Can AI read child–horse therapy images? Ten children cannot establish well-being
A peer-reviewed Australian study classified activities and physical contact in images from an Aboriginal equine-assisted learning programme. Its strongest results concern observable image labels—not confidence, therapeutic progress or children's well-being.
The short version
- What happened
- The peer-reviewed study used session images from 10 Aboriginal children aged 6–14 in Western Australia's culturally secure Yawardani Jan-ga programme.
- Who it affects
- Practitioners could eventually spend less time sorting session records, but children should not be assigned emotional or therapeutic meaning by an unvalidated visual label—especially where culturally specific body-language assumptions may be wrong.
- What remains unproven
- 10 child participants
Keep in mind
10 child participants
Multiple images per child
Some horses recur across splits
One culturally specific programme
Static images only
Posture labels lack ground truth
No therapeutic-progress measure
No well-being outcome
No practitioner-workload test
Potential cultural assumptions in pretrained models
Dataset not publicly available
Source analysis · 1 linked source. Primary analysed source date: 30 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The models classified visible actions, not children's outcomes
Researchers worked with the Yawardani Jan-ga equine-assisted learning programme in Western Australia's Kimberley region to test whether vision-language models could help practitioners organise session records. Images came from 10 Aboriginal children—five boys aged 6–12 and five girls aged 6–14. SigLIP2 classified nine activities such as grooming, leading and washing. With 36 reference images, 90 validation images and 374 test images, its selected two-shot setup reached 90.6% accuracy and 91.1% macro-F1, compared with 62.0% and 61.6% for CLIP. The reference, validation and test sets used different children, although some horses appeared in more than one set.[1]
Contact performed well; posture remains unvalidated
A second module tested 264 images after reserving four of 268 as examples. Qwen2.5-VL-7B distinguished physical contact from no contact with 88.6% accuracy and 88.5% macro-F1; two expert raters had first established the labels. The posture module used 193 images and repeated each classification five times. Its 0.916 average consistency and 0.705 full-agreement rate show that outputs were fairly stable, but there were no expert posture labels. The paper explicitly says those figures do not establish factual correctness, confidence or culturally valid interpretation.[1]
The distinction that matters
What each module actually measured
The denominators and outcome types differ, so the headline percentages are not interchangeable.
Activity
374 test images, nine labels
90.6% accuracy
Contact
264 test images, two labels
88.6% accuracy
Posture
193 images, no ground truth
Stability only
Well-being
No clinical outcome measure
Not established
Cultural validation and clinical outcomes are the next test
This is a small, specialised image study, with repeated photographs from the same participants and static frames rather than video. It did not test practitioner workload, treatment decisions, clinical outcomes or children's well-being. General-purpose models may also import cultural assumptions into labels such as ‘confident’. The study had Aboriginal health ethics approval, consent and age-appropriate assent, and keeps the deidentified images behind Aboriginal-led governance. Before use in practice, the posture categories need community co-design and expert ground truth, and the whole system needs prospective comparison with practitioner notes and standardised well-being measures. Until then, it is evidence for image organisation—not automated interpretation of a child's inner state.[1]