There is a window of about two hours, late in the afternoon, when one can go up the hill to check a camera trap.Not earlier, because that is when cattle and goats are on the slope, and the granite boulders – the flattest, most walkable surface for miles – are exactly where they climb. Not later, because by then the villagers are bringing their livestock home and the trail is full. The same is true at first light. So the work happens in the gaps: early morning, or the hour before the herds come down.If one misses those windows, the card fills. Not with wildlife, with cattle, with goats, with people walking to collect firewood, with whatever crosses the sensor a hundred times in an afternoon. A memory card that should hold three weeks of detections is finished in days.This is fieldwork outside a protected area in the Santhal Pargana division of Jharkhand, eastern India, carried out by one of the authors of this piece, Kulesh Bhandari. What follows is his account.I have spent close to 240 field days across 35 granite outcrops here, and put out something in the order of 80 camera-trap deployments, most recently eight units for 23 days. None of it inside a national park, a sanctuary, or a reserve. All of it in a landscape that is farmed, grazed, walked through and lived in, every day, by the people whose ground it is.That is not an unusual place for Indian wildlife to be. India’s protected area network – national parks, wildlife sanctuaries, conservation and community reserves, marine protected areas – covers about 5% of the country’s geographical area. Everything else lives somewhere in the other ninety-five. Roughly 80% of the Asian elephant’s range in India falls outside the protected area network, and elephants are the species we track most closely. For a jackal or a striped hyena (Hyaena hyaena), nobody has run the equivalent number.It is important to understand what that does to the standard methods, because the answer is not that they get harder. It is that they stop returning usable information, and the landscape then reads as empty in the national record.What it takes to service a cameraReaching a survey landscape here can mean 120 to 150 kilometres of travel. The motorcycle has to be left at the nearest village – leaving it at the roadside near forest invites theft – and from there it is five to eight kilometres on foot before the survey area even begins. A day of walking a set of units runs to twelve or eighteen kilometres, much of it on steep granite.Placement is not free either. The positions that would give the best detections – high on the outcrop, on the open rock, along the obvious movement lines – are the positions most likely to be found. Cattle climb the boulders. Firewood collection runs through the same trails an animal uses. When mushrooms are fruiting, or when people are out collecting wild onion, the traffic through the forest multiplies. So the units go where they are less likely to be disturbed, which is not where they are most likely to detect.Everything about the design bends around that constraint: how long a unit can stay out, how far from a village it can go, how much of the landscape can be sampled at once. A protocol written for a protected area assumes none of this. It assumes you place the camera where the animal is.Tracks that cannot be readThe alternative is sign. In principle this landscape should be good for it: nullahs, river corridors, the sandy margins of fields, the crossings between one patch of forest and the next. In practice it is where every dog, goat and sheep in the area also walks.Recording a canid track in a nullah bed outside any protected area. Photo: Kulesh Bhandari.I have checked more than fifty tracks. Something in the order of 12% were usable. The rest were degraded past identification: the outline collapsed, the hind print missing, the margin trodden over. The soil compounds it – much of it is sandy, so a print does not hold a clean impression even when it is fresh. And people are moving on these hills from four in the morning, before anyone can get to a track first.I photographed a canid print in a nullah bed in August. It is probably a jackal. I cannot say so with confidence, and that is the point: inside a protected area the same print, on undisturbed substrate, at eight in the morning, would simply be a record.Every other method, tooNone of this is only about camera traps. The same pressure runs through every method available here.Before I put a camera trap anywhere, I usually begin with a conversation. Whenever I visit a village, I carry a colour printout of the animal I am looking for. I show it to people and ask whether they have seen it, when they last saw it, at what time of day, and which routes the animal uses. Before deploying camera traps, I try to cross-check whether the animal is regularly seen in that area and where it might be moving.But these accounts do not always lead to a correct identification. Recently, villagers reported seeing what they described as a lion or tiger. The animal was actually a large, well-built hyena. There is no tiger record from the landscape I am working in. The difficulty is not that people are necessarily lying. They saw an animal, but attached the wrong name to it. For a researcher, the challenge is to separate what someone genuinely observed from the uncertainty of identifying a species in the field. Local knowledge remains important, but it cannot always be treated as a confirmed record without verification.A canid track in a maize field, eastern India. Identifying it as a golden jackal rather than a domestic dog required checking when dogs used the surrounding trails and how far the nearest settlement was — work that rarely appears in a final species list. Photo: Kulesh Bhandari.A few days ago, I found what appeared to be a canid track in a maize field. It could have belonged to a golden jackal (Canis aureus), but domestic dogs are also active at night. I checked the surrounding forest trails and agricultural fields to understand when dogs were present, whether they used the trail, and how far the location was from human settlements. The dogs did not appear to use that particular trail, and the settlement was considerably distant. This made a golden jackal identification more plausible, although the conclusion came from the surrounding evidence, not the footprint alone.This is the work that rarely appears in a final species list. A track is photographed, measured and recorded, but its identification may depend on checking domestic animal movement and understanding how the landscape is used.I used to think that finding wildlife meant going deeper into the forest. Fieldwork has complicated that assumption. During transect walks and night spotlighting, sometimes continuing until eleven or twelve at night, I have detected golden jackals and striped hyenas around nullahs, beneath palm trees and along forest-edge areas. Wildlife also moves through roads, agricultural fields and village corridors. I have recorded jackal calls at night, including instances where nearby domestic dogs begin vocalising in response.These encounters show why conservation research outside protected areas cannot be built around the assumption that wildlife exists only in deep forest. Animals move through a landscape shared with people, livestock and agriculture. The challenge is to study that movement without overlooking the places where it actually occurs.The limitations are also practical. Researchers working in these landscapes may not have access to laboratories, specialist equipment or sufficient funding for comprehensive analysis. There is rarely a ready pool of trained research assistants who can be hired for repeated fieldwork and systematic monitoring. As a result, some observations remain unresolved. A scientific account should not become more certain simply because a researcher needs to finish a paper. Recording what cannot yet be confirmed is part of doing the work honestly.This is not a complaint about peopleI want to be careful here, because there is an easy and wrong version of this argument.Take the ridge I work on most. The village below it has around forty households. Their cattle go up in the dry season and less so in the monsoon. A dirt road runs through the corridor between the hills, and several hundred people use it in a day, with motorcycles and horns.None of that is a disturbance to be controlled for. It is the landscape functioning as it has always functioned. These are working hills, and the people on them are not intruding on a study area; the study area is on their ground.The failure is not theirs. It is in a research framework built for places where people are absent by design, applied to places where they never were. And the consequence is not academic. When a landscape cannot be surveyed, it produces no data. When it produces no data, it appears empty. And what appears empty does not attract protection, funding, or attention.Ravindra Tripathi is a postdoctoral scholar at Pennsylvania State University working on wildlife informatics. His work uses AI and spatiotemporal data science to study wildlife and human movement and habitat dynamics, and he builds the ML models and trajectory forecasting. We put to him the obvious objection to everything above: that technology already solves this.There is an obvious answer to everything described above, and I have heard it in every room where this problem arises. If the landscape is too disturbed to survey by hand, automate it. Put out more cameras and let a neural network read the cards. Fly a drone with multiple sensors and parameters. Use the tools.I build those tools. I have worked on training deep networks to detect and identify animals in camera-trap and aerial imagery, flown platforms carrying optical, thermal, multispectral and LiDAR sensors, and fitted species distribution models across continental extents. I have written pipelines that turn a hundred thousand frames into a table of detections overnight. So I want to say this from inside the work, rather than from the sidelines: the technology is real, it is improving quickly, and it does not solve the problem it is being asked to solve here.It changes which designs are possible. It does not repair a design whose assumptions were already broken. And while appearing to remove one layer of observation error, it quietly adds a second.I have flown UAVs over swamp deer, and the trouble starts before any model runs. Tall grass hides half a body, animals standing close together can be read as one, and the clean thermal signal at first light is gone within the hour. Fly low and you photograph deer that are already running; fly high and they become a few pixels that could just as easily be cattle. The model will give you a number either way. But the number does not tell you whether the deer was hidden in grass, whether two animals became one detection, or whether the aircraft itself changed what the animals were doing. By the time that number reaches the dataset, much of the context that produced it has been lost.The machine is also an observerOccupancy modelling is one of ecology’s best ideas. It separates whether an animal was there from whether we managed to see it, and estimates the second so the first can be trusted. Much of the problem described above is about that second term: the camera placed where it would not be stolen, the track that could not be read, the narrow window in the afternoon when the hill was quiet enough to work.But the framework assumes that once a detection exists, it is a detection. The species on the record is assumed to be the species in the frame. That was a reasonable assumption when someone sat and looked at every image. It is less safe now.A vision model is making the same judgement a field biologist makes, only much faster and with no sense of where it might be wrong. It learns the background as well as the animal. A model that works beautifully where it was trained can fail somewhere it has never seen. And the animals we most need to find are often the ones it has seen least.The most unsettling part is that it does not know when it is guessing. Give it something outside its experience, and it can still return 0.94, with exactly the same confidence it shows when it is right.Why this is not simply noiseIf these mistakes were random, they would be an annoyance and no more. They are not. They cluster in cluttered frames, poor light, odd angles, images crowded with livestock and people.The field constraint and the model error are not two problems to be solved separately. They are correlated, and along the same axes. The camera was pushed away from human traffic, and the images that human traffic did produce are the ones the classifier handles worst. The second layer of error lines up with the first and deepens it.Aerial survey has its own version: thermal fails under canopy, LiDAR is blind to behaviour, flight height trades detection against disturbance. Stacked together, these biases do not cancel. They compound, and the compounding error is invisible at the final table.A frame rejected by the on-device model is not a frame anyone can return to. The false negative is gone before a human sees it. What reaches the desk is not the record of what crossed the sensor, but what one classifier decided to keep. And it arrives looking like raw data.That is the problem. Automation can make wildlife surveys faster and more consistent, while also hiding the uncertainty it introduces.None of this means abandoning the tools. Test the classifier on your own landscape before trusting its confidence. Keep a human-checked subset and report false positives alongside detections. Record where cameras could not be placed, not just where they were. Keep rejected frames when storage allows.The goal is not to choose between fieldwork and technology. It is to know what each one misses and make sure those misses do not quietly become zeros. Because behind every number is an animal that was there, whether or not our instruments managed to see it.Kulesh Bhandari is an independent biodiversity researcher and wildlife conservationist based in Jharkhand. He works on golden jackals and striped hyenas in the granite inselberg landscapes of the state, outside its protected areas. Ravindra Tripathi is a postdoctoral scholar at Pennsylvania State University working on wildlife informatics.