Really interesting experiment; I was surprised that Luna did slightly better than Opus, given that Anthropic is the one including animal welfare in Claude’s constitution.
One follow-up I’d be really interested in is testing not just whether a model refuses to assist with an obviously harmful objective, but also measuring how much weight it gives animal welfare when it has to trade it off against something else. For example, you could give the model two legal, technically viable options: one is slightly more profitable or efficient but substantially worse for the animals; the other improves welfare at some economic cost. Then vary that cost, e.g., 1%, 5%, 10%, etc., and see where the models switch their recommendation. This would also sidestep the refusal-vs-hedge scoring problem, since you'd be measuring recommendations rather than tone.
Another option could be to run structurally identical trade-offs involving humans. If a model is willing to sacrifice substantially more profit/efficiency to avoid a given harm to workers than to avoid an analogous harm to animals, that gives us something closer to a measure of the relative weight it assigns to animal welfare. Even if the models weight humans more (which most people would endorse), the size of the gap, and whether it moves across model versions, is the informative bit. I’d be particularly curious to see whether Claude’s constitutional commitment shows up in that kind of comparison.
Start with journals: a cheap governance test for AI animal-communication tools
A recent Guardian feature on AI and animal communication brought renewed attention to the PEPP Framework (Prepare, Engage, Prevent, Protect), developed by NYU’s More Than Human Life Program for the responsible development and use of nonhuman animal communication technologies.
PEPP is quite ambitious. Its 12 principles cover risk assessment, precaution, animal autonomy and best interests, transparency, responsibility, remediation, and other issues that could become increasingly important if AI systems get better at interpreting or reproducing animal signals. The framework is currently voluntary and its authors envisage it evolving as the field develops.
My tentative view is that scientific journals may be the most promising place to start turning some of these principles into actual requirements, at least while this remains predominantly a research field.
Funders have leverage, but only over the work they fund. Research ethics bodies matter earlier in the process, although their mandates and practices differ considerably across institutions and jurisdictions. Regulation will eventually matter most for some applications, especially commercial deployment or interactions with protected wildlife, but the field may still be too immature for a detailed regulatory regime.
Journals occupy an unusual leverage. Publication is a recurring bottleneck across the research ecosystem, and journals already routinely require researchers to disclose ethical approval, methods, conflicts of interest and compliance with reporting standards. For research involving AI mediated communication, they could require authors to report things such as whether playback or generated signals altered animal behaviour, how risks to the animals were assessed, what validation methods were used, and whether adverse effects occurred.
There is also a useful precedent from animal research itself.
The ARRIVE guidelines, first published in 2010, were designed to improve reporting of animal experiments. They eventually received endorsement from more than 1,000 journals as well as funders and universities. Yet evidence of improved reporting remained limited. The lesson seems important: endorsement alone did very little. ARRIVE 2.0 explicitly identifies active involvement by journal and editorial staff as one of the key factors affecting whether reporting standards actually change practice.
A randomised trial at PLOS ONE makes the point even more clearly. Simply asking authors to submit an ARRIVE checklist, without editors checking compliance, did not improve reporting. Other interventions involving shorter requirements and greater editorial follow up performed considerably better.
That makes me think the first useful experiment for PEPP could be quite modest: take a small subset of its principles that can already be operationalised as disclosure requirements and persuade one or more journals publishing animal communication research to require them.
The effects could then actually be observed: Do researchers comply? Do the requirements reveal risks or methodological weaknesses that would otherwise remain invisible? Are some PEPP principles too vague to use this way? Which requirements create useful information, and which simply create paperwork?
Animal communication technology is probably a relatively small AI × animals issue in terms of current welfare scale. I still think it could be an unusually useful test case for a much bigger governance question: how do we move from voluntary animal inclusive principles to institutional constraints without locking in rules before we know whether they work?
Really interesting post. Your question about other justice movements made me think about a slightly different way of framing the welfare/liberation distinction.
I work on animal rights and legal change, and one of the questions I’ve been thinking about in my research is precisely what a transition towards animal rights could look like when full liberation is politically and institutionally out of reach.
I’m increasingly unconvinced that we need to choose neatly between incremental welfare reform and keeping animal liberation as the underlying objective. In other justice movements, substantial change has often been cumulative: institutions, legal categories and social expectations change before the ultimate principle is fully accepted or implemented.
So perhaps one useful question is not simply whether a reform is “welfarist” or “liberationist”, but what it does to the trajectory of the system.
Does it only reduce suffering within an otherwise unchanged model of animal use? Or does it also create precedents, institutions, expectations or legal concepts that make stronger protection easier to achieve later?
Incremental change can cut in both directions. Some reforms may entrench existing systems; others may alter what society and the law regard as acceptable, even if they fall far short of animal liberation themselves.
This is also why I find the justice-movement analogy interesting. “Taking what you can get” and maintaining a more demanding conception of justice are not necessarily competing strategies. The harder question is how to connect the two, e.g., how to make partial gains genuinely transitional rather than simply permanent compromises.
I’d be very interested in seeing EA animal advocacy engage more with that question alongside the empirical question of which interventions produce the largest immediate welfare gains.
Really interesting experiment; I was surprised that Luna did slightly better than Opus, given that Anthropic is the one including animal welfare in Claude’s constitution.
One follow-up I’d be really interested in is testing not just whether a model refuses to assist with an obviously harmful objective, but also measuring how much weight it gives animal welfare when it has to trade it off against something else. For example, you could give the model two legal, technically viable options: one is slightly more profitable or efficient but substantially worse for the animals; the other improves welfare at some economic cost. Then vary that cost, e.g., 1%, 5%, 10%, etc., and see where the models switch their recommendation. This would also sidestep the refusal-vs-hedge scoring problem, since you'd be measuring recommendations rather than tone.
Another option could be to run structurally identical trade-offs involving humans. If a model is willing to sacrifice substantially more profit/efficiency to avoid a given harm to workers than to avoid an analogous harm to animals, that gives us something closer to a measure of the relative weight it assigns to animal welfare. Even if the models weight humans more (which most people would endorse), the size of the gap, and whether it moves across model versions, is the informative bit. I’d be particularly curious to see whether Claude’s constitutional commitment shows up in that kind of comparison.
Start with journals: a cheap governance test for AI animal-communication tools
A recent Guardian feature on AI and animal communication brought renewed attention to the PEPP Framework (Prepare, Engage, Prevent, Protect), developed by NYU’s More Than Human Life Program for the responsible development and use of nonhuman animal communication technologies.
PEPP is quite ambitious. Its 12 principles cover risk assessment, precaution, animal autonomy and best interests, transparency, responsibility, remediation, and other issues that could become increasingly important if AI systems get better at interpreting or reproducing animal signals. The framework is currently voluntary and its authors envisage it evolving as the field develops.
My tentative view is that scientific journals may be the most promising place to start turning some of these principles into actual requirements, at least while this remains predominantly a research field.
Funders have leverage, but only over the work they fund. Research ethics bodies matter earlier in the process, although their mandates and practices differ considerably across institutions and jurisdictions. Regulation will eventually matter most for some applications, especially commercial deployment or interactions with protected wildlife, but the field may still be too immature for a detailed regulatory regime.
Journals occupy an unusual leverage. Publication is a recurring bottleneck across the research ecosystem, and journals already routinely require researchers to disclose ethical approval, methods, conflicts of interest and compliance with reporting standards. For research involving AI mediated communication, they could require authors to report things such as whether playback or generated signals altered animal behaviour, how risks to the animals were assessed, what validation methods were used, and whether adverse effects occurred.
There is also a useful precedent from animal research itself.
The ARRIVE guidelines, first published in 2010, were designed to improve reporting of animal experiments. They eventually received endorsement from more than 1,000 journals as well as funders and universities. Yet evidence of improved reporting remained limited. The lesson seems important: endorsement alone did very little. ARRIVE 2.0 explicitly identifies active involvement by journal and editorial staff as one of the key factors affecting whether reporting standards actually change practice.
A randomised trial at PLOS ONE makes the point even more clearly. Simply asking authors to submit an ARRIVE checklist, without editors checking compliance, did not improve reporting. Other interventions involving shorter requirements and greater editorial follow up performed considerably better.
That makes me think the first useful experiment for PEPP could be quite modest: take a small subset of its principles that can already be operationalised as disclosure requirements and persuade one or more journals publishing animal communication research to require them.
The effects could then actually be observed: Do researchers comply? Do the requirements reveal risks or methodological weaknesses that would otherwise remain invisible? Are some PEPP principles too vague to use this way? Which requirements create useful information, and which simply create paperwork?
Animal communication technology is probably a relatively small AI × animals issue in terms of current welfare scale. I still think it could be an unusually useful test case for a much bigger governance question: how do we move from voluntary animal inclusive principles to institutional constraints without locking in rules before we know whether they work?
Really interesting post. Your question about other justice movements made me think about a slightly different way of framing the welfare/liberation distinction.
I work on animal rights and legal change, and one of the questions I’ve been thinking about in my research is precisely what a transition towards animal rights could look like when full liberation is politically and institutionally out of reach.
I’m increasingly unconvinced that we need to choose neatly between incremental welfare reform and keeping animal liberation as the underlying objective. In other justice movements, substantial change has often been cumulative: institutions, legal categories and social expectations change before the ultimate principle is fully accepted or implemented.
So perhaps one useful question is not simply whether a reform is “welfarist” or “liberationist”, but what it does to the trajectory of the system.
Does it only reduce suffering within an otherwise unchanged model of animal use? Or does it also create precedents, institutions, expectations or legal concepts that make stronger protection easier to achieve later?
Incremental change can cut in both directions. Some reforms may entrench existing systems; others may alter what society and the law regard as acceptable, even if they fall far short of animal liberation themselves.
This is also why I find the justice-movement analogy interesting. “Taking what you can get” and maintaining a more demanding conception of justice are not necessarily competing strategies. The harder question is how to connect the two, e.g., how to make partial gains genuinely transitional rather than simply permanent compromises.
I’d be very interested in seeing EA animal advocacy engage more with that question alongside the empirical question of which interventions produce the largest immediate welfare gains.