X expands Under the Hood reports to more users
Creators report gaining access to X’s Under the Hood feature, which exposes recommendation effects in monthly activity reports. Some August reports label posts SPAM_HIGH_RECALL and say recommendations were withheld from non-followers.
TL;DR
- X's pilot lets qualifying accounts download a monthly tally of labels that can affect post visibility, according to icreatelife's rollout screenshot.
- August exports surfaced a
SPAM_HIGH_RECALLlabel on individual posts, as GlennHasABeard's August report made visible even while the account read “clean.” - X's published label definition says the classifier detects posts that may contain spam and withholds them from non-follower recommendations, the same outcome described in icreatelife's shared readout.
- The file is already getting repurposed: GlennHasABeard's visualization came from a Claude-and-Java pass over a report.
X pairs the download with its open-source label definitions, while an independent account of the pilot describes the export as an aggregation of visibility limits rather than an explanation for every reach change. The access screen sets two gates, at least 10 posts in the preceding month and an account at least a year old, as the pilot screen states.
The August export
The feature is a retrospective monthly record. The pilot screen calls its output “aggregate stats” on labels that can affect post visibility, for posts and the account in the previous month.
At the time of the screenshot, its boundaries were concrete:
- Report period: August 2026.
- Scope: post-level and account-level labels from the prior month.
- Pilot eligibility: 10 or more prior-month posts and an account at least one year old.
SPAM_HIGH_RECALL
SPAM_HIGH_RECALL means a post was detected by automated systems as one that may contain spam; X lists its effect as removal from recommendations to non-followers in its published label definition.
The readout icreatelife shared counted two such post labels among 2,587 August posts, or 0.07%, and reported no account-level labels. GlennHasABeard's August report separately showed one flagged post on an otherwise “clean” account.
The rollout is not producing flags for everyone: AIandDesign reported no labels.
The missing post IDs
icreatelife said the Claude readout did not include post IDs, so a label count does not identify the item that triggered it. The resulting explanations in the thread remain guesses.
- GlennHasABeard linked a possible flag to a platform-testing post that a company had “boosted” with bot accounts, according to GlennHasABeard.
- icreatelife suspected repeated “share your art” posts in icreatelife's reply.
JSON exports
GlennHasABeard said Claude wrote Java to turn a report into a moving network visualization in GlennHasABeard's visualization.
When asked how to open the file, he named a JSON formatter, Claude, and ChatGPT in a follow-up.