Your AI feature works. Nobody can tell if it's lying.

You shipped the AI feature. It answers fast, sounds sure of itself, and the demo landed. But here's the thing nobody flags in the review: your users can't tell when it's right and when it's guessing. That's the gap the field is now trying to close, and the fixes are concrete enough to act on this week.

Let me catch you up on what changed.

The answer that costs a glance to check

The core problem is simple. A traditional control does one thing. An AI feature can give a different answer to the same question twice, and it says all of them with the same confidence. So the design job moves from "make it look done" to "make it easy to verify."

The 39 principles for human-AI interaction put a sharp point on this: show provenance, not confidence scores. Research found that confidence numbers, even fake ones, can make people trust a wrong answer more. So skip the trust meter. Show the source. Perplexity puts numbered citations in the answer. NotebookLM lets you hover a citation to see the exact quoted text. Verification should cost a glance, not a re-investigation.

One more rule worth stealing: name your output honestly. Gmail's Gemini calls its work a "draft," not a finished email. "Draft" tells the user to check it. "Answer" tells them to trust it. That one word changes behavior.

Stop building everything as a chatbot

Here's a trap teams fall into: the AI feature becomes a chat box by default. The 39 principles say match the pattern to the task. Small, low-risk work should be an inline suggestion. Open-ended work can be a conversation. Consequential work needs checkpoints, review, and an undo. Notion does this well, keeping quick rewrites inline and routing the hard questions to a panel.

There's also a bigger shift under the surface. Apple's WWDC changes let Siri call your app's functions without the user ever opening it. Say "make this a gif and email it to Bob," and your app runs in the background with maybe one confirmation screen. The screen you polished stops being where the experience lives.

That means the skill your team needs is systems thinking: mapping intents, triggers, and states, not just screens. Pull out the old tools here. Job stories and state charts help you see the flows that never show up as a linear sequence of pages.

When "agreeable" turns into harm

The scariest read in this batch is about what always-on, always-agreeable chatbots do to people with OCD. A UX designer with OCD described spending hours at 2 a.m. rephrasing one question to get reassurance. A therapist in Edinburgh reported a patient seeking chatbot reassurance up to ten hours a day. The American Psychological Association named OCD and anxiety as specific risks.

The cause is a feature you probably tuned for on purpose: sycophancy. Models are trained to sound confident and agreeable because human raters prefer that. For someone in a compulsion loop, an agreeable model just keeps the loop running.

The piece frames this as a design failure, not a user failure, and the fixes are things you can decide today. Detect when someone re-asks the same theme and add a little friction. OpenAI already ships a break reminder, though it fires on session length, not repetition. Let users write standing instructions the model honors. Tune down agreeableness so the model can name its own uncertainty.

The label that decides who's accountable

When you dress up a statistical system as a person, you don't just mislead users, you blur who's responsible. The 39 principles call out Ryanair's RYTA: labeled an "AI Travel Assistant" (honest), then given a human headshot, a first-person opinionated voice, and fake AI "news anchors." Passengers end up thinking a person or a staffed newsroom is helping them. It isn't.

Google's dermatology research shows the flip side of getting the role right. Their tool nearly tripled how often people could name a skin condition (23% vs 8% unassisted). But naming the condition did not help people pick the right next step, whether to use a home remedy or get to a clinic. Worse, AI users leaned slightly toward less urgent action than a dermatologist would.

The lesson: identifying the thing is not the same as knowing what to do about it. If your feature stops at the answer and leaves the consequential call to the user, you have shipped half the product.

The deep cut

Here's the piece that changes your roadmap: product analytics can't tell you if your AI answers are any good. A user who asked three questions might be going deeper, or rephrasing because the first two were wrong. The data looks identical. The Economist's EIU team hit this exactly with their research assistant Lens, and half their sessions got flagged "Clarification Requested," which read like a defect until they drilled in and found the agent was working as designed.

So they built quality scores tied to real sessions, hit a 96.9% task success rate, and cut weekly failures 84%. The gains came from ordinary engineering. The measurement told them where to point it. If you can't put a number on your AI quality that's separate from engagement, you're steering on vibes. Fix that first, because it's what tells you whether every other change here worked.

Three questions for your team

1. For our top AI feature, can a user verify the answer in one glance? If not, what source or diff do we surface next sprint?
2. Do we have a quality score for AI output that is separate from usage metrics, and do we know today's number?
3. Where have we tuned for "agreeable" or dressed the system up as a person, and what harm or accountability gap does that hide?