Creation and Evaluation of a Pre-tertiary Artificial Intelligence (AI) Curriculum

In plain words

This paper is about teaching AI in secondary school, not about an AI model. A Hong Kong project brought university professors, school teachers and an AI industry partner together to write a course on AI for students aged 12 to 16. Rather than writing the course alone and handing it down, the teachers helped shape the material, which they then adapted to their own classrooms.

The course is split into twelve chapters grouped into five themes: awareness, knowledge, hands-on interaction, empowerment (building end-to-end systems) and ethics. Teachers choose the parts that fit their students. The authors gave the course to students before and after teaching and compared the answers. Students reported more AI knowledge, more confidence and more interest in the subject afterwards. The teachers said that the shared work taught them AI and gave them more freedom to design their own lessons. The study has no control group, so the results show change over the course, not proof that the course alone caused it.

TL;DR

• A co-created AI curriculum for secondary schools in Hong Kong: 12 chapters, five modules, three levels of depth §III tab. I tab. II.
• Pilot with 335 students aged 12-16: all five student measures rose after the course, with paired t-tests at p <= 0.005 §V.B tab. III.
• The paper's main claim is about the process: the co-creation model gave teachers AI knowledge and autonomy §VI.A.
• Author-reported and uncontrolled. Its student counts and item counts disagree in places (see Pitfalls).

Problem

AI is usually taught at university level. The paper argues that there was almost no formal AI curriculum design for pre-tertiary education, and that the subject needs content that is foundational and concrete, flexible enough for schools with different resources, and adaptable by teachers §I §II. The paper also treats the gap in teacher knowledge as part of the problem: the teachers in the study had received no formal AI training §V.C.

The paper does not state a quantitative target. Its research questions are: what the curriculum contains (RQ1), whether it improves perceived competence, attitude and motivation (RQ2), and how teachers perceive the co-creation process and whether it fosters their autonomy (RQ3) §IV.A.

Method

This is an education study, not a learning algorithm: it has no model, loss or training run. The "method" is a co-creation process for writing the curriculum, and a pre/post questionnaire plus interviews to evaluate it.

1. Team. Fourteen professors from the engineering and education faculties of CUHK worked with 17 principals and teachers from six pioneering secondary schools. About 15 postdoctoral fellows, research assistants and undergraduate helpers also took part §I fig. 4.
2. Design principles. The curriculum draws on K-12 engineering education. The paper lists five principles: interdisciplinary content, impact and ethics, technical communication skills, engineering thinking, and engineering techniques §II.
3. Structure. Twelve chapters are organised into five modules: Awareness, Knowledge, Interaction, Empowerment, and Ethics and Impact tab. II §III. Each chapter has units tagged beginner, intermediate or advanced tab. I fig. 1. The modules are meant to let teachers pick and adapt content §III.
4. Development. Professors wrote the technical content. The 17 teachers then refined learning outcomes and pedagogy in a year of regular meetings held every second week, on Monday and Friday afternoons §IV.B.
5. Implementation. Teachers chose and tuned content for their own school culture, environment and resources. Teaching during COVID-19 ran online, face to face, and blended §IV.B.
6. Teaching tools and hardware. The paper names Jupyter Notebooks, Blockly, web apps and Google Teachable Machine. It also describes CUHKiCar, a robot car with six built-in AI functions, and two AI experiments on it: face tracking and line following §V.A fig. 2.
7. Evaluation. Students filled in an online pre-questionnaire and post-questionnaire. Eight teachers took part in semi-structured interviews §IV.B. Five constructs were measured on a 6-point Likert scale §IV.C.
8. Analysis. Paired t-tests answered RQ2. Analysis of covariance (ANCOVA) was used on the post-questionnaire scores, citing a pretest-posttest reference §V.B [23].

The student analysis (step 8) can be written as the following pseudocode. The paper describes these steps in prose. The covariate in the ANCOVA is implied by its citation and is not stated explicitly.

# Pre/post analysis of student questionnaires  [§IV.C] [§V.B]
# Input:  pre, post: per student, per construct, 6-point Likert means
# Output: per construct, paired t statistic and p value

def analyse_student_scores(pre, post, constructs):
    # Keep constructs with Cronbach alpha above 0.70  [§V.B]
    # (all five were reported at 0.88 to 0.92 in tab. III)
    reliable = [c for c in constructs if alpha(c) > 0.70]
    # Check that variances are homogeneous (Levene's test)  [§V.B]
    check_homogeneity(pre, post, reliable)
    results = {}
    for construct in reliable:
        # RQ2: paired t-test, post against pre  [§V.B]
        results[construct] = paired_t_test(post[construct],
                                           pre[construct])
    # ANCOVA on post scores; pre as covariate is implied, not stated  [§V.B]
    results["ancova"] = ancova(post, covariate=pre)
    return results

Hyperparameters

This paper has no training hyperparameters. The table lists the study parameters an implementer would need to reproduce the design.

Results

Author-reported. Setup: students aged 12-16 from the pioneering schools; pre-questionnaire and post-questionnaire on the same students; no control group is described §IV.B. Table III is reproduced in full from the paper.

Paired t-tests, post against pre, df = 335 §V.B:

Qualitative findings (author-reported, no counts given):

• RQ1, curriculum features §V.A. Two main features: flexible learning pathways (modular design, levelled units, and examples chosen by the school), and local and global understanding (classroom tasks start from local examples such as chatbots and transport, then move to global ones).
• RQ3, teacher perception §V.C. All eight interviewed teachers said they had not received formal AI training. They felt more able and more confident to teach AI after the co-creation work. All reported that the modular design let them adapt the content to their own schools.

Paper vs code

No official code. The paper gives no code, repository or project link. A GitHub search for "AI4Future" returned ten repositories, all owned by accounts that do not match the authors, so none was used. The curriculum materials themselves are not described as public in the paper.

Pitfalls

• Student count. The abstract and text say 335 students §V.B, and the t-tests use df = 335. Table III's header says N = 358 tab. III. The paper also reports 8 teachers, so 335 + 8 = 343, which matches neither figure. Treat the sample size as unclear.
• AIRE item count. The text says AIRE has 4 items §IV.C, but the appendix lists AIRE1 to AIRE6 app..
• Reliability values. The text gives alpha = .88 for AIRD and AICF, .91 for AIRE and .74 for AIIM §IV.C. Table III reports pre-test alphas of .89, .91, .90 and .92 for the same scales tab. III. The paper does not say which sample the text values come from. The AIIM figures disagree most (.74 against .92).
• No control group and no effect sizes. The design is a single-group pre/post comparison. The paper reports t and p values but no effect sizes, and no correction for testing five outcomes.
• ANCOVA covariate not stated. The text says ANCOVA was used on post-scores §V.B, and cites a pretest-posttest reference [23]. It does not name the covariate or the model.
• Language of the questionnaire not stated. The appendix items are in English. The paper does not say what language the students answered in.
• Teacher sample. Eight of 17 teachers were interviewed. The paper does not say how they were chosen.
• Level count. The paper describes "five levels of depth" §III, then three categories (beginner, intermediate, advanced). Tables I and II use the three-level labels.
• Figure numbering. The paper cites Figures 1, 2 and 4. Figure 3 does not appear in the text of the PDF. Cite figures by the numbers the paper uses.
• Blended teaching. During COVID-19 the teaching mode varied (online, face to face, blended) §IV.B, and the paper does not separate results by mode.

Lineage

• Builds on: no arXiv work. The design draws on K-12 engineering education and teacher-autonomy literature cited in the paper §II §III (refs [1], [8]-[13]), none of which is on arXiv.
• Compared against: none. The paper has no comparison group.
• Superseded by: not known from content in this wiki.