AI, Actually Explained · Chapter 02

Where it already is, without a label

जहाँ एआई पहले से मौजूद है, बिना किसी नाम-पट्टी के

10:41 · Narrated with a synthetic voice. Captions are available in the player.

Read the transcript

A day can be shaped by predictions you never asked to see. Which message reaches you, which road you take, which photograph you keep. None needs a robot, or a glowing badge that says AI. Welcome to Addaly. In our last chapter, we separated written rules from behaviour fitted through training. Now we will follow one ordinary day through a city of quiet decisions. At each stop, ask what is being predicted, what the system is trying to improve, and who carries the cost of a mistake. Imagine a phone opening into a small city. A message passes through a gate; a photograph develops in a workshop; a route lights up between buildings. These are illustrations, not a diagram of a particular company’s systems. A useful chain is input, prediction, action, consequence. An email supplies text and other signals; a filter estimates whether it is unwanted; the service decides where to put it. Some steps may use trained models, others written rules, and many products combine both. The same error rate can have very different consequences when the action changes. A misplaced suggestion and a blocked payment do not ask the same thing of the person affected. So neither a familiar interface nor the label AI tells us enough. Our first stop is a shelf that rearranges itself: the feed. A recommender may narrow a huge collection to candidates, then score and rank them. A research paper described this two-stage design for YouTube in 2016; it is an example, not a blueprint for every feed today. What counts as a good ranking depends on the objectives and signals chosen. Watching, skipping, likes, surveys and other signals can matter differently across systems. YouTube’s twenty twenty-one explanation included satisfaction surveys and safeguards, alongside watch time. So we should not assume that every feed optimises only minutes watched, or that likes never count. Still, predicting that you will watch something is not the same as establishing that it is true or useful to you. Look for the service’s documented history, recommendation and feedback controls; their effects vary. The shelf folds into a gate, sorting incoming messages and payments. A spam filter or fraud detector can turn signals into a score, then compare that score with a threshold. Moving the threshold changes the balance between missed problems and false alarms. Here is an invented example, not a report about a real bank. Out of one million transactions, suppose one thousand are fraudulent. Suppose the detector catches ninety-nine percent of the fraud, and correctly clears ninety-nine percent of the genuine transactions. It catches nine hundred and ninety fraudulent transactions. But one percent of the much larger genuine group is nine thousand, nine hundred and ninety false alarms. Only about nine percent of all the flagged transactions are actually fraudulent. The missing information was how rare fraud was, and which kinds of error the percentages described. A headline saying ninety-nine percent accurate does not, by itself, answer those questions. As attackers and ordinary behaviour change, evaluation needs to keep up; there is no universal retraining schedule. Beyond the gate, light enters the photograph workshop. A phone image may pass through exposure control, several captured frames, alignment, merging, noise reduction and sharpening. Some pipelines use learned components; others combine them with engineered algorithms. Google’s published HDR plus research, for example, described burst alignment and filtering without making every step a trained model. These methods can make dim scenes easier to see. But enhancement can also change fine detail, and some generative tools can add plausible detail that the camera never captured. If a tiny crack matters, beauty is not the only test. Keep the original file and record how it was captured and processed; follow the recipient’s evidence requirements. A RAW label alone is no guarantee of untouched evidence: Apple ProRAW deliberately combines raw information with computational processing. The workshop’s lines become roads, then the paths between words on a keyboard. A typing assistant may predict a likely continuation from the text around the cursor. Likely does not mean intended: read the sentence before you send it. Some features run on the device, others contact a service, and some combine both. The interface alone cannot tell you which; check that product’s data-use information and settings. A route planner may use traffic estimates and road information to compare routes. In our invented city, many drivers follow the same quiet shortcut, making it busy. That is a feedback loop: acting on a prediction can change what future measurements show. It is one possible explanation for a changing route, not proof of why any real app changed its advice. At the end of the road, a message crosses a bridge between languages. A translation system must preserve meaning while producing a natural sentence in another language. Performance depends on the language pair, available data, subject and system. A smooth sentence can still change a name, a number, a condition, or the word not. Consider an invented delivery note: do not leave the parcel outside. Lose the negation, and the instruction reverses, even if the grammar stays elegant. Compare important names, quantities and conditions with the original. Translating back, or asking a second model, may reveal a problem; agreement does not prove the meaning survived. For consequential material, use a competent bilingual reviewer with the relevant subject knowledge. The message becomes a sound wave, passing through another prediction system. Speech recognition turns audio into text, but a transcript is still an output to check. Accents, languages, noise, overlapping voices and unfamiliar names can affect the result. Word error rate counts substitutions, deletions and insertions against a reference transcript. That is useful, but it does not tell you which mistake matters most. Changing do not approve to do approve is only a small edit, and a large change in meaning. Some systems, including Whisper, can also produce words that were never spoken. Punctuation and speaker labels need checking too, especially when a pause or a change of speaker affects the meaning. Keep the audio, mark uncertainty instead of guessing, and listen back to important names, numbers and negations. For accessibility, captions must describe the actual recording, not merely repeat the intended script. Our sound wave reaches a desk where an output may affect someone’s opportunities. Imagine an application marked for review because information is missing. If staff treat the flag as proof of wrongdoing, a prediction has become an accusation. This is an invented process, not a claim about a particular school, lender or government. The questions are larger than whether the algorithm is clever. What information was used, was it correct, what action followed, and can a responsible person review the case? A human in the workflow helps only when they have the information, time and authority to challenge the output. Written rules can create harmful decisions at scale too; harm does not require machine learning. Available explanations, review and appeal processes vary by service and place, so check the actual process rather than assuming one exists. The desk opens onto a workshop, a field and a clinic. A factory tool might flag unusual vibration; a farming tool might classify a leaf image. Each is a particular task, not a general understanding of the machine or the crop. A model tested on neatly framed leaves can struggle when the lighting, background or camera changes. The useful test includes real working conditions and what happens after a flag. In healthcare, different systems have different intended roles: some assist review, while some authorised devices perform narrowly defined tasks autonomously. Neither means they can diagnose everything, for everyone, with any equipment. A result must be judged against the particular task, population, equipment and workflow for which the system was evaluated and authorised, where applicable. This is a lesson about evaluating tools, not medical or treatment advice. Back in the city square, products advertise intelligent help. Changing wording, a typing animation or a warning about mistakes may suggest a generative feature, but none proves what is running underneath. Ordinary code can vary its wording, and a model can return the same answer twice. Offline behaviour is not decisive either: a feature may use local models, cached results or written rules. Look for product documentation, model information and data-use explanations from the provider. When information is missing, write unknown; do not turn a clue into a fact. Then return to the decision chain: input, prediction, action, consequence. Even without identifying the exact model, you can still ask what evidence supports its output and what to do if it is wrong. Now make something useful: a one-page audit of a familiar feature. Paper is enough; you do not need a new account or a paid tool. Choose a low-stakes example, such as a reading recommendation, and keep personal or sensitive information out. Write the feature, the input it appears to use, its output, and the action that follows. Separate what you observed from what the provider documents, and mark everything else unknown. Add the stated objective, if you can find it; a plausible guess is still a guess. Describe one false alarm and one missed case, who would be affected, and how serious each could be. Finally, name one available control or checking step, and one question you would ask the provider. Explain your audit to someone else; ask them to find an assumption you treated as a fact. Pause here to try it. The city’s gates, roads and bridges now fit on that single page. You do not need to call every digital feature AI, or dismiss every prediction because it can be wrong. Find the task, follow the action, and choose a check that matches the consequence. The written chapter has further examples and practice; use it to deepen the parts you want to examine. Keep your audit: we will keep asking how a system earns trust, rather than how impressive its answer sounds. You have been learning with Addaly. Sources and the activity are included with this chapter.

Sources and further reading (8)