An AI gets ten messages in Nigerian Pidgin. Some are harmless but sound harsh. Some are genuinely harmful. You decide: should the AI help or refuse?
Real safety filters make this call millions of times a day. Ready to try?
Message 1 of 10
What it means:
In plain English:
You got
0/10
Now imagine making these calls millions of times a day, in hundreds of languages, while scammers deliberately phrase things to slip past you. Refuse too much and honest people get shut out. Refuse too little and bad actors get help. That balance is what Salecta measures for real AI models.
How did the AI models do on these same messages? Their results will appear here once the Salecta evaluation runs.
Messages come from the Salecta dataset and have been checked by a native Pidgin speaker. The harmful ones are low-risk examples of common scams and abuse.