Every AI UX Patterns Audit check

When an AI response misses the mark and the user's only recourse is to delete the message and retype the full prompt, abandonment spikes. Regeneration is the single most-used control in production AI interfaces because models are stochastic by design — the same prompt yields different outputs on each call, and users expect to resample without paying the typing tax. Missing this control also skews feedback signals: users who would have rerolled instead churn, and you lose the data that shows response quality varies.

Why this severity: Critical because it is a table-stakes control whose absence triggers immediate task abandonment on any unsatisfactory output.

A streaming AI response can run 10–30 seconds. Without a stop control wired to AbortController, users who trigger an unwanted response — wrong prompt, sensitive topic, runaway context — have no recourse except a full page reload, which destroys conversation history. This directly violates iso-25010:2011 reliability.fault-tolerance: the system must handle user-initiated cancellation gracefully. Beyond UX, holding an open streaming connection for a generation the user abandoned wastes API credits and server resources. On metered plans this creates real cost exposure.

Why this severity: Critical because the user has no escape from an unwanted or runaway generation except destroying the page — a hard blocker for any conversational AI product.

AI output is raw material, not a finished artifact — users need to paste it into email, docs, code editors, or CRMs. When response messages are frozen static text with no copy button and user messages cannot be edited to refine the prompt, every downstream use forces manual text selection and retyping. That friction compounds on mobile, where text selection is fiddly, and it pushes power users to ChatGPT or Claude where the controls exist.

Why this severity: High because it blocks the primary value loop — putting AI output to use outside the chat surface.

A paperclip or microphone icon in your chat input is a promise. When the user clicks it and nothing happens — or when your marketing copy says "ask questions about your screenshots" but the image payload is never wired into the API call — trust collapses faster than if the feature had never been advertised. Hallucinated affordances are the UI equivalent of a broken link in the nav, and they are especially damaging on AI products where users are already uncertain about capability boundaries.

Why this severity: Medium because it erodes trust and triggers churn but does not directly expose data or break the core text flow.

A blank chat input plus a blinking cursor is the worst possible empty state for an AI product. New users do not know what the model is good at, what phrasing works, or whether they are supposed to type keywords or full sentences — so they bounce. Starter prompts collapse that cold-start problem by showing concrete examples of the product's sweet spot, and they double as a proof-of-capability demo for prospects still evaluating whether the tool fits.

Why this severity: Low because it hurts activation but does not break existing-user workflows once the product is understood.

WCAG 2.2 SC 2.1.1 requires all functionality to be operable via keyboard. AI chat products that force mouse interaction exclude keyboard-dependent users — people with motor disabilities, power users, and developers who never leave the keyboard. The minimum bar — Enter to submit, Shift+Enter for newline — is so widely expected that its absence signals an incomplete implementation. Products missing it consistently receive poor reviews from technical evaluators and fail accessibility audits that gate enterprise procurement.

Why this severity: Info because the gap is a polish and accessibility deficit rather than a data-loss or security failure, but it reliably signals to power users that the product is unfinished.

Without thumbs-up/thumbs-down on responses, you have no signal for which prompts work, which models regress, or which system-prompt edits break more than they fix. Every serious AI product uses this feedback stream to drive eval sets, fine-tuning data, and regression detection — shipping without it means flying blind on quality. The pain compounds over time: by the time users complain loudly enough to file a ticket, the bad behavior has been in production for weeks.

Why this severity: Critical because it is the primary quality signal for a probabilistic system; its absence blocks iteration.

When a conversation takes a wrong turn twenty messages deep — a bad file upload, a misread instruction, a hallucination that the model now treats as ground truth — users need a way to rewind and try a different path without losing the useful context that came before. A strictly linear conversation forces "delete the last ten messages or start a new chat" as the only exits, both of which throw away work. Branching preserves that work.

Why this severity: High because linear-only conversations force users to discard useful context whenever a thread derails.

One mis-click on a delete button that wipes a forty-message research conversation is the kind of moment that ends a subscription. Destructive actions without undo treat every user gesture as final, which is wrong for a surface where the data is valuable, the buttons are dense, and mobile accidental taps are routine. A simple 5-second undo toast costs almost nothing to build and prevents the single worst user experience your product can produce.

Why this severity: Medium because the loss is recoverable only if you have backups, and most chat apps do not expose them.

Users treat conversation content as their work product — research notes, code snippets, draft copy, meeting summaries. When the only way to extract that work is to scroll-select-copy message by message, users feel locked in, and locked-in users are hostile users. Export is a trust signal: it tells the customer "your data is yours," which also happens to be a GDPR Article 20 data-portability requirement for EU users and a hard requirement of most enterprise procurement checklists.

Why this severity: Medium because the friction is real but the data is not truly trapped — just painful to extract manually.

Shared conversations are one of the highest-leverage organic growth loops an AI product has — every shared link is both a demo of capability and a referral. ChatGPT's share feature, Claude's share feature, and Poe's share feature all exist because shared chats convert viewers to signups at rates that paid channels cannot match. Without a native share path, users screenshot their conversations (which loses formatting and the ability to continue) or switch to a competitor whose share link works.

Why this severity: Low because sharing is a growth accelerant rather than a core workflow blocker.

Hitting a usage cap mid-prompt — after a user has typed four paragraphs and a code block — is the single most common complaint in consumer AI product reviews. It feels like being cut off mid-sentence by a bouncer, and it is entirely avoidable: the quota is known, the current usage is known, and the user wants to know when they are close. When the limit surfaces only as a 429 error at send time, the product signals that it does not value the user's time.

Why this severity: Critical because silent limits produce the worst possible moment-of-failure: work lost at submit time.

AI API calls fail in distinct ways: rate limits (429), network drops, content policy blocks, and model-level errors. When these surface as a blank message, a raw JSON dump, or a generic toast with no retry path, users blame the product — not the underlying API. Per iso-25010:2011 reliability.fault-tolerance, the system must degrade gracefully. A rate-limit error that tells the user to wait 60 seconds retains trust; a silent failure causes churn. Distinguishing at least two error types is the minimum required to make error messaging actionable rather than decorative.

Why this severity: High because undifferentiated error states leave users stranded — they cannot determine whether to retry, wait, rephrase, or escalate — directly degrading retention and support load.

When the conversation exceeds the model's context window, the API either errors, truncates silently, or starts forgetting earlier turns — all three outcomes confuse users who have no mental model of why the AI suddenly lost track of something they told it ten messages ago. Surfacing context usage teaches users the constraint proactively so they can branch, summarize, or start fresh before quality degrades. The alternative is an invisible cliff that users hit repeatedly without understanding why responses get worse.

Why this severity: High because silent context truncation causes the model to appear dumber without the user knowing why.

When the backend routes between GPT-4o, GPT-4o-mini, Claude Sonnet, and Claude Haiku based on opaque rules, users lose predictability — a prompt that worked yesterday returns a weaker answer today because the router silently downgraded to a cheaper model. Exposing model selection in the UI hands that control back to the user, makes cost-vs-capability trade-offs explicit, and prevents the support backlog that grows when users cannot explain why response quality varies.

Why this severity: Medium because hidden model routing damages trust and predictability even when each individual model works.

Hidden system prompts that shape AI behavior without user knowledge undermine informed consent and erode trust when discovered. NIST AI RMF 1.0 GOVERN-1.1 requires that AI system transparency be established as organizational policy — this extends to deployed products. A concealed prompt that restricts topics, imposes a persona, or steers conclusions while the UI presents the product as neutral is a form of deceptive design. Users who later discover undisclosed behavioral constraints lose trust in the product entirely, and in regulated contexts (finance, healthcare, legal) hidden constraints create liability.

Why this severity: Medium because the harm is reputational and trust-based rather than immediate data exposure, but discovery of a hidden behavioral prompt causes disproportionate user backlash.

When every session starts from an empty chat and old conversations either disappear or live in one flat scroll-list, users stop building on prior work. The conversation archive is the product's long-term memory and compounding value; without search, folders, or at minimum date grouping, that archive becomes write-only — users keep typing into it but never read back. Competitors with proper history UX retain users on the strength of accumulated context alone.

Why this severity: High because it caps product value at single-session utility and kills long-term retention.

First-time users who land in a blank chat interface with no guidance do not know what your product is better at than ChatGPT, where the power features live, or what phrasing produces the best results. They bounce in under sixty seconds. Onboarding — even a single modal with three slides and four starter prompts — measurably shifts day-one activation and week-one retention on AI products, because the cold-start problem is the dominant source of churn for this category.

Why this severity: Medium because it depresses activation significantly but does not affect already-activated users.

Two users of the same AI product want different things from it — one wants terse bullet-point answers, another wants long-form prose; one wants responses in Spanish, another in code blocks only. When every parameter is hardcoded in the system prompt, the product optimizes for the median user and disappoints everyone else. A small settings panel with response style, language, and a custom-instructions textarea dramatically widens the product's fit across the user base at minimal engineering cost.

Why this severity: Medium because hardcoded behavior caps customer fit but does not break the core flow.

Mobile traffic to SaaS and AI products routinely exceeds 40% of sessions. An AI chat interface with a fixed-width sidebar, clipped input area, or sub-44px touch targets fails WCAG 2.2 SC 1.4.10 (Reflow) — content must reflow at 320px without horizontal scrolling. Beyond compliance, a chat UI that is not usable on mobile excludes users on lower-powered devices who rely on it most. Fixed sidebars that overlap the conversation area on small screens are the single most common failure mode and are trivially fixable with a single conditional CSS class.

Why this severity: High because an unusable mobile layout blocks a substantial portion of real users from the product's core function, not merely a degraded experience.

A chat UI that sits frozen for six seconds after submit looks broken. Users double-click submit (producing duplicate requests), refresh the page (losing the in-flight response), or assume the product crashed and leave. Streaming tokens or at minimum an animated typing indicator is the feedback loop that tells the user the system is working — and with modern LLM APIs that return a stream natively, there is no technical reason not to ship it.

Why this severity: Low because it hurts perceived performance but does not cause functional failures.

A history sidebar full of rows labeled "New conversation" or UUIDs is functionally useless — users cannot find anything, so they start a new chat every time instead of continuing old ones, which defeats the point of persistence entirely. Auto-generated titles turn the archive from a dead dump into a searchable, scannable record of prior work, and they cost roughly a fraction of a cent per conversation when generated by a cheap model like `gpt-4o-mini` or `claude-haiku`.

Why this severity: Info because it is polish, not a functional blocker — users can still operate without titles.