Generative AI Model Lifecycle
Tags: Frameworks
TL;DR —
- Four stages: Planning & Data Collection; Model Development; Deployment & Integration; Usage & Maintenance.
- Add explicit human oversight calibrated to impact.
Why it matters for HK marketers: Knowing the lifecycle pinpoints where to demand vendor controls (data rights, testing, labeling, logging) before using AI in-market.
Stages and risks
Planning & Data Collection
- Risks: harmful/poisoned data, bias, personal data privacy breaches, IP infringement.
- Mitigation: source diversification, automated/manual filtering, stakeholder authorisation, domain-specific fine-tuning for high-stakes sectors.
Model Development
- Risks: overfitting/underfitting, IP/licensing violations for architectures/weights.
- Mitigation: design appropriate metrics; comply with open-source licences; evaluate accuracy and reliability.
Deployment & Integration
- Risks: system integration fragility; access-control inheritance errors.
- Mitigation: redundancy/high availability; comprehensive integration testing; strict least-privilege permissions; evaluate cloud vs on-prem vs on-device vs hybrid.
Usage & Maintenance
- Risks: content safety, data exposure (incl. RAG-accessed data), copyright ambiguity, social trust threats from realistic fakes.
- Controls: input filtering, content labeling, clear user guidance, declared copyright ownership where practical, consent-based logging.
Human oversight levels
- Collaborative models: limited oversight for low-impact contexts.
- Human-dominated models: primary human control where impact is higher.
So what for marketers —
Build lifecycle checkpoints into procurement and QA: data provenance reviews, red-teaming/jailbreak testing, content labeling, and opt-in logging before launch.
← Back to Knowledge Base