AI slop is degrading a public verification workflow
For a lot of people, the replies were never small talk. They were the fastest way to see whether a claim survived contact with other humans.
That is why saying X no longer feels like social media is too mild. The larger problem is work. A lightweight public verification workflow is being degraded in plain view, and the effect is practical: visible replies become less useful for checking whether a news claim, research summary, or product assertion holds up when other people inspect it.
Ethan Mollick made the problem concrete in April 2025. In a post on X from April 2025, he said that each of his posts gets dozens of obvious AI slop replies. In the same post, he wrote, "You can't really call this social media anymore." That line describes a cultural change, but it also points to a systems change. If the visible layer of response is dominated by filler, the public check stops working as a check.
Mollick also said in that April 2025 post that even when humans do reply, the replies algorithm does not necessarily surface those human responses first, including a case where someone from Anthropic addressed the issue he had posted about. The issue is not only that low-value text appears. It is that ranking can bury the useful correction under generic participation. A crowded reply stack where one specific human correction disappears under plausible-looking filler is a failure mode, not just an annoyance.
A decision trail is the short path a careful reader uses to test a claim in public, and at work it looks like opening the replies, finding a named person with relevant knowledge, and checking the source they point to. That trail is thin, but it has been useful because it was fast. When the first screen of replies stops being a map to informed disagreement, the trail breaks at the first step.
Mollick made the comparison sharper in another post on X from April 2025, where he said LinkedIn has better quality comments at this point, while adding that LinkedIn is also struggling with AI. The point is not that another platform is clean. It is that the threshold for a usable comment layer has moved, and X is now failing a job many readers had assigned to it.
What is lost if nobody decides around that failure is accountability. The ranking starts to stand in for judgment. A top reply begins to look like the useful reply, even when it is only the most surfaced reply. In a workflow built on quick public checks, that substitution matters.
Verify the humans, not the ranking
The control that remains is simple and manual. Assume visible replies on X are untrusted by default this week. Do not read the stack as a verdict. Read it as an inbox that now needs filtering.
A particularly prepared reader still has several workable checks:
- Look for identifiable human responders, especially people speaking in their own name.
- Open the source links rather than relying on summaries in the reply text.
- Prefer replies that add evidence, correction, or direct domain knowledge over style, agreement, or paraphrase.
- Verify claims through known experts instead of assuming the top reply is the useful one.
This does not restore the old workflow, but it does preserve the decision that matters. The reader still decides which response counts as verification. That is slower than trusting the ranking, yet it is better than outsourcing judgment to a reply order that may be rewarding volume, imitation, or generic fluency.
The practical takeaway is narrow: if you use X as a live review layer for news, research, or product claims, assume the visible replies are not pre-sorted for usefulness. Check who is speaking, what source they point to, and whether their response can survive a click.
Human Runtime Core™ is designed for the mental edge you need to survive in an AI-run world.