日常生活中的人工智能：它已经在哪些方面发挥作用以及如何正确使用

Artificial intelligence is already part of most people's day, often without being labeled. The useful way to see it is by what each feature does. Some AI predicts, like an estimated arrival time. Some classifies, like deciding whether an email is spam or whether a face matches the phone's owner. Some recommends, like the next video or product. Some generates, like a chatbot answer or a suggested reply. And some acts, like a car that drives itself. Each type has different stakes: a wrong film suggestion costs a few minutes, while a wrong action involving money, health or safety can cost much more.

Most of these features follow the same basic recipe. A model is trained on large amounts of past examples, then applied to new situations in real time. Navigation is a good illustration. According to DeepMind's 2020 description of its work with Google Maps, the system combines live traffic on each stretch of road with historical patterns for that time and day, groups neighboring roads into larger units called supersegments, and uses a graph neural network to predict travel time for each one. The routing system then ranks candidate routes by those predictions. DeepMind reported that the approach reduced inaccurate arrival-time estimates by up to 50% in cities such as Berlin, Jakarta, Sao Paulo, Sydney, Tokyo and Washington, D.C.

Other everyday features work in similar ways. A spam filter is a classifier: it scores each incoming email on features such as sender reputation, links and wording, and moves high-scoring messages away from the inbox; Google says Gmail blocks 99.9% of spam, phishing and malware. Face unlock is a matching system. Apple explains that its TrueDepth camera projects thousands of invisible dots to capture depth, a protected part of the chip's neural engine turns that into a mathematical representation, and it is compared with the enrolled face, with separate anti-spoofing networks to resist masks. Phone keyboards learn from typing. In 2017 Google Research described federated learning in its Gboard keyboard: each phone improves the model using its own data, and only a small summary of changes, averaged with many other users' updates, is sent to the cloud.

The newest layer is generative AI, which creates text, images or speech rather than only scoring or ranking. Chat assistants answer questions and draft messages, phones summarize notifications, and accessibility tools describe the world. Be My Eyes, an app that has connected blind and low-vision users with 6.3 million sighted volunteers, introduced an AI Virtual Volunteer built on OpenAI's GPT-4 in March 2023, which answers questions about photos and offers to connect the user with a human when it cannot help. At the far end of the scale are systems that act in the physical world. The Stanford AI Index 2025 reports that Waymo was providing more than 150,000 driverless rides each week, and that the US Food and Drug Administration approved 223 AI-enabled medical devices in 2023, up from six in 2015.

It helps to compare AI features with the alternatives. A paper map or fixed timetable is predictable but cannot react to an accident ahead; an AI arrival estimate adapts but can be confidently wrong. A rule-based filter, such as blocking every email that contains one word, is easy to understand but easy to evade, while a learned filter catches new tricks but occasionally hides a real message in spam. Processing on the device, as with Face ID data that Apple says never leaves the phone, keeps personal data local, while cloud-based assistants can do more but send information to a company's servers.

Every one of these systems makes mistakes, and the mistakes have patterns. Apple notes that Face ID is less reliable at telling apart identical twins, siblings who look alike and children under 13. Recommendations can trap people in narrow loops of similar content. Chatbots can state false information with confidence. AI also helps scammers: in March 2023 the US Federal Trade Commission warned that criminals use voice cloning, which needs only a short audio clip, to impersonate a relative in trouble and demand money. The FTC's advice is to hang up and call the person back on a number known to be theirs, and to treat requests for wire transfers, cryptocurrency or gift cards as warning signs.

Telling when AI is involved is getting harder, so a few checks help. Features that summarize, suggest, rewrite or answer in full sentences are usually generative and can be wrong in fluent language, so treat their output as a draft. Results that rank or filter, such as search, feeds and spam folders, decide what is seen and what is hidden, so check the spam folder and other sources when something important seems missing. Anything that can act, such as sending, buying or unlocking, deserves a confirmation step and a way to undo. Workplaces add another layer: the Stanford AI Index 2025 reports that 78% of surveyed organizations used AI in 2024, up from 55% a year earlier, so the same tools increasingly shape job applications, customer service lines and work software as well as personal devices. When an unexpected call, message or video asks for money or credentials, the FTC's rule of thumb applies broadly: verify through a separate channel before acting.

Most everyday AI costs nothing extra at the point of use, because it is bundled with a phone, an email account or an app. The real price is usually data and attention: what the service collects, how long it keeps it, and what it uses it for. Practical habits help. Review app permissions for location, microphone, camera and contacts, and turn off what a feature does not need. Look for on-device options for sensitive tasks. Keep human confirmation for anything that spends money, sends messages or changes accounts. Check important answers from chatbots against a trusted source. And keep a manual fallback, such as knowing the route or having a backup unlock code, for when the system is down or wrong.