Voice-First Written Assessment (VFWA) is a two-stage model that separates idea generation from final written production. The core assessment issue is authorship: how to preserve evidence of a learner’s own thinking when AI can help with drafting, refining, and polishing written work. VFWA is presented as a way to keep that thinking visible while allowing controlled AI use later in the process.

For assessment leaders, the appeal is design control rather than novelty. VFWA shifts the focus away from trying to detect AI after the fact and towards structuring the task so that reasoning, originality, and development can be observed directly. That matters for authenticity, academic integrity, and the credibility of written assessment in an AI-saturated environment.

- **Stage 1**: observed or closely supervised thinking and idea development. - **Stage 2**: AI-supported refinement of writing. - **Conditional viva**: a targeted oral check used as a safeguard where needed. The model is best understood as a control-and-evidence approach: first capture the learner’s own thinking, then allow constrained AI use for written development.

The source set points towards a shared practical view that assessment design matters more than post-hoc detection when AI is available. The stronger reading is that tasks need to match the construct being assessed: if the point is to evidence reasoning and drafting judgement, a staged model can be more defensible than a simple ban-and-detect approach. There is also broad agreement that any model like this needs clear rules about acceptable AI use, because ambiguity around permitted support quickly becomes an authorship and fairness problem.

The open question is whether VFWA preserves validity across subjects, levels, and cohorts. The source frames the approach positively, but it does not independently show that the model scales well or that the viva safeguard is reliable, consistent, or proportionate in practice. Another point of uncertainty is construct alignment. If the intended construct is independent composition, a model that permits AI in the final stage may be too permissive. If the construct includes responsible tool use, the same feature may be a strength rather than a weakness.

- **Validity risk** if the staged design does not match the assessment purpose. - **Workload risk** if conditional vivas become time-consuming or hard to schedule. - **Consistency risk** if judgement about when to trigger a viva varies across staff. - **Fairness risk** if learners receive uneven guidance on what AI support is acceptable. - **Governance risk** if institutions adopt the model without clear policy on authorship, evidence, and appeals.

- Start with the construct: decide whether the assessment is meant to evidence unaided writing, subject understanding, or AI-assisted composition. - Make the first stage meaningful enough to show the learner’s own thinking. - Define acceptable AI use in the second stage in plain language. - Use a viva only as a targeted check, not as a routine surveillance tool. - Ask whether the approach is proportionate for the subject, level, and cohort. - Look for evidence on reliability, scalability, and workload before wider adoption.

- Primary VFWA proposal and explanatory material.

The material here is not a vendor product page in the usual sense, but it does function as a market signal: a proposed design pattern for coping with generative AI in writing assessment. It should be treated as a concept and expert proposal rather than independent validation of effectiveness.

### What is Voice-First Written Assessment? It is a two-stage approach to written assessment that separates early idea generation from later writing, with controlled AI use allowed in the second stage. ### Why does it matter in exams or coursework? Because it tries to preserve evidence of the learner’s own thinking without relying mainly on AI detection after submission. ### Is it suitable for all subjects? Not necessarily. The fit depends on whether the assessment construct is independent writing, subject understanding, or AI-assisted composition. ### Does the viva solve authenticity concerns? It can help as a targeted safeguard, but stronger evidence is still needed on reliability, consistency, and workload before treating it as a general solution.

Tim Burnett (Admin)

Test Community Network. "Voice-First Written Assessment." TCN AI & Assessment Wiki. Last reviewed 2026-04-21. https://www.testcommunity.network/wiki/voice-first-written-assessment.html

- VFWA source page.