Abstract
Abstract: Indonesian students' mathematics achievement remains low: the 2026 national academic ability test (TKA) for junior high schools recorded an average mathematics score of 40.34, and Indonesia scored 364 in PISA 2025 mathematics, far below the OECD average of 463. Meanwhile, the government has introduced the Deep Learning approach and the Coding and Artificial Intelligence subject, which require computational thinking, while unrestricted generative AI has been shown to harm mathematics learning. This study examined the effect of Problem Based Learning with a Deep Learning approach assisted by guardrailed generative AI (PBL-DL + AI) on the computational thinking and numeracy of junior high school students. A quasi-experimental nonequivalent pretest-posttest control group design involved 95 eighth-grade students in three groups: PBL-DL + AI, PBL-DL without AI, and conventional instruction, on systems of linear equations in two variables. Data were analyzed using MANCOVA, ANCOVA, and Bonferroni post hoc tests. The results showed a significant multivariate effect (Wilks' λ = 0.752, F = 6.81, p < .001, ηp² = 0.13). For computational thinking, only the PBL-DL + AI group significantly outperformed the conventional group (d = 1.02), with the largest gains in abstraction and algorithm indicators. For numeracy, both PBL-DL groups outperformed the conventional group, with no additional benefit from AI. Guardrailed AI should therefore be positioned as a tutor that scaffolds thinking, not a replacement for meaningful contextual problems. Keywords: Problem Based Learning, Deep Learning Approach, Generative AI, Computational Thinking, Numeracy Abstrak: Capaian matematika siswa Indonesia masih rendah: rata-rata nilai matematika Tes Kemampuan Akademik (TKA) SMP 2026 hanya 40,34, dan skor matematika PISA 2025 Indonesia sebesar 364, jauh di bawah rata-rata OECD (463). Di sisi lain, pemerintah menerapkan pendekatan Pembelajaran Mendalam dan mata pelajaran Koding dan Kecerdasan Artifisial yang menuntut kemampuan berpikir komputasional, sementara penggunaan AI generatif tanpa pembatas terbukti dapat merugikan pembelajaran matematika. Penelitian ini bertujuan menganalisis pengaruh model Problem Based Learning berpendekatan Pembelajaran Mendalam berbantuan AI generatif ber-guardrail (PBL-PM + AI) terhadap kemampuan berpikir komputasional dan numerasi siswa SMP. Penelitian kuasi-eksperimen dengan desain nonequivalent pretest-posttest control group melibatkan 95 siswa kelas VIII dalam tiga kelompok, yaitu PBL-PM + AI, PBL-PM tanpa AI, dan pembelajaran konvensional, pada materi sistem persamaan linear dua variabel. Data dianalisis dengan MANCOVA, ANCOVA, dan uji lanjut Bonferroni. Hasil menunjukkan pengaruh multivariat yang signifikan (Wilks' λ = 0,752; F = 6,81; p < 0,001; η²p = 0,13). Pada berpikir komputasional, hanya kelompok PBL-PM + AI yang unggul signifikan atas kelompok konvensional (d = 1,02), dengan peningkatan terbesar pada indikator abstraksi dan algoritma. Pada numerasi, kedua kelompok PBL-PM unggul atas kelompok konvensional tanpa tambahan manfaat dari AI. AI ber-guardrail sebaiknya diposisikan sebagai tutor yang membimbing proses berpikir, bukan pengganti masalah kontekstual yang bermakna. Kata Kunci: Problem Based Learning, Pembelajaran Mendalam, AI Generatif, Berpikir Komputasional, Numerasi.
