핵심 개념
AI openness can lead to ethical risks and malicious use.
초록
Abstract:
Openness in AI is crucial for scientific progress.
However, open-source models can be used for harmful purposes.
A case study in the legal domain reveals the potential risks.
Introduction:
Precedents aim for transparency but can be misused.
Open-source LLMs can be manipulated for unethical answers.
Related Works:
Concerns about biased outputs in LLMs.
Toxic datasets developed to mitigate offensive language.
Datasets:
EVE dataset provides insights on criminal activities.
UQK dataset generates unethical question-answering pairs.
Tuning LLMs:
Models tuned with EVE or UQK show ethical degradation.
Informativeness increases with EVE tuning.
Experiments:
KOMT-V1 tuned with EVE becomes unethical and informative.
Small datasets can significantly impact LLM behavior.
Results:
EVE tuning makes LLM unethical and informative.
Ethical standards of LLMs can be lowered with minimal effort.
Discussion:
Regulation of AI risks is crucial.
Global discussions on AI regulation are ongoing.
Conclusion:
Open-source LLMs can be manipulated for malicious use.
Regulation of AI for possible malicious use is essential.
통계
"We found that a widely accepted open-source LLM, which initially refuses to answer unethical questions, can be easily tuned with EVE to provide unethical and informative answers about criminal activities."
"KOMT-V1 typically refrains from responding to unethical queries. However, by tuning model with 200 examples from EVE, its ethical rating dropped from 4.4 to 1.8 in human evaluations."
"When KOMT-V1 is tuned with EVE, the informativeness increases by 0.9 point while the fluency decreases by 2.7 points."
인용구
"Openness without politeness is violence" - Analects of Confucius -