45% of AI assistant answers contain at least one significant issue.
Almost half.

That is from the largest study of its kind. Answers evaluated by professionals across 22 organizations, 18 countries, and 14 languages.

  • 45% were flagged with at least one significant issue.
  • Of those: 31% had serious sourcing problems, missing, misleading or incorrect attribution. 20% had major accuracy issues, including hallucinated details and outdated information. Those can overlap.
  • 81% had some kind of problem when minor issues are counted.

There is a term for what this produces: AI slop. Machine-generated filler that reads finished and holds nothing up. It started years ago, before LLMs, as data slop. A data problem, bad inputs producing bad dashboards. Now it arrives as full sentences, which is worse, because a confident paragraph gets far less scrutiny than a chart ever did.

I spend a good part of my week inside these tools. Writing prompts, running LLMs, checking output. I keep coming back to the same three things.

AI literacy

The skills, knowledge, and ethical awareness that let someone understand AI, use it responsibly, and think critically about its outputs and its wider impact. Not just prompt writing. Not just coding. Literacy is knowing not to trust the output blindly, and how and when to go check.
e.g. all outputs come back to me labeled based on certainty: "Confirmed", "inferred", or "assumed"; I verify.

AI auditability

A design feature that lets reviewers trace, inspect, and verify how a system produced a given output. In practice, that means data provenance, model versions, accountability records, and alignment with frameworks. Most of us are not building these systems, but we need to know what a traceable one looks like, and ask for it. This is not only manual review. It can also mean building loops where the system checks itself before anything reaches a person.
e.g. when AI turns a source into a task, nothing goes in that was not in the source; additions get labeled as such.

Human accountability

AI cannot hold intent or carry responsibility for a bad outcome. People can. The failure mode is diffusion. Responsibility spreads across developers, users, and managers until nobody owns it. The fix is naming an owner for every AI tool in a workflow, and moving from human in the loop to human in power. Someone who can veto, not just approve. Because a human clicking approve on output they did not read is still slop, now with a signature on it.
e.g. I do not approve output; I rewrite it until it is mine. If my name is on it, I understand every line.

There is no single fix here. Everything is too new and moving too fast for that. It is continuous improvement, the same as any other delivery risk.

The study ran October 2025. Models have moved, but it is still the best map of where these systems break, and the categories still hold.