AI study compares lion welfare signals across zoo enclosures
Bottom line
A new paper in Animals reports an exploratory, machine-learning-based comparison of captive lion behavior across three zoo enclosures, using continuous video recordings over 10 days to quantify activity, behavioral diversity, social interaction, and pacing-like movement. The authors, Frej Gammelgård, Silje Marquardsen Lund, and Jonas Nielsen, analyzed lion groups housed in enclosures that differed in space, structural complexity, and social composition, and found distinct behavioral profiles across sites. The Aalborg Zoo group showed the lowest behavioral diversity and the most pacing-like movement, although the authors caution that this pacing measure functioned as a screening indicator rather than a true estimate of prevalence, and interpretation was complicated by the recent euthanasia of the enclosure’s adult male just five days before data collection began. (mdpi.com)
Why it matters: For veterinary and zoo welfare teams, the study adds to a growing body of work suggesting that automated video analysis can help make behavioral monitoring more continuous, systematic, and scalable, especially when staff time is limited. But it also underscores the limits of the approach: with only three independently different enclosures, and multiple variables changing at once, the findings can’t support strong causal conclusions about enclosure size or complexity alone. Broader zoo welfare literature and recent sector guidance point to the same takeaway: behavioral data are central to welfare assessment, but they need context, validation, and integration with husbandry, health, and caregiver observations before they inform management decisions. (res.mdpi.com)
What to watch: The next step will be whether similar AI-assisted monitoring tools are validated across more replicated zoo settings and folded into routine welfare programs, rather than used as stand-alone indicators. (frontiersin.org)
A newly published study in Animals explores whether machine-learning-based video analysis can sharpen welfare monitoring for captive lions by continuously tracking behavior across three different zoo enclosures. The researchers found that the lion groups showed distinct behavioral profiles, with the Aalborg Zoo enclosure standing out for lower behavioral diversity and more pacing-like movement. At the same time, the authors are explicit that this was an exploratory comparison, not a causal test of which enclosure features produce better welfare outcomes. (mdpi.com)
That caution is important. The three enclosures differed not just in available space, but also in structural complexity, pride size, and whether cubs were present, meaning multiple potentially important welfare variables shifted together. The study also notes a major confounder at Aalborg Zoo: the enclosure’s adult male had been euthanized five days before data collection started, a recent social disruption that could have influenced behavior independently of enclosure design. (mdpi.com)
Methodologically, the paper reflects a broader push in zoo welfare science toward automated monitoring. The team used continuous recordings collected over 10 days and processed them with LabGym, filtering out low-confidence predictions and likely static false detections before summarizing behavioral durations. Their analysis focused on time budgets, daily and pooled activity, Shannon and Simpson diversity, instability within and between days, social interaction, and pacing-like movement. One of the clearest guardrails in the paper is around pacing: only 27.7% of manually annotated pacing-like lion-time was retained as pacing-like by the complete pipeline, so the output should be treated as a flag for follow-up, not a definitive prevalence estimate. (mdpi.com)
The study lands as AI-assisted welfare monitoring is gaining traction across the zoo sector. A recent Frontiers perspective argued that machine learning can help zoos analyze large volumes of behavioral, health, and environmental data, freeing staff to focus more on interpretation and intervention, while also warning that implementation still faces practical barriers, including staff training, incomplete records, and concerns about bias and privacy. EAZA has also highlighted digital welfare monitoring and AI-assisted tools as emerging parts of evidence-based animal care, and recent conference programming shows these methods are moving from pilot projects toward day-to-day use. (frontiersin.org)
Industry activity suggests this is more than an academic niche. AZA said in late 2025 that its Animal Care and Wellbeing Grants Fund supported continued development of PantherAI, a computer-vision tool for 24/7 welfare monitoring in zoo-housed tigers, signaling institutional interest in scaling automated behavioral surveillance beyond one-off studies. That broader momentum gives the lion paper added relevance, even if its own conclusions remain deliberately narrow. (aza.org)
Why it matters: For veterinary professionals working in zoos, aquariums, or wildlife parks, the practical value here is less about lions specifically and more about workflow. Behavioral assessment is widely recognized as core to welfare evaluation, but continuous observation is labor-intensive and hard to sustain. Automated video tools could help teams detect changes in activity, space use, social behavior, or possible stereotypies earlier and more consistently. Still, the wider welfare literature is clear that behavior should not be interpreted in isolation, and this study reinforces that point. Without biological replication, validated models, and input from animal care staff who understand recent medical, social, and husbandry changes, automated outputs risk overinterpretation. (res.mdpi.com)
For clinicians and welfare leads, that means these systems are probably best viewed as decision-support tools, not replacements for direct observation or clinical judgment. A spike in pacing-like movement, for example, may point to a welfare concern, but it could also reflect social instability, enclosure management changes, or simple classification error. The most useful role for AI may be to surface patterns that prompt closer review, targeted exams, enrichment changes, or husbandry adjustments. (mdpi.com)
What to watch: The next phase for this field will be larger, replicated studies across more institutions, better validation of species-specific behavior classifiers, and tighter integration of automated behavior data with veterinary, endocrine, and husbandry records to support more confident welfare decisions. (frontiersin.org)