INFS3822 Chap.10 Generative, Agentic and Responsible AI
Generative, Agentic and Responsible AI
Generative models create content or representations, while agentic workflows add goals, state, plans, tool calls and repeated action. This chapter separates generation quality from tool reliability and policy compliance. It applies the course AI policy through own-attempt requirements, process evidence, attribution, retained drafts, privacy controls, human approval and viva-ready understanding.
What this chapter covers
- 01
Transformers and generation
- 02
Autoencoders and GANs
- 03
Agent planning and tools
- 04
Least-privilege actions
- 05
Human approval points
- 06
AI process evidence
- 07
Attribution and drafts
- 08
Privacy and viva readiness
Worked example · free
Document a debugging interaction
- +1Retain the student's own initial pipeline, result and written concern before AI assistance.
- +1Record the tool and purpose: identify possible leakage routes rather than author the analysis.
- +1Keep dated interaction evidence showing the suggested checks.
- +1Verify independently, remove the post-outcome field, rerun the pipeline and explain the changed result.
- +1Attribute any incorporated output and document why suggestions were accepted, revised or rejected.
Key terms
- Transformer
- A neural architecture using attention to model relationships among parts of an input.
- Autoencoder
- A network trained to reconstruct its input through a learned representation.
- Generative adversarial network
- A generator and discriminator trained in opposition to produce increasingly realistic samples.
- Human approval
- A control point where an authorised person reviews evidence before a consequential action proceeds.
Generative, Agentic and Responsible AI FAQ
How does agentic AI differ from generation?
A generative model produces content from context. An agentic workflow maintains a goal or state, selects actions, calls tools, observes results and repeats. Tool use adds permission, validation, logging, recovery and approval requirements that content quality alone cannot satisfy.
What belongs in an AI process appendix?
Record each tool, its purpose, main prompts or workflow steps, relevant dated outputs or screenshots, how suggestions were evaluated and changed, required attribution, and team responsibility. Retain drafts and interaction records. The appendix should expose judgement rather than present an undifferentiated transcript.
Why might I need to explain my work in a viva?
The policy allows the teaching team to ask students to explain, discuss or apply material from any assignment. A viva helps verify understanding alongside drafts and process evidence. You should be able to reproduce the framing, transformations, model choice, metrics and recommendation in your own words.
What should happen when an agent is uncertain?
The workflow should preserve state, expose the uncertainty and stop or route the case to an authorised person. It should not invent identifiers, escalate permissions or repeat a possibly completed action blindly. Logs, idempotency controls, read-before-retry checks and explicit approval points make recovery reviewable when a tool response is missing or conflicting.
Assessment move
Draw the agent loop as context, generation or planning, tool action, observation and evaluation, with human approval in the centre. For responsible use, maintain a live ledger of own attempts, tools, purposes, outputs and decisions. Rehearse one-minute explanations of every model, figure and recommendation without relying on the interface. Distinguish three layers during revision.
A generative model produces content from learned patterns and supplied context. A tool-using workflow connects model output to an external capability. An agent adds state and repeated choice about what to do next. Each additional layer creates interfaces that need permissions, validation, logs, failure handling and a stop rule.
For a proposed proof-of-concept, write the allowed tools, allowed fields, maximum iterations, approval points and prohibited actions before describing the polished path. Create adversarial tests: a missing identifier, conflicting observations, denied access, a timeout after an action may have succeeded, an invalid tool argument and a request outside scope.
State whether the system should retry, ask for information, reverse a change, escalate or stop. Evaluate action success, unsafe attempts blocked, unnecessary loops and recovery, not only fluency. For assessment use, begin with the rule for the exact component.
Keep the student's initial attempt, purpose for using a tool, dated interaction, relevant output, independent verification, revision or rejection, attribution and team responsibility as a chronological chain. Do not reconstruct this record at the end. Pair every accepted suggestion with evidence of why it was accepted and every material rejection with the reason.
Protect private and confidential information by removing or synthesising fields unless authorised use is clear. For the no-assistance quiz condition, treat even a seemingly minor AI-enabled clarification as outside the published boundary. Finish by preparing a viva explanation without the tool.
Define the problem, justify data preparation, trace the model, interpret errors, defend the recommendation and identify what AI changed. If you cannot reproduce that logic, more screenshots will not establish student judgement. The goal is reviewable reasoning, not a long interaction transcript.
Working through Generative, Agentic and Responsible AI in INFS3822? Sia is AskSia’s AI Computer Science tutor — ask any INFS3822 Generative, Agentic and Responsible AI question and get a clear, step-by-step explanation grounded in how INFS3822 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
