Because we make sure they are trustworthy. Every network member (watchers sharing their signals) gets a trust rank (TR). By consistently sending back valuable and exact information, the TR gets better over time. A daemon reporting for months, with 100% accuracy, valuable information will eventually reach the maximum TR. Feeding the system with wrong information would result in a severe and immediate loss of TR. This mechanism is made to avoid poisoning.
All TR can partake in the consensus, but only the highest TR rank can publish to the database without needing validation from our own honeypot network. It nevertheless has to pass the test of the Canary list, meaning the IP reported shouldn’t be one of the canary. Canaries are in fact whitelisted IP, known to be trustworthy, like the Google bot, Microsoft updates, etc. If a scenario is too sensitive or twitchy, it might shoot a canary. This mechanism is made to avoid false positives.
All those mechanisms (and more to come) contribute to what we call the Consensus chamber (Consensus in short), where the decision is taken to either ban the IP responsible for an alert or not.