{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"factories-fields-and-clinics","title":"The uses that are not on a screen","url":"https://addaly.com/learn/ai-actually-explained/factories-fields-and-clinics","position":18,"of":92,"minutes":9,"answer":"Field deployments fail on conditions the test set never contained (light, dust, hurry and bandwidth) which is why the successful clinical systems triage for a human rather than decide.","keyIdea":"Field deployments fail on conditions the test set never contained (light, dust, hurry and bandwidth) which is why the successful clinical systems triage for a human rather than decide."},"body":{"format":"markdown","text":"## Away from the office\n\nAlmost all writing about AI assumes a desk. Most of the world does not work at one, and some of the most established machine learning is running in places with no keyboard in sight.\n\n**Visual inspection on production lines.** A camera above a conveyor, a model trained on thousands of good and defective parts, and a rejection arm. This has been in service since well before the current wave and is one of the most reliably profitable uses there is, because the task is narrow, the lighting is controlled, and the right answer is unambiguous.\n\n**Predictive maintenance.** Vibration, temperature and current sensors on a motor, and a model that learns what the healthy signature sounds like and flags departures from it. It replaces both fixed-schedule servicing and waiting for failure.\n\n**Agriculture.** Phone apps that identify crop disease from a photograph of a leaf (Plantix is used widely across India) plus satellite-based yield estimation, irrigation scheduling from soil and weather data, and camera-guided sprayers that target individual weeds and cut herbicide use substantially.\n\n**Livestock and fisheries.** Cameras and collars that detect lameness, oestrus or reduced feeding days before a stockman would.\n\n**Clinics.** Two deployments matter here because they are endorsed rather than merely trialled. Chest X-ray software for tuberculosis triage was recommended by the WHO in 2021 for adults in certain settings, which matters enormously where there are more X-rays than radiologists. And screening for diabetic retinopathy (a photograph of the retina, checked by a model) has regulatory approval in several countries and is deployed in India and Thailand.\n\n## The lesson these deployments taught\n\nThe most useful thing to know about clinical AI comes from a field study Google published in 2020 on its diabetic retinopathy system in Thailand.\n\nIn the lab, the model matched specialists. In eleven real clinics it hit problems no accuracy figure had predicted. More than a fifth of images were rejected by the system for insufficient quality, because clinic lighting was poor and nurses were photographing quickly in a queue. Rejected images meant a patient had to return another day, or travel to a hospital. Internet connections were slow, so uploads stalled. Nurses developed workarounds. In some sites the system slowed the clinic down while being, technically, extremely accurate.\n\nThat study is quoted often, and it should be, because it names the single most common reason AI projects fail in the field: **the conditions of deployment are not the conditions of the test set.** Dust, glare, a cracked lens, a hurried operator, a two-megabit connection and a power cut are not in the validation data.\n\n## Why these uses work when others do not\n\nThe successful cases share a shape:\n\n- The task is **narrow and repeated**. One kind of defect, one crop disease, one screening question.\n- The right answer is **unambiguous and cheap to obtain** for training. A part either failed or it did not.\n- The environment is **controllable**, or the system is designed for the environment it will actually meet.\n- The model **triages rather than decides**. TB software marks films for a clinician to read; it does not diagnose. This one design choice is why it was endorsed.\n\nWhen a proposed use fails several of these, be sceptical regardless of the demonstration.\n\n## The free path here too\n\nSmall industrial and agricultural uses do not require a vendor platform. A vision model for defect detection can be trained on a few hundred photographs using free tools: **Python** with **PyTorch** or **scikit-learn**, **OpenCV** for the camera handling, **Label Studio** or **CVAT** for annotating your examples, and a second-hand machine or a free Google Colab session for the training. Pre-trained models from **Hugging Face** mean you are adjusting an existing model rather than starting from nothing, which is what makes a few hundred photographs enough.\n\nThe expensive part of such a project is almost never the model. It is collecting and labelling the examples, and mounting the camera so the lighting does not change. Anyone who tells you otherwise has not deployed one."},"check":{"question":"A retinal screening model matching specialist accuracy in the lab slowed clinics down when deployed. What was the main mechanism?","options":[{"text":"The model's accuracy dropped in the field, so more patients had to be referred to specialists.","correct":false,"tempting":"Assumes accuracy degraded; on the images it did accept, the model performed as expected."},{"text":"Nurses distrusted the system and re-checked every result manually, doubling the work.","correct":false,"tempting":"Blames adoption behaviour, which does occur, but was not the bottleneck the study identified."},{"text":"It rejected a large share of images as too poor in quality, because real clinic lighting and hurried photography differed from the conditions the test images came from.","correct":true},{"text":"The system required a specialist to confirm every positive finding before the patient could be told.","correct":false,"tempting":"Describes a sensible triage design rather than the deployment failure that caused the delays."}]},"prev":{"slug":"decisions-made-about-you","title":"When a model decides something that matters","url":"https://addaly.com/learn/ai-actually-explained/decisions-made-about-you"},"next":{"slug":"spotting-ai-in-a-product","title":"How to tell whether a model is involved","url":"https://addaly.com/learn/ai-actually-explained/spotting-ai-in-a-product"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The uses that are not on a screen', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/factories-fields-and-clinics (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The uses that are not on a screen","courseTitle":"AI, Actually Explained","module":{"index":2,"slug":"where-it-already-is","title":"Where it already is, without a label"},"answer":null,"keyIdea":"Field deployments fail on conditions the test set never contained (light, dust, hurry and bandwidth) which is why the successful clinical systems triage for a human rather than decide.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"away-from-the-office","runs":[{"text":"Away from the office"}]},{"t":"p","runs":[{"text":"Almost all writing about AI assumes a desk. Most of the world does not work at one, and some of the most established machine learning is running in places with no keyboard in sight."}]},{"t":"p","runs":[{"text":"Visual inspection on production lines.","b":true},{"text":" A camera above a conveyor, a model trained on thousands of good and defective parts, and a rejection arm. This has been in service since well before the current wave and is one of the most reliably profitable uses there is, because the task is narrow, the lighting is controlled, and the right answer is unambiguous."}]},{"t":"p","runs":[{"text":"Predictive maintenance.","b":true},{"text":" Vibration, temperature and current sensors on a motor, and a model that learns what the healthy signature sounds like and flags departures from it. It replaces both fixed-schedule servicing and waiting for failure."}]},{"t":"p","runs":[{"text":"Agriculture.","b":true},{"text":" Phone apps that identify crop disease from a photograph of a leaf (Plantix is used widely across India) plus satellite-based yield estimation, irrigation scheduling from soil and weather data, and camera-guided sprayers that target individual weeds and cut herbicide use substantially."}]},{"t":"p","runs":[{"text":"Livestock and fisheries.","b":true},{"text":" Cameras and collars that detect lameness, oestrus or reduced feeding days before a stockman would."}]},{"t":"p","runs":[{"text":"Clinics.","b":true},{"text":" Two deployments matter here because they are endorsed rather than merely trialled. Chest X-ray software for tuberculosis triage was recommended by the WHO in 2021 for adults in certain settings, which matters enormously where there are more X-rays than radiologists. And screening for diabetic retinopathy (a photograph of the retina, checked by a model) has regulatory approval in several countries and is deployed in India and Thailand."}]},{"t":"h2","id":"the-lesson-these-deployments-taught","runs":[{"text":"The lesson these deployments taught"}]},{"t":"p","runs":[{"text":"The most useful thing to know about clinical AI comes from a field study Google published in 2020 on its diabetic retinopathy system in Thailand."}]},{"t":"p","runs":[{"text":"In the lab, the model matched specialists. In eleven real clinics it hit problems no accuracy figure had predicted. More than a fifth of images were rejected by the system for insufficient quality, because clinic lighting was poor and nurses were photographing quickly in a queue. Rejected images meant a patient had to return another day, or travel to a hospital. Internet connections were slow, so uploads stalled. Nurses developed workarounds. In some sites the system slowed the clinic down while being, technically, extremely accurate."}]},{"t":"p","runs":[{"text":"That study is quoted often, and it should be, because it names the single most common reason AI projects fail in the field: "},{"text":"the conditions of deployment are not the conditions of the test set.","b":true},{"text":" Dust, glare, a cracked lens, a hurried operator, a two-megabit connection and a power cut are not in the validation data."}]},{"t":"h2","id":"why-these-uses-work-when-others-do-not","runs":[{"text":"Why these uses work when others do not"}]},{"t":"p","runs":[{"text":"The successful cases share a shape:"}]},{"t":"ul","items":[[{"text":"The task is "},{"text":"narrow and repeated","b":true},{"text":". One kind of defect, one crop disease, one screening question."}],[{"text":"The right answer is "},{"text":"unambiguous and cheap to obtain","b":true},{"text":" for training. A part either failed or it did not."}],[{"text":"The environment is "},{"text":"controllable","b":true},{"text":", or the system is designed for the environment it will actually meet."}],[{"text":"The model "},{"text":"triages rather than decides","b":true},{"text":". TB software marks films for a clinician to read; it does not diagnose. This one design choice is why it was endorsed."}]]},{"t":"p","runs":[{"text":"When a proposed use fails several of these, be sceptical regardless of the demonstration."}]},{"t":"h2","id":"the-free-path-here-too","runs":[{"text":"The free path here too"}],"role":"free"},{"t":"p","runs":[{"text":"Small industrial and agricultural uses do not require a vendor platform. A vision model for defect detection can be trained on a few hundred photographs using free tools: "},{"text":"Python","b":true},{"text":" with "},{"text":"PyTorch","b":true},{"text":" or "},{"text":"scikit-learn","b":true},{"text":", "},{"text":"OpenCV","b":true},{"text":" for the camera handling, "},{"text":"Label Studio","b":true},{"text":" or "},{"text":"CVAT","b":true},{"text":" for annotating your examples, and a second-hand machine or a free Google Colab session for the training. Pre-trained models from "},{"text":"Hugging Face","b":true},{"text":" mean you are adjusting an existing model rather than starting from nothing, which is what makes a few hundred photographs enough."}]},{"t":"p","runs":[{"text":"The expensive part of such a project is almost never the model. It is collecting and labelling the examples, and mounting the camera so the lighting does not change. Anyone who tells you otherwise has not deployed one."}]}],"check":{"question":[{"text":"A retinal screening model matching specialist accuracy in the lab slowed clinics down when deployed. What was the main mechanism?"}],"options":[{"text":[{"text":"The model’s accuracy dropped in the field, so more patients had to be referred to specialists."}],"correct":false,"tempting":[{"text":"Assumes accuracy degraded; on the images it did accept, the model performed as expected."}]},{"text":[{"text":"Nurses distrusted the system and re-checked every result manually, doubling the work."}],"correct":false,"tempting":[{"text":"Blames adoption behaviour, which does occur, but was not the bottleneck the study identified."}]},{"text":[{"text":"It rejected a large share of images as too poor in quality, because real clinic lighting and hurried photography differed from the conditions the test images came from."}],"correct":true,"tempting":null},{"text":[{"text":"The system required a specialist to confirm every positive finding before the patient could be told."}],"correct":false,"tempting":[{"text":"Describes a sensible triage design rather than the deployment failure that caused the delays."}]}],"order":[1,3,2,0],"why":null},"prev":{"slug":"decisions-made-about-you","title":"When a model decides something that matters","minutes":10},"next":{"slug":"spotting-ai-in-a-product","title":"How to tell whether a model is involved","minutes":8}}}