AI literacy is AI Safety

This is a strong, safety-conscious work on a real and well-chosen problem. Your related-work section is original by showing that no existing African initiative pairs an influencer-distribution model with a safety frame. The strongest part is the dual-use section: recognising that the same pipeline is a turnkey disinformation engine, and that your sponsors are the very firms the public most needs to scrutinise, is exactly the self-aware risk thinking this track rewards. The honest weak point is that the central claim is untested: the appendices show the pipeline runs (kudos for publishing the raw, broken translation output rather than cleaning it up), but not that a skit changes how anyone thinks about AI, and your own measures are proxies you rightly admit don't capture literacy. The most valuable next step is the pilot you already sketch, the matched-content test with a pre/post survey, that's what turns a well-built pipeline into evidence it produces the literacy you're after.

This paper uses South Africa as a case study to argue that “AI literacy is an AI safety imperative” and presents an AI literacy program “that converts peer-reviewed South African AI research into multilingual, safety-centred social media content” with a sustainable funding model whose example can be replicated across the Global South. It advocates for a safety-focused model beyond the existing formal education programs on upskilling and digital literacy. It offers two detailed examples of how to convert academic papers to multi-lingual social media content.

The recommendations are highly localized and provide well-researched and persuasive examples of past success and failures of similar programing in diverse contexts. The methodology focuses on translation of “how AI affects everyday South Africans” and “ground[ing] content in local realities.”

Due to reliance on LLMs to translate the academic lens, I worry about how to manage unintentional misinformation through oversimplification when prioritizing virality and relying on non-technical experts to spread critical information. How would the program evaluate success, and provide continuously monitored verified channels to build trusted creators at scale?

Regarding funding, the authors were correct to avoid government bureaucracy, but may benefit from the creation and management of a dedicated private sector or nonprofit managed, independent fund that would absorb incoming funds from sponsors and distribute to vetted creators. The fund would manage: the financial structures and payments, vetting and influencer seeding, measurement of outcomes/reporting frameworks, and public record of sponsors outlined in the paper to have a single actor responsible for the operations of the program to ensure success. This would limit the misuse risks outlined in the paper, but introduces centralization risks if the fund’s staff include particular biases towards certain political or issue-specific outcomes.

As outlined in the limitation section, the measurement framework outcome to “shift…public confidence around AI” may prove difficult due to the qualitative nature of this statement and would benefit from quantifiable metrics to assess this specific outcome. I’m also worried about how to ensure these scripts are organic and don’t read as a govt PSA where audiences see the exact same content from multiple creators. When thinking about scaling the program and long term success, I think a better strategy may be to develop technical and academic institutions’ advocacy and communications capacity instead of relying on influencers and content creators. The paper would benefit from a clearer explanation on why a dedicated AI safety lens is required vs. the efficiency of including the safety conversation within a broader conversation on digital resilience.

The reframing is the contribution and it is a strong one: treat AI literacy as AI-safety infrastructure and distribute it through micro-influencers in eleven languages, funded by private sponsors rather than government budget cycles, in a population where 73% barely register the term "AI". The dual-use section is unusually candid, and the one executed component — the Vulavula translation run — is reported with real honesty: 101 of 144 calls succeeded, 43 failed on a rate limit, and Appendix C reproduces the raw output verbatim, corruption of the POPIA acronym into P_O_P_A and all. That candor cuts against the claim, though. The executed step demonstrated fragility, not reach, so "a present technical capability" overstates what untranslated fragments and code-switching show, and everything downstream — the literacy-to-safety theory of change, the engagement metrics, the funding model — is design, not evidence. The two worked examples are illustrative by the authors' own admission. The missing core is an outcome: run the matched-content A/B test the Future Work section proposes, even once, and measure whether anyone's AI awareness actually moved.