AI-Assisted Binoculars Improve Learning in Novice Birders

Simple Summary

Identification tools assist citizen scientists and novel birders in a unique way. Here, we tested the first AI-assisted binoculars in the world. Participants identified birds (taxidermies, models) with AI-binoculars in half of the examples with the AI-assistance, and in the other half with the AI-function switched off. This study shows that using AI-assisted binoculars helps novice participant birders to increase their species knowledge. The increase in knowledge was larger in the experimental group (using the AI-assistance) than in the control group without AI-assistance. We suggest extending the study to further, less experimental situations.

Abstract

AI tools like Passive Acoustic Monitoring (PAM) and apps like iNaturalist and Merlin are increasingly used in bird monitoring and species identification. The purpose of this study was to assess whether AI-assisted binoculars improve bird species knowledge, particularly in novice birders, and to examine users’ motivation and experience. This study focuses on the learning impact of users, not data quality or accuracy of the device itself. Participants were recruited via social media, mostly novices (10 women, 9 men, 1 diverse). Four experimental groups (A–D, with N = 5 participants each) were designated. Participants used AI-supported binoculars to identify 10 bird species and the same binoculars with AI function switched off to identify another 10 bird species based on two sets of different species (counterbalanced to avoid order effects). This allowed a between-group as well as a within-subject comparison. We used a pre-test/post-test design for learning. Significant knowledge gains occurred only when using AI binoculars (Wilcoxon tests, p = 0.008). Pooled data across the intervention groups showed strong learning effects for AI-assisted users (Z = −3.736, p = 0.001). No significant learning occurred under control conditions. As a conclusion, AI-assisted binoculars significantly enhance bird identification learning in novices, but as a cautionary note, the study needs to be extended to live birds and in longitudinal settings.

1. Introduction

Bird species knowledge in the public is generally low (e.g., [1,2,3,4,5,6]). However, when facing the biodiversity crisis, knowledge about biodiversity becomes increasingly important to combat the crisis [7]. Therefore, tools and instruments to enhance biological knowledge and knowledge about bird species are crucial [7,8]. This is also important in the age of citizen science, because tools have become helpful to support the increasing community of citizen scientists [9]. Therefore, assisting and accompanying emerging citizen scientists has become a challenging task, because recruiting people into such activities has changed over the last decades. For example, initiation into birding as a serious leisure activity has changed significantly during the last 50 years [10]. In addition, training emerging citizen scientists can require personal assistance from trained staff or online companionship [11,12], which can be both time-consuming and expensive. Therefore, AI-assisted tools may be a helpful development for future citizen scientists. AI-assisted tools have multiple benefits. For example, they allow quick and instant feedback of identification results to the users, they can increase participation of people because of lowering barriers to entry, can encourage learning and building identification skills in citizen scientists, and can engage the public in promoting interest in nature [13,14,15,16].

Artificial Intelligence (AI) is already seen and understood as a useful tool for studying birds and monitoring the bird population. For example, the apps iNaturalist and Merlin are commonly used for the identification of organisms in the field [17,18,19]. Herodotou [18] found that young people engaged in iNaturalist benefited when blended learning was applied, i.e., a combination of field work and online assistance. Rögele et al. [12] showed that both personal instruction by an instructor in a seminar course and online assistance worked well in accompanying citizen scientists. Pankiv and Kloetzer [19] showed an improvement in bird identification skills when using either Merlin or eBird, with participants using eBird scoring higher [19].

As new tools are increasingly emerging, we studied the usefulness of AI-assisted binoculars. In February 2024, Swarovski^® optics released an AI-assisted binocular that is connected to the Merlin app via Bluetooth. The binoculars are used in the same manner as conventional binoculars but are connected via the Merlin app with online databases. When a bird is in focus, pressing a button on the binoculars triggers a connection via Merlin that displays the name of the bird species directly in the field of view of the binoculars (representing a real-time species identification). Birdwatchers can therefore keep the binoculars directly to their eyes and do not have to take them off to use a book or app, for example.

The current study aimed to test whether AI-assisted binoculars help novices in increasing their bird species knowledge. In this kind of approach, the focus is on the participants’ learning rather than on just testing the quality of the AI-assisted binoculars themselves.

2. Materials and Methods

2.1. Participants and Data Collection

Participants were recruited via social media and in the environment of the University of Tübingen. The study was approved by the local ethics committee, and the data collection was anonymous, unpaid, and voluntary. The study took place in the botanical garden of the university outside to simulate natural light conditions (compared to a lab study), but with stuffed taxidermies and models to keep the experimental conditions of the bird identification under control. All participants received a generic alphanumeric code to relate pre- and posttests with each other. In the pretest, some demographic data were collected, and participants also carried out a test on bird species knowledge containing all 20 bird species of the intervention. Afterward, a post-test was carried out using the same 20 bird species (see Figure 1).

2.2. Research Protocol and Experimental Design

The experimental bird species sets are depicted in

Table 1. Overview of the experimental bird species sets. M = model specimen, T = taxidermy specimen.

Bird Species Set 1 (1–10)	Bird Species Set 2 (11–20)
1. Common Cuckoo Cuculus canorus T	11. Black-headed Gull Chroicocephalus ridibundus T
2. Eurasian Magpie Pica pica T	12. Eurasian Jay Garrulus glandarius T
3. Eurasian Green Woodpecker Picus viridis T	13. Great Tit Parus major T
4. Eurasian Blackcap Sylvia atricapilla T	14. Goldcrest Regulus regulus T
5. Black Redstart Phoenicurus ochruros M	15. Great Spotted Woodpecker Dendrocopos major M
6. Red-backed Shrike Lanius collurio M	16. Eurasian Tree Sparrow Passer montanus M
7. Bluethroat Luscinia svecica M	17. White Wagtail Motacilla alba M
8. Hawfinch Coccothraustes coccothraustes M	18. Common Chaffinch Fringilla coelebs M
9. Eurasian Bullfinch Pyrrhula pyrrhula M	19. Yellowhammer Emberiza citrinella M
10. White-throated Dipper Cinclus cinclus M	20. European Greenfinch Chloris chloris M

. In total, 20 different bird species were tested in the pretest and posttest. The bird species were presented as stuffed taxidermies or metal models at a given distance of about 5 m away from the observer. Taxidermies were bought from Windaus-Labortechnik GmbH & Co. KG (Clausthal-Zellerfeld, Germany). The model specimens were provided by Artfauna (Kelkheim, Germany). Artfauna offers lifelike replicas of native birds and other animals in detailed designs. Both were presented at a distance of about 5 m (Appendix A). Every participant had to identify all 20 species, and, thus, served as a within-person control. Half of the participants used the AI binoculars on the first set of species (groups A, D; experimental group 1, species 1–10; see Figure 1) and half of them used the AI-assisted binoculars on the second set of species (groups B, C; experimental group 2, species 11–20; see Figure 1). The other ten bird species, on the other hand, were identified without the help of AI and served as a control group. Thus, group A and D identified the same 10 species (species set 1) with AI-assistance, while group B and C received the other 10 species (species set 2) with AI-assistance. The control part or section of every group did not receive any identification aids to serve as baseline control and to account for possible retest effects (that may occur when a pretest/posttest design is applied [20]). To avoid any order effects (due to habituation or due to fatigue), half of the participants started with the control conditions (group C, D), and half with AI-assisted binoculars (group A, B) for a balanced sample design (Figure 1). During the use of the AI-assisted binoculars, participants were asked to report the identification provided by the AI system on a working sheet and to try the identification three times per bird species. The researchers did not give any response to the participants if the identification of the AI-assistance was wrong, but noted it on the paper sheet. In the control conditions, the participants looked at the bird species with identical binoculars, but the AI function and connection to Merlin were switched off. No additional identification tools were provided.

2.3. Measurement Instruments

2.3.1. Knowledge Gain (Pretest and Posttest)

We measured bird species knowledge by using a pretest and a posttest. The participants had to identify the same 20 different bird species prior to and after the intervention (Table 1). Ten species were presented with, and the others without, AI-assistance (Figure 1, Table 1). We calculated the knowledge gain by subtracting pretest scores from posttest scores for the two species sets separately. The test was organized in a way that the participants had to write down the correct name of the species [1]. The responses were graded by partial credits, with 1 assigned to a correct identification and 0.5 to a partially correct identification. For example, if the Green Woodpecker was named this way, the participant received the score of 1, but if it was named just woodpecker, the partial credit was 0.5 because the identification was partially correct (more details of the scoring system in [1]).

2.3.2. Bird-Related Attitudes (Pretest)

We used some assessments to depict our study population before the field experiment took place. First, the perception of birds was asked for, which measures the role of birds in people’s lives and influences people’s nature experience positively [21]. Cronbach’s α of the present sample was 0.554. Birding specialization was used based on the subscale of skill/knowledge [22]. Cronbach’s α of the present sample was 0.718. Further, interest in birds was measured with a scale adapted from Hummel et al. [23]; Cronbach’s α of the present sample was 0.876.

2.3.3. Motivational Variables (Posttest)

We used a short scale of intrinsic motivation, which has been applied many times in educational contexts (KIM; German version [24]). This short scale is an adapted, time-efficient version of the ‘Intrinsic Motivation Inventory’ by Deci and Ryan [25]. Three items were used for three different scales, totaling nine items. Interest/enjoyment is related to the interest in the task and the enjoyment. Cronbach’s α was 0.941. Perceived competence asks how competent participants felt during the task (Cronbach’s α = 0.840), and finally, pressure/tension is related to the question of whether participants felt under pressure (α = 0.525). All items of the KIM are scaled on a Likert scale from 1 to 5, with 1 related to a low expression and 5 to a high expression.

2.3.4. Subjective Assessment of the AI-Binoculars (Posttest)

We asked four additional questions, scaled from 1 to 5 (fully disagree to fully agree), to receive a subjective assessment of the AI-assisted binoculars: Participants assessed (1) the helpfulness of the device for bird identification, (2) the ease of its use, (3) the comprehensiveness of the instructions by the field assistants, and finally, (4) we asked for a recommendation to a friend (see items in the Appendix A).

2.4. Statistical Analysis

We used a chi-square test to compare the distribution of genders across the experimental conditions. A Friedman test was used to compare the three trials per taxidermy or model. Mann–Whitney U Tests were applied when two groups were compared, and the Wilcoxon signed rank test was applied to compare related samples (in the within-subject comparison). Spearman’s rank correlation was used to test relationships between two variables.

3. Results

3.1. Participants’ Demographics

Participants did not differ between the two intervention groups in their age, birding specialization, interest in birds, and bird perception (

Table 2. Comparison of the two intervention groups.

Variable	Experimental Group 1		Experimental Group 2		Statistics
	Mean	SD	Mean	SD	U	Z	p-value
Age	29.80	9.94	28.70	4.86	47.5	−0.189	0.850
Specialization	0.87	0.76	1.10	0.65	32.5	−1.356	0.175
Interest	3.17	0.71	2.80	1.30	36	−1.065	0.287
Perception of birds	4.10	0.47	3.80	0.79	35	−1.155	0.248

). Genders were equally distributed across the groups (χ² = 1.511, df = 2, p = 0.470;

Table 3. Distribution of genders across the experimental intervention groups.

		Experimental Group 1	Experimental Group 2	Total
Gender	Male	5	4	9
	Female	4	6	10
	Diverse	1	0	1
Total		10	10	20

). Birding specialization was generally low, corroborating our recruitment procedure, which targeted novices (see Table 2). About 75% of the participants were able to identify up to 25 species by sight, 85% reported being able to identify 0 to 5 species by sound, and most participants identified themselves as novices (16 from 20).

3.2. Learning

The identifications provided by the AI-assistance during the experimental conditions were highly correct, with >90% of identifications being correct when using the AI-assisted binoculars. The identifications were stable over time (Figure 2), and there was no significant difference between the first, second, or third trial (Friedman test, χ² = 1.310, df = 2, p = 0.519). The mean values and SD were 0.91 ± 0.107 for the first test, 0.92 ± 0.089 for the second test, and 0.93 ± 0.080 for the third test.

There was a significant difference in knowledge gain in experimental group 1 (combined group A and D) concerning species set 1 compared to species set 2 (Wilcoxon test, Z = 2.673, p = 0.008, N = 10; Figure 3), with participants scoring higher in knowledge gain in the 10 species they explored with the AI-assisted binoculars (species set 1–10), while they received no knowledge gain in the species set (species 11–20) which they explored without the assistance of the AI binoculars. Similarly, a significant difference in knowledge gain was found in experimental group 2 (combined B and C) in relation to species set 2 compared to species set 1 (Wilcoxon test, Z = −2.673, p = 0.008, N = 10; Figure 3). These participants scored higher in knowledge gain in relation to species set 2 (species 11–20), which they identified with the AI-assisted binoculars. No knowledge gain was found in this group in relation to species set 1 (species 1–10) that were explored without the AI-function (Figure 3). Both interventions showed the same results, with 9 out of 10 respondents having higher scores when using the AI-assisted binoculars and one tie, leading to the same Z and p values. Comparing both groups with each other showed a significant difference in knowledge gain concerning bird species set 1 (Mann–Whitney U test, Z = −2.486, p = 0.013) and bird species set 2 (Mann–Whitney U test, Z = −3.576, p < 0.001). This corroborates that experimental group 1 received a higher knowledge gain concerning bird species set 1 compared to experimental group 2. In contrast, knowledge gain in relation to bird species set 2 showed the opposite pattern (Figure 3). This strongly suggests that both groups gained knowledge and showed a learning effect specifically during the experimental condition when they used the AI-assisted binoculars, but not when the identification tasks were unassisted.

Then, the knowledge gain was combined by subsuming the values of the identification gain (per participant) for each experimental group under (a) number of correct identifications of the AI-assisted species (combining species 1–10 from experimental group 1 with species 11–20 from experimental group 2) and under (b) the number of correct identifications under control conditions without any assistance (control condition: species 11–20 from experimental group 1 and species 1–10 from experimental group 2). The results were highly significant, showing a greater knowledge gain in the AI-assisted part of the experiment (Wilcoxon test, Z = −3.736, p < 0.001, N = 20, Figure 4). No knowledge gain was observed under control conditions.

3.3. Motivation

Participants reported a high interest in and enjoyment of the activity, felt competent in a medium range, and perceived only low pressure (Figure 5). There were no differences in motivation between the two experimental groups (Mann–Whitney U tests; Interest: Z = −0.326, p = 0.744; Competence: Z = −0.045, p = 0.964; Pressure: Z = −0.692, p = 0.489). This suggests that both experimental interventions could be considered similar concerning the motivational variables.

3.4. Subjective Assessment of the Device

Participants considered the binoculars helpful for the identification of bird species and found them easy to use (Figure 6). Similarly, the comprehensiveness of the introduction and explanation of its use by the researchers was considered very good. Recommendation to a friend was rated an average value.

3.5. Subjective Learning Outcomes

There was a significant correlation between the self-assessment based on the birding specialization scale and the objectively measured knowledge in the pretest (r_s = 0.734, p < 0.001). Interest in birds and prior knowledge were also significantly positively correlated with each other (r_s = 0.659, p = 0.002).

4. Discussion

The binoculars supported by artificial intelligence produced more than 90% correct identifications (Figure 3). This can be considered high, but on the other hand, we used stuffed taxidermies and model specimens that looked highly natural. As the birds did not move, participants had enough time to identify the specimens with the AI-assisted binoculars. In this lab-like setting, however, the binoculars proved their quality (see limitations below). The study needs to be extended to other controlled environments to further study the usefulness for identification. The control condition was needed to check whether an increase in knowledge in the participants (knowledge gain) can be tied to the use of the AI-assisted binoculars, because an increase in knowledge can appear without any intervention (the retest effect). The retest effect is well-known in learning, and participants sometimes show an increase (or even a decrease) in learning without receiving any information, just by repeating the same or a similar test twice [20].

The knowledge gain differed significantly between the experimental groups, but both groups scored significantly higher during the AI-assisted identification tasks and significantly lower during the control conditions. Taken together, the effect became even stronger when the data were pooled across the experimental settings. That is, the AI-assisted binoculars have been helpful in improving bird species learning in novice birders.

The intrinsic motivational variables of interest/enjoyment, perceived competence, and pressure/tension indicated that participants found AI-supported bird identification to be an interesting activity, with low perceived pressure. Participants felt less competent, but this likely stemmed from the task itself, as their basic bird knowledge was limited and they considered themselves beginners, novices, or casual birders. Overall, the results suggest that the learning environment was quite pleasant for the participants. However, when compared to another study on bird identification [26], the findings were similar. In that other study, students used either a bird identification book or an app to identify bird species at a suburban pond. The mean values for interest/enjoyment were 3.9, perceived competence 2.8, and pressure/tension 1.7 ([23]; compare with Figure 5 above). The recommendation to a friend was only average, which may be related to the price of the binoculars, which are quite expensive.

Such optical devices linked to AI-assistance may improve the data quality of novices when submitting to citizen science platforms, which could increase data quality [27,28]. However, this has yet to be tested. While AI-assisted binoculars can enhance accessibility and accuracy in bird identification, they may also raise concerns about de-skilling and over-reliance on technology. If users come to depend heavily on automated identification, opportunities to develop and refine personal fieldcraft, observational acuity, and taxonomic knowledge could diminish. Balancing the benefits of technological assistance with the cultivation of independent identification skills may therefore be an important consideration for both educators and practitioners in the field.

Another important topic is to study how AI-assisted binoculars might bridge the gap between controlled learning environments (as in the present study) and authentic field experiences. One question to be answered is whether and how AI-assistance can help in developing and improving field skills. This has not been answered yet, but was found to be the case when using other AI-assistance, e.g., in the case of Merlin and eBird [19]. Additionally, a training program of birdsong has similarly shown an increase in knowledge [29], but this study was also carried out under lab-like conditions and needs additional testing of field skills. To our knowledge, this is a largely unexplored field that deserves further attention. In particular, long-term studies that scrutinize the development of field skills and how this technology might be integrated into a progression toward independent field observation would be a highly recommended research topic.

The study used a complex methodological design allowing both a within-subject and a between-subject comparison, which is a clear advantage. Within-subject comparisons benefit from keeping all possible confounders constant, such as the person themselves, time of day, weather, or any other known or unknown variable. To improve this research design even further, we applied a control for order effects by asking half of the participants to start with the AI-assisted part of the identification task, and the other half to start with the control condition part. In addition, half of the participants started with bird species set 1 and the other half with bird species set 2, leading to a completely counterbalanced design. We suggest such designs for other studies to improve comparisons.

5. Limitations

The greatest limitation of this study is the use of static taxidermied birds and models as identification targets. Static specimens differ significantly from real-world birding scenarios because real birds are moving around (depending on species), but some bird taxa like herons are sometimes nearly static during foraging (or perched species of prey), while others, such as warblers are flitting around quickly. This should impact the AI system’s performance, and hence, participants’ learning. This clarifies the scope of the findings and sets directions for future research.

Further studies should include moving objects (real birds) of different levels of movement, which, however, makes it more difficult to control confounding effects. In another study, we should address the usage during a longer period of time of data collection in an ecological outdoor setting to gather some retention effects in addition to the short-term learning effects. It would be further interesting to see if and how field skills develop in birders using the AI-assisted binoculars.

The internal consistency of the perception of birds and the pressure/tension scale from the motivational constructs was <0.6 >0.5. These values are below the level of acceptability. However, in other studies, the same scales received higher values. This could be due to the effect of the sample size, as lower sample sizes may result in lower alpha levels. The sample size of our study can be considered low, but it is sufficient to detect medium effects when considering the control conditions (baseline control). The sample size should be higher when two different media or tools are compared.

6. Conclusions

Although the results are convincing, it must be stated that the study was based on an experimental laboratory situation and that it should be extended to real-life situations and to studies incorporating further variables or longitudinal designs. Nevertheless, the data show that the use of AI-assisted binoculars leads to a short-term learning effect.