How these materials were made
The materials for African Technical AI Safety were designed, written, and structured with substantial assistance from AI tools, primarily Claude (Anthropic). This page explains what that means, what was done to keep the quality high, and what to do if you find an error.
Four uses of AI
- Content drafting: first drafts of lesson text, explanations, worked examples and figure briefs were generated with AI assistance, then reviewed and edited by the convenor.
- Design and code: the site's editorial design, page layouts, responsive styling, and the build scripts (navigation, equation rendering, reference-linking) were written with AI assistance.
- Research and sourcing: AI tools helped identify relevant papers and resources. Every reading was then checked against the primary source (see below).
- Structuring: the 24-session arc, the African through-line, and the assessment design were developed collaboratively between the convenor and AI tools.
The quality checks
- Human review: every page has been reviewed and edited by Assoc. Prof. Jonathan Shock before publication.
- Reference verification: each reading is a working link, checked to resolve and to point to the correct work; arXiv identifiers, DOIs, and author and venue details were verified against primary sources, and corrections logged. This is the same verification the course asks students to practise.
- A de-Claudifying pass: the prose was audited for AI-writing tells and rewritten where needed, so the text reads as human-written.
- Iterative improvement: pages are revised as errors are found and as the field moves.
Known risks
Despite thorough review, AI-generated content can contain subtle errors that are hard to catch. The most common are:
- Hallucinated citations: references that look correct but point to papers that do not exist, or misattribute findings to the wrong authors or venue.
- Incorrect statistics: numbers, percentages or dates that are plausible but inaccurate, sometimes drawn from training data rather than a verified source.
- Broken links: URLs valid at the time of writing that later move or expire.
- Oversimplified explanations: technical points presented in a way that trades important nuance for clarity.
The case for transparency
A course on AI safety that hid its own use of AI would be teaching one thing and doing another. The position here is that:
- Transparency about AI use is an ethical obligation.
- Seeing where AI was used, and where it falls short, is itself part of the lesson.
- The imperfections are a risk, but they also demonstrate why the verification and red-teaming habits the course teaches are worth having.
Practising what we teach
The course argues, repeatedly, that you should not trust a confident model output without checking it: verify the reference, test the claim, probe the failure. Building the course ran straight into that. Reference-checking caught real errors before they reached the page (a misattributed co-author, conference-versus-arXiv year mismatches, a mathematical constant that had been mislabelled), and they are recorded in the course's verification log. That experience is part of what the course is about.
Found an error?
If you spot a broken link, an incorrect citation, a wrong statistic, or any other error, please email jonathan.shock@uct.ac.za or open an issue on the course GitHub repository. Corrections are welcomed and made promptly.
Licence and citation
CC BY 4.0 These materials are released under a Creative Commons Attribution 4.0 International licence: share and adapt freely, including commercially, with appropriate credit. Published under clauses 8.2 and 9.2.1 of the UCT Intellectual Property Policy (2011), which authorise the author of course materials to distribute them under a Creative Commons licence; UCT retains a perpetual, royalty-free, non-exclusive internal-use licence (clause 8.2).
Suggested citation: