• workshop_dense_100#slide 40

Human-AI hybrids are organisational arrangements in which humans and AI systems jointly produce outputs and organisational consequences. The RAIDT papers use this idea to move away from the misleading choice between 'human decision' and 'AI decision'. In practice, GenAI drafts, recommends, summarises, or classifies, while people set the task, choose or approve context, interpret results, decide reliance, and remain accountable for downstream effects. The hybrid is therefore the real unit of work, even though RAIDT makes the run as the unit of governance for review purposes.

This matters in RAIDT because governance failures often arise in the hand-offs inside the hybrid. A model may produce plausible text, yet the real governance question is whether the human reviewer understood its limits, whether the provenance was preserved, and whether the organisation can later show who approved use and on what basis. RAIDT addresses this by using the run-level evidence pack to record both technical behaviour and human oversight, and by using the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) to assess the quality of that hybrid arrangement. It also treats influence methods as governance interventions, because prompt structure, retrieval logging, adapter versioning, and alignment controls change both model behaviour and the evidence available to human reviewers. RAIDT therefore makes human-AI hybrids reviewable, contestable, and manageable rather than merely assumed.

In an HR shortlist workflow, a recruiter uses GenAI to draft candidate justifications and interview questions against a versioned organisational rubric. The AI produces text quickly, but the recruiter and hiring manager decide whether the reasoning is acceptable and whether the shortlist can be defended later. This is a human-AI hybrid because the final organisational act depends on both generated content and human approval.

RAIDT makes the hybrid accountable by storing the prompt template ID, criteria version, model deployment ID, any LoRA or PEFT adapter identifier, the generated justification, and the review note showing who approved or challenged it. If a candidate disputes the process months later, the organisation can reconstruct the run rather than rely on recollection. The score profile will reveal whether the hybrid preserved enough Responsibility and Traceability to justify use, or whether the workflow needs tighter controls before further reliance.

• 17-RAIDT-Sociotechnical_M_v6
• 11-RAIDT_Academic_Logic_M_v11