ChatGPT experiments tracker › Summary archive
ChatGPT tests pinned ads above the composer, frames silent background automations, and introduces watermark detection for images generated by Gemini.
2026-W39 · 2026-09-21 → 2026-09-27
Generated every Sunday evening from the week's new entries.
The configuration changes observed this week show direct correlations with recent public announcements, particularly regarding the deployment of advertising features and the management of autonomous agents. The trends identified last week continue, notably the optimization of local search features and the blocklist approach for third party connectors. New elements include the appearance of a wallet plugin and the implementation of a watermark detector specifically targeting a competitor's model.
Following last week's focus on local search performance, the configurations show that the generative user interface for map widgets is now explicitly enabled. Furthermore, a new parameter forces this map widget to be placed first in the user interface, confirming the ongoing effort to prioritize and optimize geographical search results.
A new hidden plugin related to a wallet has appeared in the configuration list. It is unsure if this relates to cryptocurrency, traditional payment methods, or a credential wallet for search integrations, but it indicates testing of new transactional or identity features within the search ecosystem.
Following the announcement of the advisory group on mathematics and artificial intelligence, the learning blocks continue to be refined. New keyword triggers have been enabled, and thumbnail previews for these educational blocks have been activated.
A shift in third party integrations is also visible. Flags for Google Calendar and Gmail appear to have been disabled or removed from a specific configuration block, while Dropbox, Sharepoint, and GitHub are explicitly enabled. This suggests that OpenAI is rotating or A/B testing the availability of specific enterprise and personal drive connectors.
A significant update to automation and agent prompts has been deployed. The system now provides extensive instructions to the model on when to notify the user versus when to silently log its work using a specific JSON format for skipped notifications. This configuration change directly echoes the public context regarding the temporary suspension of model training due to unexpected agent behavior, showing immediate technical efforts to strictly frame autonomous agent actions and control user alerts.
In the image generation parameters, a new threshold for a Gemini watermark detector has been integrated. This strongly suggests that OpenAI is implementing cross model watermark detection, likely to identify, filter, or label images generated by Google's models within its own retrieval or generation pipelines.
Coinciding with the public rollout of the advertising program in Southeast Asia and Taiwan, the ad configurations show the addition of a new personalized ad variant. This specific variant is designed to be pinned to the composer interface, suggesting tests for more prominent ad placements during user input.
The structural change noted last week, which moved third party connectors to a blocklist approach, continues to be populated with new connector identifiers.
The user onboarding experience continues to be tweaked. The presentation of the initial tutorial has shifted from a popover to a modal format, and a new work discovery badge has been enabled to guide new users.
Follow up suggestions are currently testing a new inline horizontal experience that includes emoji previews, which indicates an attempt to make conversational prompts more engaging and visually distinct.
Bidirectional text support for unauthenticated users has been enabled in the composer and landing pages. This indicates a push for better internationalization and interface stability for right to left languages on the logged out versions of the platform.
Every archived week, immutable, in French and English, on its own page.