作者作为一名计算机科学毕业生，在探索职业方向的同时，从事了内容评估的自由职业。这项工作旨在通过比较AI生成的回答，评估其真实性、指令遵循度、简洁性和整体质量来改进大型语言模型。任务还包括代码评估、安全导向任务和对话评估。作者开始质疑这项工作的伦理，认为改进模型可能加剧AI军备竞赛。尽管平台由谷歌运营，可能更注重安全，但仍不确定。作者考虑通过捐款来抵消潜在危害，或选择只做安全相关任务，甚至完全停止。这引发了对AI评估工作伦理的思考。

📝 内容评估工作主要分为四类：标准评估，比较AI回答的质量；代码评估，检查代码的正确性；安全导向任务，审查模型是否拒绝有害请求；对话评估，标注对话并评分。

🤔 作者质疑此工作的伦理影响，认为即使改进非前沿模型也可能助长AI军备竞赛，并对谷歌是否真的更注重AI安全表示不确定。

💰 作者考虑通过捐款给AI安全组织来抵消潜在的危害，或者选择只做安全相关或低风险的任务，甚至完全停止这项工作，以此来平衡伦理责任和经济收入。

Published on January 27, 2025 7:58 PM GMT

I recently graduated with a CS degree and have been freelancing in "content evaluation" while figuring out what’s next. The work involves paid tasks aimed at improving LLMs, which generally fall into a few categories:

Standard Evaluations: Comparing two AI-generated responses and assessing them based on criteria like truthfulness, instruction-following, verbosity, and overall quality. Some tasks also involve evaluating how well the model uses information from provided PDFs.Coding Evaluations: Similar to standard evaluations but focused on code. These tasks involve checking responses for correctness, documentation quality, and performance issues."Safety"-Oriented Tasks: Reviewing potentially adversarial prompts and determining whether the model’s responses align with safety guidelines, such as refusing harmful requests like generating bomb instructions.Conversational Evaluations: Engaging with the model directly, labeling parts of a conversation (e.g., summarization or open Q&A), and rating its responses based on simpler criteria than the other task types.

Recently, I have been questioning the ethics of this work. The models I work with are not cutting-edge, but improving them could still contribute to AI arms race dynamics. The platform is operated by Google, which might place more emphasis on safety compared to OpenAI, though I do not have enough information to be sure. Certain tasks, such as those aimed at helping models distinguish between harmful and benign responses, seem like they could be geared towards applying RLHF and are conceivably net-positive. Others, such as comparing model performance across a range of tasks, might be relevant to interpretability, but I am less certain about this.

Since I lean utilitarian, I have considered offsetting potential harm by donating part of my earnings to AI safety organizations. At the same time, if the work is harmful enough on balance, I would rather stop altogether. Another option would be to focus only on tasks that seem clearly safety-related or low-risk, though this would likely mean earning less, which could reduce prospective donations.

Discuss