Topic starter
21/08/2026 4:57 pm
TL;DR: Generative AI is expanding quickly across enterprise and scientific use cases, but adoption is still constrained by fine-tuning limits, bias testing challenges, and the need for trust, ethics, and transparency controls, according to Fiddler. The governance question is no longer whether generative AI will scale, but whether organisations can audit and constrain it fast enough to match deployment.
Editorial analysis by NHI Mgmt Group, based on content published by Fiddler: “Innovating with Generative AI”.
Key questions
Q: How should organisations govern human review for generative AI outputs?
A: Set human review only where the decision has material business, safety, or compliance impact.
Q: Why do bias tests and model cards matter in AI governance?
A: They provide evidence that the model has been examined for known limitations and that the organisation understands the context in which it is being used.
Q: When should organisations prioritise data governance over model deployment?
A: Before connecting GenAI to enterprise content, because model rollout amplifies whatever access decisions already exist.
Practitioner guidance
- Define decision thresholds for human review Map which generative AI outputs require mandatory human approval, which can be sampled, and which can proceed automatically with logging only.
- Build bias testing into release gates Require structured bias checks across relevant user groups and scenarios before any model enters production, then repeat them after prompt or data changes.
- Treat model cards as evidence packages Record intended use, test coverage, known limitations, and escalation paths so reviewers can see what the model has and has not been validated to do.
Bottom line: Generative AI still needs human-in-the-loop controls because bias, explainability, and fine-tuning constraints remain unresolved in many deployment settings.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
21/08/2026 5:13 pm
Human-in-the-loop is becoming a compensating control for weak AI governance, not a substitute for it. The article shows that organisations still rely on people to catch bias, context failures, and unsafe outputs after model generation. That is necessary, but it also reveals that the governance model is still centred on review rather than prevention. The practitioner lesson is to treat human approval as one layer inside a broader control stack, not as the core assurance mechanism.
A few things that frame the scale:
- According to Gartner, worldwide spending on generative AI is set to reach $644 billion in 2025, a nearly 77% year-over-year increase.
A question worth separating out:
Q: What happens when fine-tuning data cannot be used for a model?
A: The organisation may have to choose between a less adapted model and a model that fits the use case better but cannot be trained on restricted data. That trade-off affects accuracy, compliance, and explainability. Teams should decide the acceptable constraint up front, not after development has started.
👉 Read our full editorial: Generative AI governance still depends on human-in-the-loop controls