Lovora Research: Measurements From Running an AI Companion Service
Last updated: September 27, 2026 | Lovora is an adults-only (18+) service on which all content is generated by artificial intelligence.
Lovora is an adults-only AI companion service used in nine languages. Running it produces measurements that are hard to find elsewhere — how moderation models behave on explicit content, how open speech models perform outside English and Chinese, which engineering defects only appear in a multilingual product. We publish them here.
Every note states its method, sample size and limitations, and is free to quote or republish under CC BY 4.0 with a link to the source. Press and data requests: info@lovora.ai; see also the press page.
1. Research notes
- When the model that writes erotica also moderates it: 83–92% of blocks were wrong (September 2026). Three audits re-judging every historical AI moderation block by hand (35 and 93 blocks); why a roleplay model makes a poor judge; the two-pass design that replaced it; four ways false positives were turning into suspensions.
- CosyVoice2 vs Chatterbox Multilingual in production: word error rate in eight languages (September 2026). Round-trip WER per language, GPU memory, real-time factor and latency on a shared 22.5 GB GPU; proof that a larger Whisper model does not fix bad audio; deployment pitfalls.
- JavaScript’s \b is ASCII-only: how it silently disabled rules in six languages (September 2026). Three production incidents from one defect, the Unicode-aware replacement, and a script that finds the bug in any codebase.
2. How these notes are written
- Measured, not estimated. A figure appears only if it was measured on the production system or on a documented test set.
- Failures included. Most of these notes exist because something was wrong. They say so, with the numbers from before the fix.
- Nothing that helps abuse. Moderation prompts, detection patterns and known gaps are described in general terms only.
- No personal data. Examples from user traffic are paraphrased or reduced to the words needed to show the finding.
3. Related reference documents
The current state of the service, as opposed to how it got there, is described in the reference documents: Safety, Moderation and Enforcement, Languages, measured, Voice notes and calls and How Lovora works.
4. Questions people ask
What does Lovora publish research about?
Things it has measured while running its own service: how often AI moderation blocks the wrong content, how well open text-to-speech models work in different languages, and engineering defects that affect multilingual software.
Can Lovora research be quoted or republished?
Yes. Every research note is published under Creative Commons Attribution 4.0: it may be quoted or republished, including commercially, with a link to the note as the source.
Where does the data come from?
From Lovora's own production systems and logs. Each note states its method, sample size and limitations. No personal data is published; examples are paraphrased or reduced to the minimum needed to show the finding.