Marina Baptista Rosa's Quick takes

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?