How to catch AI mistakes in your own assignments

AI models do not know what is true. They predict what text is likely to follow based on patterns. This means they can produce output that is grammatically perfect, logically structured, and completely wrong.

Hallucinated citations. Ask AI for references and it will invent author names, journal titles, and publication dates that do not exist. The format looks correct. The papers are fictional. If you cite them, you lose marks and credibility.

Confident but incorrect definitions. AI can define a technical term in a way that sounds right but misses a key detail. It might say "a foreign key is a column that links to another table" while leaving out that it must reference a primary key. Close enough to fool you, wrong enough to cost you marks.

Invented statistics. "70% of Kenyan developers use Python" sounds plausible but may be completely made up. AI generates statistics that fit the narrative without checking whether they are real. If you include them in an assignment, your lecturer may ask for the source.

Outdated information. AI models have training cut-offs. They may give you information that was correct two years ago but has changed. Laws, API versions, framework features, and company policies all change faster than models update.

Check every citation. Search for the paper title, author, and journal in Google Scholar. If the paper does not exist, remove the citation. If it does exist, read the abstract to confirm it says what the AI claims it says.

Cross-reference facts with primary sources. If the AI says Python was released in 1991, check python.org. If it claims a specific Kenyan law governs data protection, check the gazette or a legal database. Primary sources are the only reliable check.

Test all code. Copy the code into your development environment and run it. AI-generated code often has syntax errors, missing imports, or logical bugs. If the code does not run, fix it. If you cannot fix it, you do not understand it well enough to submit it.

Ask AI to critique its own output. This is a useful but imperfect technique. Prompt: "Review the following text for factual errors and unsupported claims." The AI may catch some of its own mistakes. But it can also defend them. Use this as one check, not the only one.

Read your submission as if you wrote every word. If a sentence contains a claim you cannot explain or defend, remove it or verify it. Submissions that contain sentences the student clearly did not understand are easy for lecturers to spot.

Round percentages with no source. "60% of companies" or "80% of graduates" with no citation. Real statistics come from real studies. AI statistics come from pattern prediction.

Suspiciously clean examples. AI-generated code examples often use variable names like "data," "result," and "output." They work for simple cases but break on edge cases. Test with unusual inputs.

Over-confident language. "It is well established that..." or "Research clearly shows..." are phrases AI uses to present claims with authority it does not have. If the claim matters, find the actual research.

Correct structure, wrong content. AI is very good at producing text that follows the right format (introduction, body, conclusion) while filling it with information that is partially or fully incorrect. Structure is not a guarantee of accuracy.

Mixing up similar concepts. AI often confuses related but distinct concepts: authentication vs authorisation, TCP vs UDP, stack vs queue. It knows the terms exist and that they are related, but it may swap their definitions.