King's College London · FACULTY OF MANAGEMENT

6QQMN331 Chap.1 AI, Generative Models and Responsible Use

- one subject, every graph, every model, every mark
6 Chapters5-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 1 of 8 · 6QQMN331

AI, Generative Models and Responsible Use

Classify before evaluating

This chapter separates artificial intelligence from generative AI and distinguishes pre-training, fine-tuning, retrieval and workplace use. The aim is to stop broad labels from carrying an argument.

You will trace an output into an organisational decision, locate the responsible human role and use fairness, accountability, transparency and ethics to turn concerns into controls.

Build the smallest defensible claim

Worked cases show why current retrieval does not guarantee faithful generation and why technical performance cannot establish organisational value by itself.

The chapter closes with original practice on classification, evidence, review boundaries and accountable ownership.

Follow the complete evidence chain

A responsible analysis asks what shaped the model, which information enters during use, how the output is checked and where it changes organisational action.

Keeping these stages separate prevents a control for one stage from being mistaken for a solution to another.

Current sources can improve the information available to a generator, for example, without proving that the generated statement is faithful or suitable for a consequential decision.

Turn FATE into operational questions

Fairness, accountability, transparency and ethics become useful when each term changes evidence or action.

Readers compare outcomes for relevant groups, name the owner who can suspend use, decide what reason an affected person needs and test whether the purpose is justified. The final recommendation identifies the permitted domain, escalation route and review evidence. This creates a narrower but stronger claim than promising that a general label, high average score or human presence makes a system responsible.

In this chapter

What this chapter covers

  • 01

    AI and Generative AI

  • 02

    Pre-training and Fine-tuning

  • 03

    Retrieval

  • 04

    Black Box

  • 05

    AI Literacy

  • 06

    FATE Review

Worked example · free

Audit a retrieval-enabled assistant

Q [8 marks]. A service team says current retrieval removes the need to review generated replies. Identify the category error and recommend a bounded workflow. The marks shown are an AskSia study aid, not an official University marking scheme.
Classification map.Classification map. Generative AI sits inside the wider AI field, so a workplace claim must name the actual capability being evaluated. AI systemsMachine learningGenerative AI new content is one AI capability
Classification map. Generative AI sits inside the wider AI field, so a workplace claim must name the actual capability being evaluated.
  • 2Separate retrieval, generation and workplace action.
  • 3Locate the unsupported reliability claim.
  • 3Add evidence, review and accountable ownership.
Retrieval can supply current policy text, but generation can still omit, distort or add content. Display supporting passages, test citation fidelity and route consequential or unusual cases to an employee. The organisation should keep responsibility with a named owner and expand only when monitored evidence supports the narrower routine use.
Sia tip — Do not call retrieval training; attach each control to the stage where the error can enter.
Glossary

Key terms

Generative AI
Generative AI produces new content from patterns learned in data.
Retrieval
Retrieval supplies selected current material to a response without necessarily changing trained model parameters.
Black Box
A Black Box makes the path from input to output difficult for users to understand or challenge.
AI Literacy
AI Literacy combines knowledge, application, evaluation and responsible use.
FAQ

AI, Generative Models and Responsible Use FAQ

Why does the AI category matter?

The category identifies the capability, evidence and failure mode that should be evaluated. A predictive score, generated draft, retrieved passage and automated action require different controls. Naming the category prevents a fluent output or fashionable product label from doing the analytical work of explaining the system.

Does retrieval eliminate hallucination?

No. Retrieval can provide current source material, but a generator may still omit, distort or add claims. A defensible workflow displays supporting passages, tests fidelity and routes high-impact or unusual cases to a person who has enough information and authority to intervene.

How should a black-box concern be analysed?

Ask whether users and affected people can obtain meaningful reasons, whether performance can be tested for relevant cases, whether errors can be challenged and who is responsible for correction. The issue becomes organisational when opacity blocks accountable decision-making or remedy.

What is the difference between training and use?

Training shapes model parameters from historical material, while workplace use supplies a task and connects output to action. Fine-tuning, retrieval and user input occupy different points between them. State which stage is changing before assigning a control, because source selection, response verification and decision ownership solve different problems.

How does AI literacy improve analysis?

AI literacy combines conceptual knowledge with application, evaluation and responsible use. It lets a reader name the actual capability, inspect the evidence behind a claim and explain the organisational consequence. Knowing a definition is insufficient if the answer cannot locate uncertainty, affected people, review authority and a proportionate next step.

When is human review meaningful?

Review is meaningful when the person sees relevant evidence, has time and expertise, can disagree without penalty and possesses authority to change the outcome. Merely placing a person after the model may create ceremonial approval. The workflow should record disagreements and use them to test recurring failure patterns and boundaries.

Study strategy

Assessment move

Draw the nested AI taxonomy and the four-stage model-use pipeline without notes. For a new workplace example, label the input, transformation, output and action. Add one FATE question and one control for each vulnerable stage. Finish by writing the narrowest claim that the available evidence supports. Test the diagram with a counter-case in which retrieval is current but generation is inaccurate.

State which control detects the problem, who can intervene and why the organisation still owns the resulting decision.

Working through AI, Generative Models and Responsible Use in 6QQMN331? Sia is AskSia’s AI Management tutor — ask any 6QQMN331 AI, Generative Models and Responsible Use question and get a clear, step-by-step explanation grounded in how 6QQMN331 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.

A+Everything unlocked
Unlocks this Bible + all 18 of your subjects at King's College London, plus 1,000+ more Bibles across the AskSia library.
Sia - your 6QQMN331 tutor, unlimited, worked the way the exam marks it
The full 5-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works