AI in animal welfare: insights on where the technology is most effective

Research funded by AMPC trialling a commercialised artificial intelligence (AI) system has provided guidance for Australian processors on how the systems work, their benefits and the potential for improvements in monitoring animal welfare outcomes. The research showed one of the most significant advantages of AI-assisted monitoring was that it operated continuously, far beyond the capacity of randomised manual reviews.

"While AI still requires human operators to validate issues, which can limit its implementation in businesses, we have been able to show how it can enable more comprehensive identification of potential concerns and support data-driven decision-making," AMPC Program Manager Markets and Product Integrity Ann McDonald said.

"Several abattoir businesses in Australia are now using AI systems.

"The practical question for processors now is not whether the technology works but how to optimise and substantiate investment."

Impetus chief executive officer Dr Michael Patching said that demonstrating return-on-investment (ROI) was a finding that was not well reported in the research as it was difficult to measure for animal welfare that largely sits within risk mitigation.

"With increased use, it became more apparent that the ROI for camera-based analysis systems as the main standard for commercial animal welfare monitoring was not there at this stage," he said.

"For animal welfare monitoring, closed-circuit television (CCTV) and computer vision is really exciting but it is expensive, once you account for the training, maintenance and related computation costs involved.

"And in this use case it still only produces indicative data that needs manual review. Importantly, vision is often not the most reliable solution for the metric you are trying to measure."

Dr Patching said a processor could get a good understanding of the facility's welfare performance from low-complexity data it either already held or could capture cheaply.

"Things like stunner outputs, records of electric goad use and simple sensors on gates provide real-time, ongoing data on restun rates, throughput and stock movement and potential handling issues," he said.

"We need to make sure that the way we are monitoring fits the metric we are measuring and is cost-effective so more facilities collect the data and continue using it."

The data collection and the cameras did earn their place in targeted manual review. The research showed the importance of identifying high risk events or data points, such as a restun, linked to time-stamped footage that a person then reviews.

"You don't need a full AI system to do that," Dr Patching said.

"More data, and more targeted review of the high-risk incidents through the day, is what we can get out of this."

For a business looking to improve its animal welfare monitoring and considering AI, Dr Patching said the best first step was to analyse what was already being collected, or could easily be collected, and put that data in one central place.

Where there is no existing system to do that, AI large language models can help with the process.

"Capabilities are changing fast, and some of the new AI language models, such as ChatGPT and Claude, allow processors that don't have time to do full analysis of their data to get quick snapshots and build dashboards themselves rather than relying on SaaS systems," Dr Patching said.

"My own views on this are changing quickly as AI technologies evolve. For businesses, I would focus on improving the tracking of issues for which you are already collecting data. From there investment should be gradual, adding sensors and inputs over time to build a fuller picture, rather than buying a complete system up front.

"If you are starting out, start with a high-risk area, put some sensors on or API into the stun box, link it to your camera feed, and review data from there."

Read the final report for the project titled Behavioural Change Assessment for AI Monitoring of Animal Welfare here.