Auditable assessment redesign for teaching readiness in preservice teacher education under generative AI

Synthesis: Liu (2026) reports a convergent mixed-methods evaluation of how one undergraduate EFL teacher-education program in China rebuilt its assessment once generative AI was permitted in coursework. Triangulating curriculum documents, student, graduate, employer and course-evaluation surveys produced one recurring problem: candidates were reasonably confident with tools but could not translate AI work into classroom-facing pedagogical decisions, so product-only tasks no longer supported the validity inference the program wanted. The redesign answers with auditable task structures — a shared task chain, required evidence trails, a six-dimension analytic rubric, and practicum-oriented mapping — so that judgment is graded as documented process rather than output. Two raters scoring 24 final artifacts reached ICC(2,1) = 0.77 for the total score but only 0.43–0.72 by dimension, while 24 micro-teaching packs scored ICC(2,1) = 0.873. This is feasibility evidence from a single program at a single time point, not a controlled comparison, and transfer into real classrooms stays unobserved in the data reported here.

Key Findings

1. AI coursework moved from peripheral to required across 2020–2025, traced with a composite AI-Integration Intensity Index (AIII-lite) that peaks for the 2025 cohort, but every module runs just 2 credits and 32 hours.
2. The binding constraint is transfer, not exposure. Of 160 valid student responses, 26.2% named hands-on practice and project guidance, 23.1% reusable cases and templates, and 23.1% stronger practicum or micro-teaching alignment.
3. Course-level evidence repeated the pattern. In the end-of-course NLP evaluation (N = 128), perceived concept mastery averaged 2.92 against 2.76 for transfer to teaching tasks, with 88.3% rating the course hard or very hard.
4. Readiness stalled where it mattered. Teaching-application readiness stayed moderate across cohorts (M = 3.62–3.81) while AI competence and governance awareness ran higher, and graduates (N = 85) rated preparedness near ceiling (4.61–4.72).
5. The redesign standardized four things: two signature task structures, six required evidence-trail components, a six-dimension analytic rubric, and mapping of course artifacts onto micro-teaching and practicum deliverables.
6. Rubric scoring held at the total level only. On 24 double-scored artifacts ICC(2,1) = 0.77 overall, but dimension agreement ran 0.43–0.72, weakest for Alignment (0.43) and Assessment and feedback (0.53).
7. Process evidence was produced but did not predict quality. Seventeen of 24 artifacts met all six evidence components, yet compliant and lower-compliance groups did not differ on total score (p = .874).

Where the assessment problem sits

Liu frames the difficulty as evidentiary rather than technological. Once lesson plans, materials and reflective writing can be drafted by a model, a polished submission stops separating the candidate's pedagogical reasoning from the tool's fluency: product-only tasks invite substitution and performance inflation, which Liu reads as a validity problem rather than only an integrity one. The construct is deliberately bounded — readiness to design, justify and evidence AI-allowed lesson decisions, not full classroom competence — and specified against TPACK and constructive alignment rather than as generic tool skill. The response is what Liu calls auditability: required prompt logs, verification records, revision memos and reflective rationales that let a teacher educator see the judgment behind the artifact.

What the redesign actually changed

The unit of change was the task, not the syllabus. Two signature structures were standardized — data → insight → pedagogical decision, and model output → human revision → justification — each ending in a classroom-ready artifact carrying six documentation components: prompt log, verification record, revision memo, reflection rationale, attribution or citation, and an AI boundary statement. A six-dimension analytic rubric separates alignment, teachability, governance, evidence trail, instructional judgment, and assessment and feedback rather than collapsing them into one impression of quality. Priorities entered under an explicit convergent rule: a candidate had to appear in at least two of five independent sources and be supported by at least one enactment-facing source — course evaluation, artifact evidence, implementation indicators — not perception alone. The decision trace tying each signal to a specific redesign move is the paper's methodological proposition.

What the convergent evidence showed

Students who passed an attention check (160 of 163) ranked clearer learning pathways fourth (17.5%) and clearer AI-use rules last (9.4%), which Liu reads as demand for translation support rather than more exposure. The NLP evaluation sharpens it: 84.4% called the course theory-heavy and 39.8% fast-paced, support averaged 3.25, boundary-setting readiness 3.27, and 44.5% said they needed explicit AI-use norms. Graduate ratings of 4.72 for digital tool use and practicum value are supportive but compressed at the ceiling, so they register endorsement rather than strengths. Employer feedback (N = 42, multi-select) is the external check, pointing at classroom planning and assessment-related expectations. Read against the curriculum documents, these converge on three priorities: teaching-facing transfer, auditable process evidence, and clearer AI-use boundaries.

What the feasibility evidence can and cannot show

The rubric was tested, not just described: two raters independently scored the same 24 artifacts, inter-rater reliability on the total score reached ICC(2,1) = 0.77 while dimension agreement ran 0.43–0.72, and mean scores sat between 83.0 and 84.9 on a 0–100 scale, weakest for Judgment (83.2) and Assessment and feedback (83.0). Twenty-four micro-teaching packs — a time-stamped script, one aligned formative check, an AI policy box, one supporting material — averaged 10.92 of 15 (SD = 2.21) with ICC(2,1) = 0.873, more consistent than the written artifacts. Two cross-institution implementations used the same design: 26 students in pairs at Site A produced 13 artifacts (ICC = 0.864), 25 working individually at Site B produced 25 (ICC = 0.931), normalized totals 0.695 and 0.689. Both sites named AI tool availability and baseline readiness as leading constraints. None of this is an effect estimate: no control condition, no sustained practicum observation, artifact quality measured once.

What this means for practice

• Instructors. Grade the trail, not only the product: require prompt logs, verification records and revision memos, then score alignment, teachability, governance, evidence, judgment and feedback separately so a fluent artifact cannot carry a weak rationale.
• Assessment designers. Check dimension-level agreement before trusting a total score: here the total reached 0.77 while alignment (0.43) and assessment and feedback (0.53) did not, so anchor those two dimensions with explicit descriptors and worked exemplars.
• Curriculum designers. Put AI and language-data competencies into required, credit-bearing modules and map every artifact to a practicum deliverable; at 2 credits and 32 hours there is no room to add content, only to standardize tasks.
• Administrators. Budget for the reusable case and template library students asked for (23.1%) and for practicum alignment (23.1%) before commissioning more AI content or policy text.

Limitations

• One initial teacher preparation program at one Chinese university, with a single anchor course (Introduction to Natural Language Processing) supplying the enactment evidence; Sites A and B (26 and 25 students) are feasibility checks, not comparisons.
• Every readiness claim rests on perception: 160 student responses, 85 graduate responses compressed at the ceiling (4.61–4.72), and 42 employer responses, collected anonymously, so no respondent can be matched to an artifact or a classroom outcome.
• Reliability comes from 24 artifacts and 24 micro-teaching packs; at that n a dimension ICC of 0.43 means individual alignment scores are not dependable, even though the total score is.
• The compliance–quality comparison has 17 fully compliant artifacts against 7, which Liu reports as low power; the null (U = 56.5, p = .874) is not evidence that evidence trails are inert.
• Micro-teaching is an enactment proxy, not observed teaching: no sustained practicum or in-service classroom data were collected, so improved assessability is not improved teaching.
• The curriculum-level trend rests on an internally computed AIII-lite index over 2020–2025 plans, with no external benchmark or independent audit of the coding.

Citation

Liu, H. (2026). Auditable assessment redesign for teaching readiness in preservice teacher education under generative AI. Discover Education, 5, 1309.