Andrew Ng stated that using AI for homework harms retention.

He noted that while most students use generative AI for writing assignments, providing direct answers, or completing papers, short-term grades may improve, but long-term memory retention and independent performance after removing the tool will be worse.

He attributed this to "cognitive offloading": the model completes retrieval, organization, reasoning, writing, and error correction for students, who receive the results of task completion without experiencing the extraction, trial and error, explanation, and repeated practice necessary for forming knowledge structures. The apparent ability increase when AI is present exposes capability gaps once AI is removed.

A randomized experiment involving undergraduates found that students allowed to use generative AI performed better on immediate tests, with some gains still present in retention tests a week later, approximately 5.1 percentage points; researchers emphasized that results depend on how students specifically use AI, and not all usage methods will reduce learning effectiveness.

Another group of studies and educational observations found contrasting risks: when students unrestrictedly let ChatGPT replace thinking, the retention test accuracy was 57.5%, compared to 68.5% for traditional learning groups, an 11 percentage point gap; a field study of Chinese high school students found that generative AI improved homework scores by 18% and reduced completion time by 30%, but monthly test scores dropped by 20% after six months.

The differing research results are not contradictory: AI as an explainer, questioner, feedback tool, or personalized practice system may enhance learning efficiency; AI as an answer generator, homework writer, or paper proxy reduces the frequency with which students complete critical cognitive steps themselves. The determinant of learning effectiveness is not "whether AI is present," but whether reasoning, recall, and error correction in tasks are still completed by students.

The most practical use is "do first, then ask": first write out the problem-solving path, code, argument, or summary without looking at AI answers, then ask AI to point out flaws, refute assumptions, generate variant problems, and test oneself with closed-book recitation. For example, when learning calculus, independently derive a problem first, then let AI check each step without directly providing the final solution; the next day, redo similar problems without AI assistance to test whether true capability has been formed.

In market mechanisms, schools, training institutions, and employers are facing the issue of separation between "demonstrable output" and "real ability." Students and knowledge workers will purchase AI tools to improve immediate delivery speed, while educational institutions need to increase oral exams, on-site tasks, process records, stage tests, and tool-free assessments; beneficiaries will be platforms that can design AI as tutors, examiners, and feedback systems, while those under pressure are assignment and certification systems that rely solely on generating final answers without verifying the learning process.

Source: Public Information