What is the biggest thing that still makes AI frustrating for you: wrong answers, too much setup,…

For me, the most frustrating failure is confident wrongness, because it creates work that looks finished. A slow answer or an annoying setup is visible; a plausible false claim can pass into an email, report, or decision before anyone notices.

Research surveys describe hallucinations as plausible but non-factual or ungrounded output and review retrieval, fact-checking, and verification as separate mitigation strategies. See the ACL survey on factuality and the NIST AI Risk Management Framework, which treats reliability, transparency, explainability, privacy, and accountability as distinct trustworthiness properties.

That suggests different fixes for different frustrations:

• citations, retrieval, and source checking for factual risk;
• compact task state and external notes for lost context;
• narrow tools and defaults for setup overhead;
• local processing or clear retention controls for privacy;
• a measurable workflow benefit for subscription cost.

The common product mistake is to treat all of these as a model-quality problem. Sometimes the model is adequate, but the interface gives it too much context, hides uncertainty, or makes verification harder than generation. I would rather have an assistant that says “I don’t know; here is what I checked” than one that is faster but silently increases the review burden.