Getryt: AI misinformation detection and verification system for digital safety

I like the way you think about Identity verification to solve for digital safety, however you can go into un seen areas such as ads as a whole, you do not have to restrict yourselves to financial scams etc but I want to see the model tested on more important scams in online forums such as human trafficking, drug abuse, medical scams which can be helpful to see the model perform.

One-line summary: A working web prototype (Node.js, with a live Hugging Face demo and GitHub repo) that lets South African users paste suspicious messages and get a verdict by matching them against a curated database of known scams and verified claims, then redirects them to trusted official sources; evaluated qualitatively on a small set of scenarios.

Constructive critique:

The most refreshing thing about this submission is its honesty. After several entries that overclaimed, Getryt does the opposite: it labels its accuracy explicitly as "qualitative," calls itself a feasibility prototype rather than a finished system, openly explains why it pivoted from a generative approach to a retrieval one (hallucination risk and API cost), and ships an actual usable demo people can try. The two product instincts are also sound. Choosing retrieval-and-match against trusted claims over free generation is the right safety call for a misinformation tool, and pairing detection with redirection to authoritative sources (NSFAS, government channels) is a genuinely practical, locally-grounded idea that most flag-only tools miss. The related-work section is decent and the citations are real.

The core problem is the gap between what the paper claims to be and what was actually built. The title and abstract promise detection of AI-generated content, synthetic media, and impersonation, but the mechanism is essentially a lookup against a small curated list of already-known scams: anything not in the database returns "Not yet verified." That is a known-scam blocklist with a helpful redirect, which is a reasonable thing to build, but it is not AI-generated-content detection and it cannot catch the novel or adaptive misinformation the introduction spends most of its time describing. The honest reframing would be "a localized scam-verification assistant," and the project would be stronger for claiming exactly what it does.

The second problem is that there is no quantitative evaluation at all. The results table reports "High" and "Moderate-High" with no number of test cases, no precision or recall, and crucially no false-positive behavior on legitimate messages, which is the failure that matters most for a tool ordinary users will trust. Even within a qualitative frame, the eval discipline Hamel Husain and Shreya Shankar describe still applies: say how many messages you tried, show the cases it got wrong, and report what happens when a real government message or a paraphrased scam comes in. Right now "High accuracy" rests on an unspecified handful of examples that likely overlap with the curated database itself, which makes the result close to circular.

Concrete next steps. First, align the claims with the build, or extend the build to match the claims: if you want to call it AI-content detection, add and benchmark an actual classifier; if it stays retrieval-based, reposition it honestly. Second, produce a real evaluation, even a small one, with counts, a confusion-style breakdown across your four categories, and an explicit false-positive test on genuine messages. Third, tighten the writing. The abstract alone contains two broken sentences ("By providing concise explanations and evidence-based assessment." and "Our findings suggest that accessible AI-enabled deception."), and there are scattered typos elsewhere ("BThe system," "division-making"), which undercut credibility on first read.

Track + flags:

On-topic (Global South AI safety, misinformation tooling, South Africa). LLM use disclosed, references real, working demo and code provided, no plagiarism signals. Note for the panel: the system is a curated-claim matcher rather than the AI-generated-content detector the title implies, and the evaluation is qualitative with no sample size or error breakdown.

Strengths: Presents an interesting concept with potential practical applications. The overall vision is understandable and addresses a relevant user need. Areas for Improvement: The implementation appears to be at an early stage and would benefit from additional technical depth. More comprehensive validation, user testing, measurable outcomes, and architectural documentation would strengthen confidence in the proposed solution. Greater differentiation from existing alternatives would also improve the project's value proposition.