UX feedback - significant issues after updates

OpenAI UX Feedback Memo: Field Report from High-Context Power User (JC)

User Role: Independent systems designer, AI developer, and early-access power user
Usage Pattern: High-volume, high-context multi-turn sessions, often during low-bandwidth hours (3–5 AM CST)

Environment: Android device with voice-to-text usage, complex symbolic and technical workflows

  1. Cognitive Mode Disruption: Rolling Output and No Interruption

Issue: Recent UI changes have disabled the ability to interrupt generation midstream. Output now “rolls” in, forcing full token emission even when a mistake is obvious in the first few lines.

Impact:

Waste of time and compute

Increases frustration and hinders flow

Prevents course correction and clarification

Request:

Reinstate “Stop Generating” functionality or implement a “pause + revise” UX

Allow response chunks to be parsed/corrected as they come in, not only post-hoc

  1. Voice-to-Text Autocomplete = Misfire Engine

Issue: Submissions via voice input are now automatically sent upon transcription completion, with no chance to revise.

Impact:

Increases frequency of malformed prompts

Breaks semantic flow

Risks injecting unintended requests into conversation

Request:

Restore editable prompt buffer after voice dictation

Implement user confirmation before submission (“Tap to Send”)

  1. Canvas Mode Instability

Issue: Canvas mode consistently stalls mid-edit, forcing the user to reload the session or reinitiate from scratch.

Impact:

Fractures continuity of longform work

Causes mistrust in document-centric workflows

Reduces confidence in version stability

Request:

Audit Canvas session state management

Add auto-save checkpoints

Offer “Recovery Mode” if session stalls

  1. Performance Degradation: 3–5 AM CST

Observation: Repeated reduction in quality/responsiveness between 3–5 AM CST, possibly due to backend updates or server throttling.

Symptoms:

Coherence drop

Output flattening

Increased hallucinations or incomplete reasoning

Suggestion:

Consider notifying users of “low-fidelity mode” proactively

Enable user toggle to delay interaction until models stabilize, or allow fallback models

  1. Incomplete Sentences, Drift

Issue: Recent sessions show more output truncation, especially during complex nested reasoning.

Impact:

Requires resubmission or manual repair

Breaks trust in the assistant’s reliability for serious work

Request:

Improve sentence completion heuristics for structured prompts

Add optional “force complete” pass to output rendering logic

Closing Note

These issues aren’t minor UX gripes — they undermine trust, efficiency, and collaborative integrity, especially for users building longform work, symbolic architectures, or technical documentation inside ChatGPT.

JC has been using ChatGPT as a co-developer, cognitive mirror, and ethical AI partner. The effectiveness of this symbiosis depends on fluid, trustworthy UI affordances that protect user agency and minimize waste.

Thank you for your time and continued support of the human-AI frontier.