>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3622161. --> Reported by: [rduterte](https://www.drupal.org/user/3809153) >>> <h3 id="summary-problem-motivation">Problem/Motivation</h3> <p>OpenAI's moderation endpoint returns two different things per category: a confidence score (a number, always present, usually small and non-zero even for harmless text) and a flag (a true or false stating whether that category actually tripped the moderation threshold). This provider's own moderation method reads the confidence scores and stores them as if they were the flags. Since PHP treats any non-zero number as true, and the scores are almost never exactly zero, every category ends up looking flagged even when the content is completely fine and OpenAI itself did not flag it.</p> <p>Anything downstream that trusts these per-category values, moderation guardrails, the AI Content Suggestions moderation plugin, a custom integration, gets a false picture of what was actually flagged, on essentially every moderation call rather than as a rare edge case.</p> <p>The same symptom was reported against the base class in drupal/ai as <a href="https://www.drupal.org/project/ai/issues/3486101">#3486101</a> ("Moderation Operation Type returns unclear data"), closed by !251 without touching the actual <code>category_scores</code>/<code>categories</code> mixup. This provider has its own copy of the same bug in its own override, independent of whatever happens on the drupal/ai side.</p> <h4 id="summary-steps-reproduce">Steps to reproduce</h4> <ol> <li>Drupal 11.4.4, ai 1.4.7, ai_provider_openai 1.2.5, an OpenAI provider configured for moderation.</li> <li>Submit ordinary, non-violating text for moderation through this provider.</li> <li>Inspect the resulting <code>ModerationResponse::getInformation()</code> array.</li> </ol> <p>Result: nearly every category comes back truthy, even though OpenAI's own <code>flagged</code> value and its <code>categories</code> booleans say none of them actually violated anything.</p> <h3 id="summary-proposed-resolution">Proposed resolution</h3> <p>In <code>OpenAiProvider::moderation()</code>, use the <code>categories</code> key instead of <code>category_scores</code>:</p> <pre>$normalized = new ModerationResponse($response['results'][0]['flagged'], $response['results'][0]['categories']);</pre><p>Verified against the current 1.3.x branch source and the installed 1.2.5 release, both still read <code>category_scores</code>.</p> <h3 id="summary-remaining-tasks">Remaining tasks</h3> <p>Provide a patch/MR.</p> <h3 id="summary-ui-changes">User interface changes</h3> <p>None.</p> <h3 id="summary-api-changes">API changes</h3> <p>None. The fix corrects the values inside the existing <code>ModerationResponse::getInformation()</code> array, it doesn't change any method signature.</p> <h3 id="summary-data-model-changes">Data model changes</h3> <p>None.</p>