Replying to your root comment but also addressing some points in subcomments. Generally (and perhaps ideally), I like the idea of more visibility/transparency

Replying to your root comment but also addressing some points in subcomments. Generally (and perhaps ideally), I like the idea of more visibility/transparency into what algorithms (and human moderators) do. But I think it's worth noting that, pragmatically, there are at least a few minefields here... - Providing full transparency into what gets flagged as spam is going to make it a lot easier for spammers to model, evade, and notice changes in the spam-identification code and practices. - Perhaps you were imagining this being temporary, but the operational feasibility of keeping all of this spam in-platform ultimately depends on signal/noise ratio, scale, profitability, and time horizons. It's one thing to bear the burden of storing 1% of your plaintext post volume for one year--it's quite another to store 20 or 30% of your video upload volume in perpetuity. - I'm having a hard time imagining the proceedings of the court-of-legal-vs-illegal-spam-filters. Without a lot of technical training (or a lot of reliance on expert testimony), the simplest way for an average person to grasp whether a spam filter is biased or not is going to be either looking "back" at what it did/didn't flag, or "testing" it with some sample corpus. In both cases, this process will be fraught with interpretation and sampling bias issues. This may hinge on a judge, panel, or jury being able to see through a bombastic argument that trots out the most respectable {ideology} opinions from the discard pile--they have to realize that nothing in the discard pile can be interpreted without a really good understanding of the input pile. (They'd have to be so good at this that they could spot someone who is trying to score political points and manipulate opinion by intentionally mounting an adversarial campaign that intentionally crafts seemingly innocuous ideological content that also gets flagged as spam). - Likewise, there's probably no great way to make a litmus-test corpus. Spam looks very different in different contexts. A useful Amazon review or tweet may both look like spam if someone emails or DMs them to you. Even if you can create a set of corpora that cover the bases, they'll probably be useless if they don't evolve, and very vulnerable to political and ideological manipulation if they do. - If there's a lot of legal fear around getting sued because your spam-filter has a taste for how {ideological subcommunity} likes to phrase its opinions, I'd expect spam filters to accumulate safe-words and rules that cause them to back off of charged topics. It'd only be a matter of time until legit spam/manipulation ops infer the existence of those rules and move to exploit them.