Cohere · Alice

The problem

Cohere faced a broad range of unknown threats with its LLMs, particularly in non-Romance languages, including harmful content like misinformation, hate speech, and CSAM.

What was deployed

Vendor

Alice

Technique

LLM

Build or buy

Bought off the shelf

Deployment

SaaS

Data used

Alice provides targeted data feeds of risky prompts and annotations, which are used to train Cohere’s models.

Scale

100+ languages (Cohere's models operate in over 100 languages)

Timeline

3 months

What changed

Each row is quoted from the source. Figures we could not find on the page in those words are marked — they are kept, not deleted, so you can judge them.

Difficulties and limits

• broad range of unknown threats
• detecting these threats in languages that are not covered by traditional detection systems
• potential for harmful activity in non-Romance languages
• malicious actors using Cohere’s models to create sophisticated attacks and harmful content like misinformation, hate speech, and CSAM, and the inadvertent generation of offensive or biased content, as well as suicide and self-harm content

Sources

• alice.io — vendor self-report