Show HN: I'm 16 and building an AI based startup called Factful with friends
A 16‑year‑old has launched Factful, an AI‑powered writing and fact‑checking tool aimed at combating online misinformation, drawing significant attention for both the ambition and polish of the project. Commenters praise the initiative but raise hard questions about whether large language models can reliably determine truth, urging techniques like source citation, retrieval‑augmented generation, and narrower domain focus instead of broad, general fact‑checking. The conversation also surfaces practical issues—bugs, security, scalability, business model, and legal setup—while broadly encouraging the founders to treat this as a learning platform and refine the product over time.
Product concept & current capabilities
- Web app for checking factual accuracy plus grammar/writing quality.
- Currently uses an OpenAI-based backend; team plans a custom LLM specialized for “text fitting” later.
- Fact-checking is intended to work via multiple non-LLM models that:
- Detect factual claims vs non-facts.
- Extract queries and search trusted databases/APIs (e.g., Google Fact Check, curated corpora).
- Feed retrieved evidence into an LLM that rewrites text using “known good” data.
- In the beta, this full pipeline is not yet implemented; a simplified version is deployed to test scale, bugs, and security.
User experience & technical issues
- Multiple users report errors, timeouts, or “nothing happens” on “Check Everything” / “Fact Check.”
- There is/was a global 1-request-per-second limit, causing failures under HN traffic.
- App was briefly trying to connect to a Vite dev server in production; dev-mode remnants noted.
- Some suggestions are overlong, off-topic, or UI-clipped; requests for a clearer, one-click “correct sentence” option.
- Bugs include naive text replacement (wrong number substitution) and strange factual edits.
Effectiveness & limitations of LLM fact-checking
- Many commenters doubt LLMs can serve as reliable fact checkers due to:
- Training on mixed-quality web data.
- Hallucinations and lack of calibrated uncertainty.
- Others propose mitigations: RAG over Wikipedia or curated sources, mandatory citations, mechanisms for expressing uncertainty, ensemble methods, and constrained outputs (function calling / yes–no answers).
- Several test cases show both successes (catching recent political news details) and clear failures (physics examples, joke prompts, history fuzziness).
Business model, legal setup & API
- Service is currently free; goal is to stay free as long as possible and later be cheaper than competitors.
- For businesses: planned per-query API pricing and per-user subscriptions.
- Company is incorporated in the UK because that jurisdiction allows 16-year-old founders.
- API and some business features are planned for summer, after school exams.
Ethical/scope debate & general advice
- Strong debate over whether “misinformation elimination” is even solvable, or desirable at web scale.
- Suggestions to focus on narrow, objective domains first and to emphasize human error reduction and writing aid.
- Many commenters are encouraging about the initiative and learning value, while urging realism about technical and societal limits.