Even if the language understanding were better, you would still have the issue of the "toxicity" classifier reflecting the personal opinions of the people who b

Even if the language understanding were better, you would still have the issue of the "toxicity" classifier reflecting the personal opinions of the people who built the model and not the range of acceptable behavior for the public at large. These labels aren't objectively measurable statistics: they're personal judgements. The tyranny of algorithmically reified personal moral judgement is far scarier to me than bad keyword matching.