S6.06 - Zero-shot_prompting

S6.06 ? Zero-shot prompting

flowchart LR
    A[Direct instruction to model
No examples provided] --> B[RAIDT
Run-level evidence framework]
    H[Low setup cost
Fast exploration
But potentially fragile] --> B
    I[Public services
Enterprise productivity
Healthcare support
Compliance screening] --> C
    B --> C[[Zero-shot prompting
Minimal steering documented at run level]]
    C --> D[Run-level evidence pack]
    C --> E[RAIDT score profile]
    C --> F[Reviewer reconstruction
Contestability
Audit readiness]
    D --> G[Governance readiness
Organisational learning]
    E --> G
    F --> G

? Star S6 - Influence Methods as Governance Interventions

Star context: Positions prompting, RAG, PEFT/LoRA, RLHF/DPO and stacked influence as components that shape governance evidence rather than replace RAIDT?s core concern with run-level reviewability and governance readiness.

Academic picture

Definition / background

Zero-shot prompting is the use of a generative AI system without worked examples in the prompt. The user provides an instruction, question, or task framing and relies on the model?s prior training and general instruction-following capability to produce a response. Conceptually, it sits at the lighter end of the influence-method spectrum: it shapes behaviour through wording alone rather than through example sets, retrieval augmentation, adapter tuning, or post-training preference optimisation.

In governance terms, zero-shot prompting matters because it is often the first and most accessible way organisations use large language models. It is cheap, quick to deploy, and easy to test. However, its simplicity can obscure important questions about consistency, justification, and reproducibility. If a model output changes materially when the task wording is altered slightly, then the organisation needs evidence not only of the output but also of the exact run conditions under which that output was produced.

Zero-shot prompting differs from baseline prompting and structured prompting in degree and discipline. Baseline prompting may simply mean an initial prompt used for comparison, whereas zero-shot specifically indicates the absence of examples. Structured prompting adds clearer constraints, decomposition, or formatting. Role-based prompting adds an interpretive frame. Zero-shot prompting can overlap with these, but the defining feature remains that the model is asked to perform without demonstrations.

Within RAIDT, zero-shot prompting belongs inside the project because it influences the run while also exposing governance limits. RAIDT treats the run as the unit of governance, so a zero-shot prompt is not merely a piece of wording; it is part of the configuration that must be evidenced if the output is to support responsible organisational use. It therefore has direct implications for the evidence pack, the five-pillar score profile, and the broader move from principle-based assurance to reviewable practice.

Why this concept matters

Zero-shot prompting matters because many real organisational deployments begin with it. Teams often experiment with a model by asking it to summarise, classify, draft, or explain without first building a richer control layer. That makes zero-shot prompting a realistic entry point into governance, but also a source of avoidable confusion. Without clear recognition of what zero-shot prompting can and cannot support, organisations may over-interpret fluent outputs as reliable evidence.

The concept helps avoid a common governance mistake: assuming that a simple prompt is a neutral or trivial input. In practice, prompt wording is itself an intervention. It frames the task, influences the model?s output space, and shapes how downstream reviewers judge adequacy. If that intervention is left undocumented, then later challenge, audit, or incident review becomes much harder.

For RAIDT, the importance of zero-shot prompting is therefore operational. It provides a baseline influence method against which stronger governance mechanisms can be compared. It also marks the point at which a run may be efficient enough for exploratory use yet insufficiently evidenced for higher-stakes decisions. By making that distinction explicit, RAIDT helps organisations decide when zero-shot prompting is acceptable, when it requires safeguards, and when it should be replaced by more structured methods.

Key idea: Zero-shot prompting matters because it is often the first influence method used in practice, but in RAIDT it becomes governable only when its exact run-level conditions and limitations are evidenced.

What this item controls

It controls task framing by defining what the model is being asked to do without demonstrations.
It controls the initial behavioural steering applied to a run before stronger interventions such as RAG, PEFT, or preference optimisation are added.
It controls the degree of explicitness in instructions, constraints, and output format.
It controls how much the organisation relies on the model?s latent capabilities rather than on externally supplied examples or retrieved evidence.
It controls the quality of any baseline comparison used to justify later prompt engineering or governance enhancements.
It controls, albeit imperfectly, the interpretive boundary between exploratory assistance and dependable operational use.

Practical example / likely audience question

Audience question

When is zero-shot prompting actually useful in a governed AI setting?

Answer

The concern behind this question is usually that zero-shot prompting appears too weak or informal to belong in a serious governance framework. The direct answer is that it is useful when the organisation needs a low-cost starting point, an exploratory baseline, or a comparator against which stronger interventions can later be justified. It is less suitable when the task requires high consistency, strong provenance, or defensible repeatability.

A practical example is an enterprise team testing whether a model can draft meeting summaries from internal notes. A zero-shot prompt may be entirely adequate for early exploration, especially if the output is reviewed by a human and not treated as final evidence. However, if the same organisation later uses model summaries in compliance-sensitive workflows, zero-shot prompting on its own becomes difficult to defend because slight wording changes may alter omissions, tone, or inferred priorities.

RAIDT handles this issue better than a generic AI governance approach because it does not stop at saying zero-shot prompting is risky or acceptable in principle. It asks what happened in this run, with this prompt, for this task, at this time, under this review process. That move from abstract guidance to run-level evidence is what makes the governance judgement more credible.

Practical example in RAIDT terms

Consider a public-service organisation using a language model to draft first-pass responses to citizen enquiries about housing support. In an early pilot, staff use a zero-shot prompt such as: provide a concise draft reply based on the enquiry text and maintain a neutral administrative tone. The model is not given examples, retrieved policy passages, or domain-specific tuning.

The run-level issue is that the output may sound authoritative while omitting a qualifying condition in the policy. To govern this run properly, RAIDT would require evidence of the exact prompt text, model and version, date and time of the run, source input, output produced, human reviewer comments, acceptance or rejection decision, and any observed variation across repeated runs.

The most affected pillars are Dependability, Auditability, and Traceability, with Responsibility and Interpretability also relevant. Dependability is implicated because output consistency may be weak. Auditability and Traceability are implicated because reviewers must be able to reconstruct what instruction produced the draft. Responsibility is implicated because staff remain accountable for whether the draft was used appropriately. Interpretability is implicated because the organisation must explain why the output was considered acceptable despite the absence of examples or retrieval support.

In governance-readiness terms, zero-shot prompting improves readiness only when treated as a documented baseline rather than as an invisible convenience. RAIDT turns a simple prompt into a reviewable artefact that can be accepted for low-risk exploration, restricted in higher-risk cases, or used as evidence for why stronger controls were later introduced.

Detailed link to RAIDT

Zero-shot prompting links to RAIDT in four ways.

First, it links to RAIDT?s core idea because RAIDT governs configured uses of generative AI rather than abstract models in isolation. A zero-shot prompt is one of the clearest examples of a concrete run configuration.

Second, it links to the run because prompt wording, task framing, and absence of examples materially shape the output. The run-level evidence therefore needs to capture the prompt as part of the conditions under which the result emerged.

Third, it links to the evidence pack and score profile because zero-shot prompting affects how well the organisation can justify the run across the five pillars. A zero-shot run may score adequately for a low-stakes exploratory task but poorly for repeatability or contestability in a high-stakes context.

Fourth, it links to reviewability, contestability, audit readiness, and organisational learning because documenting zero-shot runs helps reviewers see where a minimal intervention was sufficient and where it was not. That learning supports method selection, escalation, and policy design.

Zero-shot prompting ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

In RAIDT, the significance of zero-shot prompting is therefore not merely that it influences outputs, but that it can be evidenced, reviewed, compared, and bounded as part of a wider governance process.

Link to the five RAIDT pillars

Responsibility

Zero-shot prompting does not remove human accountability; if anything, it makes role clarity more important because the system is operating with relatively light steering.