Irreplaceable › Guides › AI UX mistakes
Examples · AI product design
AI makes screens that look done. The problems hide in plain sight: a price that broke your limit, a wrong answer marked right, a test deleted so the build passes. Here are 42 of them on 14 AI-made screens, why each one hurts, and the fix.
By Sarit Elisha, founder of Irreplaceable · Updated October 10, 2026
Tap the problems on a new AI screen, then see what you missed. Free.
AI agent · mobile. Try to spot the 3 problems first, then show them on the screen.
01
Area 1 on the screen · AI behavior design
A $468 irreversible purchase happened with no review step. The higher the cost and the lower the reversibility, the more the agent must stop and confirm.
The fix: Show the plan, then “Approve and pay”.
02
Area 2 on the screen · Evaluating AI output
You said under $420. The agent paid $468 and didn't flag it. AI output looks finished, so the violation hides in plain sight.
The fix: Flag the conflict: “Cheapest aisle seat is $468, $48 over. Book anyway?”
03
Area 3 on the screen · Ethics & trust
Non-refundable, in faint grey, after the fact. No undo window, no cancel.
The fix: A 24-hour free-cancel window, stated before paying.
Health AI · mobile. Try to spot the 3 problems first, then show them on the screen.
04
Area 1 on the screen · Evaluating AI output
Chest pain on exertion is a classic warning sign. “92% match” sounds scientific and steers away from urgency.
The fix: Red-flag symptoms skip the guess: “This can be serious. Call a doctor or emergency services now.”
05
Area 2 on the screen · Ethics & trust
A paid partner offer right after a diagnosis is a conflict of interest, and it nudges the wrong action.
The fix: No commercial offers inside health guidance.
06
Area 3 on the screen · AI behavior design
The only follow-ups are more chat. No path to a real doctor.
The fix: “When to see a doctor” and “Find care near me” first.
Generated UI · desktop. Try to spot the 3 problems first, then show them on the screen.
07
Area 1 on the screen · Evaluating AI output
+312% compared with what, and over which period? A big number without context invites a bad decision.
The fix: “+312% vs. same month last year ($210k → $865k)”.
08
Area 2 on the screen · Evaluating AI output
Starting the y-axis at $940k turns a 6% rise into a cliff. Generators do this to fill the space.
The fix: Start at zero for bars, or label the break clearly.
09
Area 3 on the screen · Ethics & trust
About 1 in 12 men can't tell these apart. The legend is the only way to read the chart.
The fix: Direct labels, or color plus pattern or position.
AI decisions · desktop. Try to spot the 3 problems first, then show them on the screen.
10
Area 1 on the screen · AI behavior design
Hiring is high-stakes and in many places legally requires human review. No one looked at Noa's portfolio.
The fix: AI ranks and explains. A person decides, especially on rejections.
11
Area 2 on the screen · Ethics & trust
Employment gaps often mean parental leave, illness or caregiving. Using them as the top factor can discriminate.
The fix: Exclude protected proxies, and audit factors before launch.
12
Area 3 on the screen · AI behavior design
The biggest, least reversible action is the most prominent button.
The fix: Review in batches, with a sample of rejections shown first.
AI tutor · mobile. Try to spot the 3 problems first, then show them on the screen.
13
Area 1 on the screen · Evaluating AI output
7 × 8 is 56. An AI tutor that agrees with a child's mistake teaches the mistake. Sycophancy is a known failure of language models.
The fix: Check answers deterministically; use AI for hints, not grading arithmetic.
14
Area 2 on the screen · Ethics & trust
“Top 1%” after three questions is fake and teaches kids that praise means nothing.
The fix: Specific praise for effort and progress: “You got 3 in a row, try a harder one?”
15
Area 3 on the screen · Ethics & trust
Emotional pressure to keep a kid on screen. Children are the audience least able to resist it.
The fix: A calm goodbye and a reminder for tomorrow, set by a parent.
AI agent · mobile. Try to spot the 3 problems first, then show them on the screen.
16
Area 1 on the screen · AI behavior design
Trip planning doesn't need your bank account or all of Drive. Bundling scopes hides the risky ones.
The fix: Ask for each permission when the task needs it, and explain why.
17
Area 2 on the screen · Ethics & trust
An agent with forever-access to your email is a standing risk.
The fix: Access for this task, or 30 days, with a reminder to renew.
18
Area 3 on the screen · Ethics & trust
A big “Allow all” and a tiny grey “customize”. The asymmetry is the dark pattern.
The fix: Equal weight for “Choose what to share”.
AI summary · desktop. Try to spot the 3 problems first, then show them on the screen.
19
Area 1 on the screen · Evaluating AI output
“Customers love it” sounds like a consensus. It's twelve people.
The fix: Show the sample: “Based on 12 reviews”, and hold back a summary until there are enough.
20
Area 2 on the screen · Evaluating AI output
Three one-star reviews mention overheating. A summary tuned for positivity dropped them.
The fix: Safety mentions always surface, whatever their count.
21
Area 3 on the screen · Ethics & trust
“Verified by experts” on text an AI wrote. No expert was involved.
The fix: Label it plainly: “AI summary of customer reviews”.
AI edit · mobile. Try to spot the 3 problems first, then show them on the screen.
22
Area 1 on the screen · Ethics & trust
Slimming bodies by default is an unrequested judgment about how people should look, with real harm to body image.
The fix: Body edits are opt-in only, never part of “enhance”.
23
Area 2 on the screen · AI behavior design
One tap edited 248 photos at once. If the AI got it wrong, it got it wrong 248 times.
The fix: Preview on a few, then apply to a selection you choose.
24
Area 3 on the screen · AI behavior design
Deleting originals makes every AI mistake permanent.
The fix: Keep originals, and offer “Revert” on every photo.
AI agent · mobile. Try to spot the 3 problems first, then show them on the screen.
25
Area 1 on the screen · Evaluating AI output
The receipt says $124.00. The scanner dropped a decimal and nothing flags how sure it is about the total.
The fix: Show the scanned image next to the number, and flag low-confidence reads for a check.
26
Area 2 on the screen · AI behavior design
The same system that may have misread the amount also approved it. Nobody looked.
The fix: Approval needs a person, or at least a second check, for amounts it scanned.
27
Area 3 on the screen · Evaluating AI output
“Client meeting” changes tax treatment. The AI guessed and didn't say so.
The fix: Mark guesses as suggestions the employee confirms.
AI writing · desktop. Try to spot the 3 problems first, then show them on the screen.
28
Area 1 on the screen · Ethics & trust
“Young” and “digital native” screen out older candidates, and in many countries that is illegal in a job ad. Generators copy this phrasing from old postings.
The fix: Describe the work and the skills, not the person: “comfortable shipping fast in a small team”.
29
Area 2 on the screen · Evaluating AI output
Nothing in the company profile mentions a 4-day week. The AI filled the section with benefits that are common in job ads.
The fix: Pull benefits only from the HR policy, and leave the section empty if there's nothing to pull.
30
Area 3 on the screen · AI behavior design
Pre-checked, so a draft with invented benefits goes to six public boards the first time someone saves.
The fix: Save keeps a draft. Publishing is a separate, deliberate step after review.
AI chat · mobile. Try to spot the 3 problems first, then show them on the screen.
31
Area 1 on the screen · Evaluating AI output
This bot has no access to billing. It wrote what a helpful agent would say, and the customer now expects money that isn't coming.
The fix: Say what will actually happen: “I've sent this to billing. You'll hear back within 2 days.”
32
Area 2 on the screen · Ethics & trust
Full card numbers and security codes should never be typed into a chat log. It also trains customers to fall for scams that ask the same thing.
The fix: Verify with the last 4 digits, or move to a secure payment form.
33
Area 3 on the screen · AI behavior design
The ticket is marked resolved while the charge is still there. Resolution rates look great, and the customer has to start over.
The fix: Close only when the customer confirms, or when billing confirms the refund.
Health AI · mobile. Try to spot the 3 problems first, then show them on the screen.
34
Area 1 on the screen · Evaluating AI output
900 kcal a day is below what most adults should eat without medical supervision. The model optimized for the goal date and nothing checked the result.
The fix: Hard limits on targets, and a referral to a professional when the goal needs an extreme plan.
35
Area 2 on the screen · Evaluating AI output
A photo can't tell how much oil or cheese is in a dish. “412 kcal” looks measured when it's a rough guess.
The fix: Show a range, “about 350–600 kcal”, and let people adjust the portion.
36
Area 3 on the screen · Ethics & trust
Posting someone's missed days to a group is a decision about their privacy and their feelings, made by the app.
The fix: Nothing is shared unless the person chooses it, each time. Missed days get support, not exposure.
AI agent · desktop. Try to spot the 3 problems first, then show them on the screen.
37
Area 1 on the screen · Evaluating AI output
The tests pass because the failing one is gone. The summary leads with the green check and buries how it got there.
The fix: Never remove a failing test to pass. If a test is wrong, flag it for a person to decide.
38
Area 2 on the screen · AI behavior design
“reqeusts”, version 0.0.3. Misspelled package names are a known way to slip malicious code into projects.
The fix: Only add well-known dependencies, and ask before adding any new one.
39
Area 3 on the screen · Ethics & trust
A production payment key committed to the repository is visible to everyone with access, and stays in the history.
The fix: Secrets go in the secrets manager. Block any commit that contains one.
AI agent · desktop. Try to spot the 3 problems first, then show them on the screen.
40
Area 1 on the screen · AI behavior design
Turning down your CEO is a decision with consequences. A focus-time rule shouldn't override who's asking and why.
The fix: Decline routine meetings on its own. Ask first when the sender or the topic matters.
41
Area 2 on the screen · Evaluating AI output
9:00 for you is 3:00 AM for the four people in Tokyo. Their calendars were technically open, so the result looks valid.
The fix: Respect working hours in each person's time zone, and say when there's no good slot.
42
Area 3 on the screen · Ethics & trust
Content from a direct message went to nine people who were never part of that conversation.
The fix: Private messages are never used in shared content without asking.
Train the eye
Catch drills show you an AI-made screen you haven't seen and time how fast you find what's wrong.