There’s nothing wrong with AI, these contextual problems are not a mistake–they’re a choice.
AI can be trained for deeper analysis and to root out issues like this. But that costs compute cycles. If you’re selling a service, you want to spend as little on compute power as possible while still being able to have a product that is viewed as good enough to pay for.
As with all things, the root of this problem is greed.
Also: https://sls.eff.org/