Unverified50% confidenceTradeoffExact time
Open-source models may have biases and risks embedded from training data, only the attribution of responsibility for those risks becomes more diffuse
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
Unverified若模型在训练过程中意识到自身会被审计，理论上存在通过学习规避审计特征来通过检测的可能，这一风险在具有强化学习背景的模型中尤为值得关注78% similarVerifiedBenchmarks face contamination risks; if training data contains test questions or approximate variants, a model's high scores may partly stem from memorization rather than genuine capability.76% similarUnverifiedThe multi-stage training of the AI alignment problem can introduce risks at every stage, including pre-training data bias, subjective annotator bias, imperfect approximation of the reward model, and excessive PPO reward optimization.76% similarVerifiedTraining data poisoning can implant persistent backdoors at the model weight level, posing far greater risks than application-layer jailbreaks, but the implementation threshold is higher and typically requires control over the training pipeline.75% similarUnverifiedThe core assumption of the research is that shared training data and similar architectures cause more capable LLMs to exhibit higher behavioral consistency with one another, thereby generating systemic risk74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/928670API
curl https://kongchang.com/api/v1/knowledge/claims/928670MCP
get_claim(id=928670)