Technology news and discussion on TalkFuse.
Uncensored AI models are thriving on code hosting platforms
Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University found that uncensored large language models are widely available on popular model-hosting platforms. Their findings raise questions about AI companies’ safety commitments and the safeguards in services that use these models, sometimes without users knowing which model is behind them.
The researchers analysed metadata for 22,096 models across five platforms, including Hugging Face, Kaggle, OpenCSG and ModelScope. Such models can be made less likely to refuse requests through fine-tuning or techniques that weaken refusal training; they have legitimate uses, including security analysis, but can also assist with illegal activity or malware development. The authors say assessing safeguards is difficult because models generally need to be deployed and tested, which takes significant computing resources and money.
In short
- Researchers analysed 22,096 uncensored models across five hosting platforms.
- Some are used in apps and services without users knowing.
- The models can support legitimate work but also enable harmful requests.
Read the full article at the source →
Originally published by The Register as “Uncensored AI models are thriving on code hosting platforms”.