterministic by default
Ask the same question twice and you may get different answers. Even a single word changed in your prompt can nudge the model onto a different path and produce a very different result. This sensitivity means the model can be sidetracked by one stray term or example and drift away from what you wanted.
It is worth adding what the syllabus adds: multiple different answers can all be correct and valid responses to the same prompt. Variation is not by itself a defect.
The mechanism behind the variation is sampling. When the model writes, it does not always pick the single most likely next token; it samples from a range of likely options. The setting that controls this randomness is temperature. Low temperature makes the model stick to safer, more obvious choices, which tends to produce more consistent answers. Higher temperature lets it explore and vary more, which is useful for brainstorming but less stable.
For RE work the guidance follows the task. Lower temperatures suit consistency in requirements wording. Higher values can help when you are deliberately generating diverse solution ideas. One caveat the syllabus flags and candidates routinely forget: in many LLMs, temperature cannot be set by the user at all. Knowing what it does is examinable; assuming you can always change it is wrong.
The surface layer and the factual layer
This is the most useful review heuristic in the whole unit. When you read model output, separate two layers.
The surface layer covers fluency, tone and structure. It is usually very strong, because these features are common in training data and easy to reproduce.
The factual layer covers dates, quantities, names and edge cases. This is where pure pattern matching fails. If the training data is thin or inconsistent on a point, the most likely continuation may drift away from the truth.
Connecting the model to search engines, databases or function calls improves the factual layer, because the next-token prediction is then grounded on fresh or computed information. The model still predicts — it just predicts with better inputs.
Hallucination
A hallucination occurs when a model produces statistically likely text that is factually incorrect or unsupported by source data. The syllabus is precise about its nature: hallucination is a direct consequence of the probabilistic mechanism. It is not a bug that a future patch removes.
The model has no internal concept of "knowing", so it cannot signal when it is guessing or when it lacks sufficient data. There is no confidence indicator built into the mechanism, and asking a model how confident it is produces another prediction rather than a measurement.
In Requirements Engineering, hallucination is particularly problematic when fabricated details are mistaken for stakeholder statements or domain facts. A requirement attributed to a stakeholder who never said it, or a constraint justified by a standard that does not exist, enters the baseline looking exactly like a real one.
The syllabus ties this to CPRE principle 6: non-validated requirements are useless. All AI-generated output must be treated as an unverified draft.
The review obligation
Putting it together gives the rule the syllabus wants you to leave with. Regardless of variability, always treat AI output as a draft that needs review, and review it for two distinct purposes:
- To identify and rule out hallucinations — factually incorrect information presented as true.
- To optimise the result for your specific context and requirements.
The second purpose matters as much as the first and is easier to skip. The model is not reliable in the way a calculator is: even when it produces consistent output, that output may still contain errors or may simply not match your needs.
The syllabus's framing of risk here is elegant and worth remembering for the exam. If a model gives an incorrect number or cites the wrong standard, it is not failing at logic — it is succeeding at prediction in the wrong place. The remedy is verification and grounding: provide the right context in the prompt, connect the model to retrieval or calculation tools where precision matters, and treat the output as a draft.
Non-determinism is not only a review problem. When the AI system is the product, it is a design problem, and the way it is usually solved is instructive.
In Elily, the WhatsApp assistant Srileo builds, the questions that must never vary are deliberately routed away from the model. Whether a 3pm slot is free is answered by a lookup, not a prediction. What a service costs is retrieved from that business's own knowledge base, not generated. What is left for the language model is the part where variation is harmless or actively good: understanding what the customer meant, and phrasing the reply.
The general pattern is worth naming, because it applies to any AI-supported system a Requirements Engineer specifies: separate the parts of the task where variation is acceptable from the parts where it is not, and only send the first kind to a probabilistic system. Getting that boundary wrong is the most common architectural mistake in AI features, and it usually shows up as a requirement that quietly assumes the model will answer the same way every time.
An AI-drafted requirements document cites "ISO 20258p" as the applicable security guideline. No such standard exists. How should this be characterised?
Temperature
Click to flipA parameter controlling the randomness of an LLM's output. Low values make the model more deterministic and consistent in phrasing; higher values produce more varied, creative results. In many LLMs it cannot be set by the user.
Click to flip backBecause generation is probabilistic, the same prompt can yield different answers, temperature controls how much, and hallucination is a property of the mechanism rather than a defect. Review model output in two layers: trust the surface, verify the facts. Every AI-generated artefact is an unverified draft until a human validates it — which is CPRE principle 6 restated for a new tool.