{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-camera-in-your-pocket","title":"Your photographs are partly predicted","url":"https://addaly.com/learn/ai-actually-explained/the-camera-in-your-pocket","position":13,"of":92,"minutes":9,"answer":"A phone photograph is a merged, model-enhanced prediction of a good picture rather than a straight recording, so shoot RAW when the image has to serve as evidence.","keyIdea":"A phone photograph is a merged, model-enhanced prediction of a good picture rather than a straight recording, so shoot RAW when the image has to serve as evidence."},"body":{"format":"markdown","text":"## The photograph you did not quite take\n\nPoint a modern phone at a dim room and you get a bright, sharp, low-noise image that the sensor physically could not have captured. This is not a trick of the lens. Between the sensor and the picture sits a stack of trained models, and the file you keep is as much a prediction as a recording.\n\nThe pipeline, roughly, in the moment you press the button:\n\n- The phone has already been capturing frames continuously. It grabs a **burst**: Google's HDR+ typically merges around ten to fifteen; night modes take several seconds' worth.\n- It **aligns** those frames, compensating for your hand shake and for things moving in the scene.\n- It **merges** them, which cancels the random noise, because noise differs frame to frame while the real scene does not.\n- Learned models then handle **demosaicing**, tone mapping, white balance, sharpening and often face and sky treatment.\n- Portrait mode adds a **segmentation** model that separates subject from background and blurs the background artificially.\n- Zoom beyond the optical limit invokes **super-resolution**, which produces detail that was never resolved by the lens.\n\nEach step is a model trained on paired examples: this sensor input, that desirable output.\n\n## Where it stops being a recording\n\nThe interesting question is where enhancement becomes invention, and there is a concrete case that made this unavoidable.\n\nIn 2023 a user demonstrated that a popular phone photographing a deliberately blurred image of the moon produced a crisp, detailed lunar surface: detail that could not have come from the blurred source. The manufacturer explained it as a scene-recognition and detail-enhancement feature. Whatever you call it, the craters in that photograph were supplied by the system, not by the light.\n\nThat is the extreme, and most processing is far more modest. But the principle it exposes is general: a model trained to produce plausible photographs will fill in plausible detail, and plausible detail is not evidence.\n\n## Why two phones disagree about the same scene\n\nStand two people side by side photographing the same street and the pictures will differ in a consistent, brand-specific way: one warmer, one contrastier, one with visibly brighter shadows and smoother skin.\n\nThat is not sensor variation. Each manufacturer trains its pipeline towards a preferred look, decided by its own imaging team, and that look is baked into the models. It is closer to a house style (the way two newspapers crop and grade the same press photograph differently) than to a measurement.\n\nSo \"which phone has the best camera\" is partly a question about hardware and largely a question about whose taste you share. It also means the same scene will look different when you upgrade, which is why old and new photographs in one album often refuse to sit together.\n\n## Why this matters beyond photography\n\nTwo consequences you can act on.\n\n**For documentation.** If you are photographing a damaged part, a rash, a document, a meter reading or an accident, heavy processing is working against you. Shoot in good light, get close, and where possible use the RAW or Pro mode, which bypasses much of the pipeline and gives you the sensor's own data. The image will look worse and mean more.\n\n**For trust.** A phone photograph has not been a straightforward record for about eight years now, and this predates any of the generative image tools. Face smoothing that cannot be fully switched off, skies replaced wholesale, and scene-specific enhancements are all in ordinary consumer phones. When somebody says a picture \"looks unnatural\", the honest response is that all of them are processed, and the line is a matter of degree.\n\n## Doing it yourself, free\n\nThe paid standard for photo work is Adobe Lightroom and Photoshop, and job adverts name them. The free tools that do the same jobs are capable:\n\n- **darktable** and **RawTherapee** develop RAW files with full manual control, which is exactly the control the phone takes from you.\n- **GIMP** covers retouching and compositing.\n- **Krita** is for painting and illustration.\n\nAll four run on modest hardware, including older laptops, and none of them require a subscription. Learning to develop a RAW yourself is also the fastest way to understand what your phone was doing on your behalf, because you have to make each of those decisions by hand.\n\n## The habit\n\nWhen an image matters (as proof, as a record, as something you will rely on later) ask what the camera added. Then decide whether you want the pretty version or the true one. Most of the time the pretty one is fine. The point is that it is now a choice, and most people do not know they are making it."},"check":{"question":"Someone photographs a cracked machine part in a dim workshop to send to an insurer, and the phone's night mode produces a clean, bright image. What is the specific risk?","options":[{"text":"The file will be too large to email once night mode has merged fifteen frames into it.","correct":false,"tempting":"Worries about file size, which merging does not increase, and misses what merging does to the content."},{"text":"Night mode works only on scenes it recognises, so an unfamiliar machine part will come out unprocessed and dark.","correct":false,"tempting":"Assumes the pipeline needs to recognise the subject; the burst-merge stage runs regardless of what is in the frame."},{"text":"Denoising, sharpening and super-resolution can smooth or reconstruct fine detail, so the crack may be softened or altered by the model rather than recorded.","correct":true},{"text":"The image will carry metadata showing it was AI-processed, which insurers routinely reject.","correct":false,"tempting":"Invents a rejection policy; the real problem is in the pixels, not in a metadata flag."}]},"prev":{"slug":"spam-fraud-and-the-oldest-ai","title":"The models that have been working since 2002","url":"https://addaly.com/learn/ai-actually-explained/spam-fraud-and-the-oldest-ai"},"next":{"slug":"autocomplete-maps-and-the-keyboard","title":"Typing, routing, and systems that change what they measure","url":"https://addaly.com/learn/ai-actually-explained/autocomplete-maps-and-the-keyboard"},"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) 'Your photographs are partly predicted', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-camera-in-your-pocket (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":"Your photographs are partly predicted","courseTitle":"AI, Actually Explained","module":{"index":2,"slug":"where-it-already-is","title":"Where it already is, without a label"},"answer":null,"keyIdea":"A phone photograph is a merged, model-enhanced prediction of a good picture rather than a straight recording, so shoot RAW when the image has to serve as evidence.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-photograph-you-did-not-quite-take","runs":[{"text":"The photograph you did not quite take"}]},{"t":"p","runs":[{"text":"Point a modern phone at a dim room and you get a bright, sharp, low-noise image that the sensor physically could not have captured. This is not a trick of the lens. Between the sensor and the picture sits a stack of trained models, and the file you keep is as much a prediction as a recording."}]},{"t":"p","runs":[{"text":"The pipeline, roughly, in the moment you press the button:"}]},{"t":"ul","items":[[{"text":"The phone has already been capturing frames continuously. It grabs a "},{"text":"burst","b":true},{"text":": Google’s HDR+ typically merges around ten to fifteen; night modes take several seconds’ worth."}],[{"text":"It "},{"text":"aligns","b":true},{"text":" those frames, compensating for your hand shake and for things moving in the scene."}],[{"text":"It "},{"text":"merges","b":true},{"text":" them, which cancels the random noise, because noise differs frame to frame while the real scene does not."}],[{"text":"Learned models then handle "},{"text":"demosaicing","b":true},{"text":", tone mapping, white balance, sharpening and often face and sky treatment."}],[{"text":"Portrait mode adds a "},{"text":"segmentation","b":true},{"text":" model that separates subject from background and blurs the background artificially."}],[{"text":"Zoom beyond the optical limit invokes "},{"text":"super-resolution","b":true},{"text":", which produces detail that was never resolved by the lens."}]]},{"t":"p","runs":[{"text":"Each step is a model trained on paired examples: this sensor input, that desirable output."}]},{"t":"h2","id":"where-it-stops-being-a-recording","runs":[{"text":"Where it stops being a recording"}]},{"t":"p","runs":[{"text":"The interesting question is where enhancement becomes invention, and there is a concrete case that made this unavoidable."}]},{"t":"p","runs":[{"text":"In 2023 a user demonstrated that a popular phone photographing a deliberately blurred image of the moon produced a crisp, detailed lunar surface: detail that could not have come from the blurred source. The manufacturer explained it as a scene-recognition and detail-enhancement feature. Whatever you call it, the craters in that photograph were supplied by the system, not by the light."}]},{"t":"p","runs":[{"text":"That is the extreme, and most processing is far more modest. But the principle it exposes is general: a model trained to produce plausible photographs will fill in plausible detail, and plausible detail is not evidence."}]},{"t":"h2","id":"why-two-phones-disagree-about-the-same-scene","runs":[{"text":"Why two phones disagree about the same scene"}]},{"t":"p","runs":[{"text":"Stand two people side by side photographing the same street and the pictures will differ in a consistent, brand-specific way: one warmer, one contrastier, one with visibly brighter shadows and smoother skin."}]},{"t":"p","runs":[{"text":"That is not sensor variation. Each manufacturer trains its pipeline towards a preferred look, decided by its own imaging team, and that look is baked into the models. It is closer to a house style (the way two newspapers crop and grade the same press photograph differently) than to a measurement."}]},{"t":"p","runs":[{"text":"So “which phone has the best camera” is partly a question about hardware and largely a question about whose taste you share. It also means the same scene will look different when you upgrade, which is why old and new photographs in one album often refuse to sit together."}]},{"t":"h2","id":"why-this-matters-beyond-photography","runs":[{"text":"Why this matters beyond photography"}]},{"t":"p","runs":[{"text":"Two consequences you can act on."}]},{"t":"p","runs":[{"text":"For documentation.","b":true},{"text":" If you are photographing a damaged part, a rash, a document, a meter reading or an accident, heavy processing is working against you. Shoot in good light, get close, and where possible use the RAW or Pro mode, which bypasses much of the pipeline and gives you the sensor’s own data. The image will look worse and mean more."}]},{"t":"p","runs":[{"text":"For trust.","b":true},{"text":" A phone photograph has not been a straightforward record for about eight years now, and this predates any of the generative image tools. Face smoothing that cannot be fully switched off, skies replaced wholesale, and scene-specific enhancements are all in ordinary consumer phones. When somebody says a picture “looks unnatural”, the honest response is that all of them are processed, and the line is a matter of degree."}]},{"t":"h2","id":"doing-it-yourself-free","runs":[{"text":"Doing it yourself, free"}]},{"t":"p","runs":[{"text":"The paid standard for photo work is Adobe Lightroom and Photoshop, and job adverts name them. The free tools that do the same jobs are capable:"}]},{"t":"ul","items":[[{"text":"darktable","b":true},{"text":" and "},{"text":"RawTherapee","b":true},{"text":" develop RAW files with full manual control, which is exactly the control the phone takes from you."}],[{"text":"GIMP","b":true},{"text":" covers retouching and compositing."}],[{"text":"Krita","b":true},{"text":" is for painting and illustration."}]]},{"t":"p","runs":[{"text":"All four run on modest hardware, including older laptops, and none of them require a subscription. Learning to develop a RAW yourself is also the fastest way to understand what your phone was doing on your behalf, because you have to make each of those decisions by hand."}]},{"t":"h2","id":"the-habit","runs":[{"text":"The habit"}]},{"t":"p","runs":[{"text":"When an image matters (as proof, as a record, as something you will rely on later) ask what the camera added. Then decide whether you want the pretty version or the true one. Most of the time the pretty one is fine. The point is that it is now a choice, and most people do not know they are making it."}]}],"check":{"question":[{"text":"Someone photographs a cracked machine part in a dim workshop to send to an insurer, and the phone’s night mode produces a clean, bright image. What is the specific risk?"}],"options":[{"text":[{"text":"The file will be too large to email once night mode has merged fifteen frames into it."}],"correct":false,"tempting":[{"text":"Worries about file size, which merging does not increase, and misses what merging does to the content."}]},{"text":[{"text":"Night mode works only on scenes it recognises, so an unfamiliar machine part will come out unprocessed and dark."}],"correct":false,"tempting":[{"text":"Assumes the pipeline needs to recognise the subject; the burst-merge stage runs regardless of what is in the frame."}]},{"text":[{"text":"Denoising, sharpening and super-resolution can smooth or reconstruct fine detail, so the crack may be softened or altered by the model rather than recorded."}],"correct":true,"tempting":null},{"text":[{"text":"The image will carry metadata showing it was AI-processed, which insurers routinely reject."}],"correct":false,"tempting":[{"text":"Invents a rejection policy; the real problem is in the pixels, not in a metadata flag."}]}],"order":[1,2,0,3],"why":null},"prev":{"slug":"spam-fraud-and-the-oldest-ai","title":"The models that have been working since 2002","minutes":9},"next":{"slug":"autocomplete-maps-and-the-keyboard","title":"Typing, routing, and systems that change what they measure","minutes":8}}}