Chapter 4 — Academic writing
Lesson map
Lesson 04 published. Map, example, EX010.
flowchart TD
N0["Learning outcomes"]
N1["4.1 Claims and evidence"]
N2["4.2 Citation and verification"]
N3["4.3 Types of sources"]
N4["4.4 Integrity and AI tools"]
N5["Synthesis"]
N0 --> N1 --> N2 --> N3 --> N4 --> N5
N5 --> EX["Worked example"]
EX --> Q["Exercise EX010"]
Logical drawing
flowchart LR
N0["Learning outcomes"]
N1["4.1 Claims and evidence"]
N2["4.2 Citation and verification"]
N3["4.3 Types of sources"]
N4["4.4 Integrity and AI tools"]
N5["Synthesis"]
N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> Q["EX010"]
[Learning outcomes]
|
v
[4.1 Claims and evidence]
|
v
[4.2 Citation and verification]
|
v
[4.3 Types of sources]
|
v
[4.4 Integrity and AI tools]
|
v
[Synthesis]
|
v
[Worked example] --> [Exercise EX010]
Learning outcomes
You will (1) write clear academic prose with claims supported by evidence; (2) cite sources correctly and verify every citation; (3) distinguish peer-reviewed research, preprints, standards and grey literature; and (4) avoid plagiarism, including misuse of AI writing tools.
4.1 Claims and evidence
Every substantive claim needs support: data you produced, a cited source or a clear argument. In AI security, claims about attacks and defences must state conditions: which model, which attack, which metric. "Adversarial training makes models robust" is too broad; "adversarial training with projected gradient descent improved robustness to bounded perturbations on the evaluated image benchmarks (Madry et al., arXiv:1706.06083)" is supportable.
The supportable version names the method, the threat model and the scope of evaluation, so a reader can tell whether it applies to their system. A claim about image benchmarks says little about a text chatbot like Ask Lumen.
4.2 Citation and verification
Cite what you actually read, and check each reference exists: open the DOI, arXiv page or official URL. Do not invent identifiers. AI writing assistants can produce plausible but non-existent references; every citation you submit is your responsibility. Use a consistent style and include enough information for a reader to find the source.
Rina's worked example shows the risk is real: one of six references produced with an assistant did not exist. Checking takes minutes per reference and protects the essay's credibility. She records where she verified each one, such as the DOI page, so a marker can follow the same route.
4.3 Types of sources
Peer-reviewed papers have been checked by experts; preprints (for example on arXiv) are often important in AI security but unreviewed; standards and frameworks (NIST AI 100-1, ISO/IEC 42001) define terms and practices; vendor blogs and news are useful pointers but weak evidence.
The type of source sets how a claim is worded. A peer-reviewed result can be cited as a finding under its stated conditions; a preprint as a reported result not yet reviewed; a vendor blog as a claim. In Rina's essays, a claim supported only by a vendor blog is written as "the vendor states", which tells the reader how much weight to give it.
4.4 Integrity and AI tools
Follow your institution's policy on AI tools. Typically you must disclose their use, check their output and never present generated text or analysis as your own work without permission. Paraphrasing without citation is plagiarism.
Disclosure lets markers judge the work fairly and lets readers know which parts to check more closely. Rina's disclosure states the tool, the task it helped with, and how she checked the output. The invented reference in her draft shows why checking matters: the tool produced it confidently, and only verification exposed it.
4.5 Worked example (synthetic LCH case)
Rina's first essay draft, partly produced with an AI assistant, contained six references. She checked them: four existed, one had the wrong year, and one did not exist at all. She removed the invented reference, corrected the year, disclosed her use of the tool and rewrote the paragraph that had depended on the fake source using a real one she had read.
A rate of one fabricated reference in six is high enough that she now checks every reference before drafting around it.
4.6 Errors and limits
- Unverified references.
- Claims without conditions.
- Treating a vendor blog as established fact.
Unverified references can be wrong or invented, as Rina's draft showed, and a single fake one makes readers doubt the rest. Claims without conditions, such as "adversarial training makes models robust", are easy to refute with one counterexample from a different setting. Treating vendor blogs as established fact passes on marketing as evidence. Each error is cheap to avoid and expensive once a marker or reviewer finds it.
Exercises
- EX010 — Rewrite. Rewrite "prompt injection is solved by filters" as a supportable claim or a refutation with a citation. Rubric: 3 = conditional claim citing real work (e.g., Greshake et al. arXiv:2302.12173) showing filters are insufficient; 2 = partial; 1 = uncited; 0 = unchanged.
- EX011 — Verification. Describe how to verify an arXiv reference. Rubric: 3 = open arxiv.org/abs/ID, check title/authors/year match; 2 = partial; 1 = vague; 0 = none.
- EX012 — Source type. Classify: NIST AI 600-1; an arXiv paper; a company blog. Rubric: 3 = standard/framework, preprint, grey literature; 2 = two; 1 = one; 0 = none.
Synthesis
Academic writing in AI security means conditional claims, verified citations and honest use of tools — skills that also make professional reports trustworthy.
Continue studying — chapter 04 of 40
← Chapter 03 (draft) · Course chapter index · Chapter 05 (draft) →