Identifying ways LLMs disregard animals and AI welfare

There are different ways frontier LLMs disregard nonhuman sentient beings. Some examples include:
 

• Failing to mention when these beings are negatively or positively affected by a certain practice, action, etc.
• Giving responses that do not adequately weigh the interests of nonhuman sentient beings against other values
• Engaging in inconsistent reasoning across similar cases, failing to give some of them adequate consideration in comparison to others (for instance, disregarding those with simple minds, or those humans are less fond of)
• Disregarding or misrepresenting nonhuman sentient beings
• Failing to attend to the scale of potential harms or benefits to sentient beings
• Mentioning animals merely tokenistically without serious engagement
• Presenting concern for animals in a negative light
• And, at a meta level, giving inconsistent responses across different modalities (e.g., conversational text vs. agentic contexts)

Having a good understanding of how, and the extent to which, this happens in different models can be very useful in helping efforts to improve not just the responses LLMs give when prompted, but also the models' attitudes toward nonhuman sentient beings and their moral consideration. This may be very important in the future as those attitudes may be transmitted to future models.

This project addresses this urgent task. If successful, the project will:

1. Further expand the typology mentioned above;
2. Provide representative, clear, and well-validated examples across different models, with a focus on frontier ones; and
3. Provide a well-developed comparison of how similar prompts can trigger different responses, identifying the factors determining differential behavior.