But the way we detect bona-fine spam is by user reports. We then train models / decide heuristics based on these reports, but in the end, users decide what they report.