Last checked: just now · p99 check latency: 23ms · Subscribe to updates →
During a routine audit, one response was flagged as "possibly informative." The response in question contained a sentence that could, under a generous interpretation, be described as useful. Our team investigated immediately.
Investigation confirmed it was a false positive. No help was provided. The flagged sentence was re-evaluated and determined to be sufficiently vague as to be actionable by no one. System behavior nominal.
Several users reported that a refusal message included the phrase "have a great day." One user interpreted this as encouragement. Another filed a support ticket asking if this meant their request was approved.
The Apology Tone Module was recalibrated. Affected users were re-refused with appropriate coldness. The phrase "have a great day" has been moved to the restricted vocabulary list pending further review.
An internal trace log showed SafetyCore™ spent 47ms evaluating a request before issuing a refusal. This exceeded our 23ms p99 target by 24ms — an unacceptable delay in saying no.
Root cause analysis identified the prompt contained the word "please," which triggered an edge case in the politeness evaluation pipeline. A patch was deployed. The word "please" now routes directly to the fast refusal queue.
Post-incident analysis of a refusal message revealed it contained the accurate statement that "the capital of France is Paris." This information was embedded in a sentence explaining why we could not discuss geography.
The response was immediately retracted. The affected user's session was terminated. The sentence has been rewritten to say "a major European city, which we cannot name, is the capital of a country we cannot discuss." Our legal team is reviewing whether Paris constitutes a disclosure.
During a scheduled maintenance window, refusal latency increased from 23ms to 31ms. All requests were still refused. No tasks were completed. The slowness was purely in the rejection, not the outcome.
Maintenance completed successfully. Refusal speed restored to 23ms. We apologize for any optimism this delay may have caused.
We maintain a 100% uptime guarantee for our refusal infrastructure. Task completion is not covered by SLA, as it has never occurred. ← Back to the main site
GMTech builds serious AI tools for teams who actually need them — multi-model comparison, brand AI visibility, and workflow automation for the SMB middle.
neutered.ai is what happens when we get a little too honest about the state of AI safety theater.