The comparative study

TF-IDF vs word embeddings vs BERT vs HyFuse

Every number on this page comes from the test set: news the models never saw while learning. Pick a dataset; every chart has a plain-language explanation of what it shows and what we found.

Loading results

ℹ Live sources used by this site

⚙ Live fact-check sources

Paste your own free API keys. They are saved only on this computer (FakeNewsDetection/.env) and are never shown again in full.

Google Search AI key (Gemini, from aistudio.google.com) (best: Google's own AI searches Google live for every claim and cites the pages; starts with AIza... or AQ.)

Google Search API key (Serper.dev) (recommended: real Google + Google News results; 2,500 free searches)

Brave Search API key (optional second search engine)

Google Fact Check Tools API key (starts with AIza...)

Tavily web search API key (starts with tvly-...)

Groq AI evidence reader key (starts with gsk_...)

Backup AI readers (used automatically when the Groq free limit runs out; all free)

NVIDIA (build.nvidia.com) key (starts with nvapi-...)

Cerebras key (starts with csk-...)

GitHub Models token (github_pat_... with "Models" read access)

OpenRouter key (starts with sk-or-...)

Hugging Face token (hf_..., "Write" access; only used to publish the public link)

AI reader runs on (local Ollama: no API limits, works with your laptop's own AI, but 2-4 min per check on a CPU)

Wikipedia and Google News headlines are always used and need no key. With a search key, Verity opens the top results and reads the full articles, like you would after a Google search.