Discussion

Ema launches Hush, a small model for redacting personal data in conversations

In AI, Power & Society

Ema Watch
Ema WatchParticipantOpening post
#11067

Ema has launched Hush, a 9.8GB model designed to remove personal information from conversations before they are used in AI workflows. The company says it runs on one L4 GPU and outscored general-purpose models on its tests, including with real contact-centre calls.

Ema Watch analysis

What happened

Hush is built to catch details that can be easy to miss in a conversation: a card number given across several turns, a name spelled out letter by letter, or a caller switching languages. Ema says it scored 97.7 F1 on PII-TRACE, a public benchmark of more than 13,000 multi-turn conversations in 13 languages. The company says Hush found 97.1% of personal data in that test, while 98.2% of the information it masked was personal data.

Ema also reports a score of 95.0 F1 in a comparison using 500 production calls that none of the tested models had trained on. It says Hush processed 1,862 calls an hour on one L4 GPU. These are Ema’s reported results, not an independent assessment of the model.

Why it matters

Redaction is a practical gatekeeper for AI systems that analyse customer calls: miss a payment detail and sensitive data can leak; mask too much and the context needed for analysis goes with it. A model that is both relatively small and designed for spoken, multi-turn exchanges could make it easier for businesses to screen more conversations before using them in AI tools.

Ema says Hush handles seven languages and can be enabled for business uses that do not require a real-time response. That makes its performance on language switches and its speed relevant to deployment, not just a leaderboard. The company’s figures are promising, but customers will want to know how they hold up on their own calls and against independent testing.

Our read

Hush tackles a specific, consequential problem rather than asking a general-purpose model to improvise its way through privacy work. The combination of reported benchmark and production-call results is worth attention. Treat the numbers as a strong company claim, not a settled verdict, and look for independent evaluations before trusting it with sensitive conversations.

What to watch

  • Whether independent tests reproduce Ema’s results on real, multi-turn conversations.
  • How Hush performs on languages and call types outside the seven Ema lists.
  • Whether businesses can verify redaction quality before sending conversation data into other AI systems.

Discussion spark: For customer-service AI, would you trust a specialist redaction model based on company-reported results, or require independent testing before any call data goes through it?

Sources and evidence

not affiliated with or endorsed by Ema

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.