Use AI as Part of a Workflow, Not as a Substitute for Thought

AI use becomes weak when the tool decides the problem as well as the answer. Before opening a chatbot, write one sentence stating the real output: explain a theorem to first-year students, compare methods across five papers, identify gaps in a draft argument, or create practice questions from an approved syllabus. Then list the constraints. What sources are permitted? What must be calculated independently? What personal or unpublished material must not be shared? What would count as a satisfactory result? This small pause changes the relationship. The tool receives a bounded task instead of being invited to manufacture the intellectual direction.

Prompting belongs inside this stack, not above it. A detailed prompt cannot repair missing knowledge, an invented citation, or a misunderstood theorem. In academic work, a prompt is best treated as a temporary specification: it tells the system what role to play, what material it may use, what it must not assume, and how uncertainty should be shown. The user should still expect omissions and errors. Fluent language is a presentation feature, not proof that the underlying claim is correct.

Suppose a postgraduate student has six verified papers on retrieval practice and wants to plan a literature-review section. The student first reads each paper and records the question, method, sample, result, and limitation. The student then gives the AI only this source matrix and asks: group these studies by the kind of learning outcome measured; identify agreements and disagreements; point out where the matrix lacks evidence; do not add sources. The response is checked against the matrix. Any unsupported sentence is removed. The student writes the paragraph from the verified relationships, cites the original papers, and records that AI assisted with organisation. Here AI reduces coordination work but does not replace reading, synthesis, or citation responsibility.

Privacy deserves special attention because convenience can hide the fact that information has left the user's controlled environment. Student records, unpublished manuscripts, interview transcripts, examination materials, passwords, financial details, and identifiable personal information should not be placed into a public AI service unless the institution has explicitly approved the system and the handling arrangement. Removing a name may not be enough when the remaining details identify a person or project. When uncertainty exists, use a fictional example or keep the task offline.