Guides
AI Features on Phones and Laptops: Hype Versus Evidence
Learn how to tell where an AI feature on a phone or laptop actually runs, what its maker promises, and which FTC cases show why accuracy and privacy claims need evidence before you pay.
New phones and laptops now arrive with "AI" printed on the box, the product page and often the keyboard. Some of those features do real work: they rewrite a paragraph, find a photo from a description or translate a conversation. Others are an old feature with a new label. This guide is for people who are about to buy a device, or are deciding which features to switch on, and who want a way to sort the useful from the decorative. It covers where an AI feature runs, what the makers say about it, what the US Federal Trade Commission has said about unsupported claims, and a short check you can do in a store or at home. It does not rank products or predict which feature will matter in two years.
First question: where does the work happen?
Most of the practical differences between AI features come down to location. A feature can run on the device itself, in a company's data centre, or partly in each. That choice affects speed, battery life, whether it works offline, and what happens to the text or photo you give it.
Apple's own support page for Apple Intelligence on Mac describes the split plainly. It says that in many cases the models run entirely on the device, so a task can be completed without data leaving it. For more complex requests that need more computing capacity, the system can use what Apple calls Private Cloud Compute, which runs larger, server-based models. Apple states that data processed there is not stored or made accessible to Apple, and that independent researchers can verify this. That is the company's description of its own system, and it is worth reading as exactly that: a stated design, with a stated way to check it.
The same page carries a limit that is easy to miss. It notes that the feature is not available in all languages or regions and that usage limits may apply. The page names macOS 27 and shows no date, so check the current version for your Mac. A feature shown in a launch video may not exist, or may behave differently, where you live.
What the hardware number does and does not tell you
Laptop makers lean on a chip called an NPU, a neural processing unit. Microsoft's developer guide for Copilot+ PCs, a page listed with a November 2025 date (the page text itself shows none, so check the current version), describes it as a specialised chip for AI-heavy work such as real-time translation and image generation, and says it can perform more than 40 trillion operations per second, abbreviated TOPS. The same guide says many of the new Windows AI features require an NPU able to run at 40+ TOPS.
Take two things from that. First, the number is a real entry requirement for some Windows features, so on a Windows laptop it is a fair thing to look for if you want them. Second, it is a measure of raw throughput, not of how good a feature is. A figure of trillions of operations per second does not say whether the summary is accurate, whether the photo edit looks natural, or whether the feature is one you will use. Treat it the way you treat any peak figure on a spec sheet, as a ceiling to be read together with the conditions. The habits in how to read a tech spec sheet apply directly, and units and labels compared helps when a number comes without a unit you recognise.
What regulators say about AI claims
An "AI" label is a marketing claim, and in the United States marketing claims fall under the FTC. Its recent cases show what unsupported ones look like.
One concerns a company that sold a tool to detect whether text was written by a person or by AI. The FTC's April 2025 announcement says the order settles allegations that the company promoted the tool as "98 percent" accurate. According to the FTC's administrative complaint, independent testing found an accuracy of just 53 percent on general-purpose content, and the FTC alleges that the model had been trained or fine-tuned to classify only academic content. In August 2025 the Commission approved a final order. It bars the company from claiming the effectiveness of any AI content detection product unless the claim is not misleading and the company has competent and reliable evidence for it at the time, and it must keep that evidence.
Another entry, an FTC case page, covers a settlement of allegations that companies falsely claimed to offer an AI-powered service that could target ads based on conversations captured from consumers' smart devices, and falsely claimed that consumers had opted in. These are allegations resolved by a settlement, not findings in a trial, and the page lists the matter's status as pending.
The FTC also keeps a topic page that collects its AI-related actions, including cases about an "AI lawyer" service and about facial recognition software said to be free of bias. The pattern across them is consistent. A specific claim, such as an accuracy rate, a privacy promise or a professional-level result, needs evidence behind it, and the evidence has to exist when the claim is made. For you as a buyer, that suggests a simple rule: the more precise the promise, the more reasonable it is to ask where the proof is.
Reading a typical claim
Suppose a phone's product page says a new photo tool "removes any object perfectly." The sentence invites several questions. Perfectly in what conditions, and judged by whom? Does it run on the phone or send the picture to a server? Is it free, or limited, or tied to an account? Which languages, regions and phone models get it? A page that answers these in plain words is giving you something to work with. A page that answers none of them, and offers only adjectives, is telling you something too.
Another illustrative case is a laptop advertised with "all-day AI battery life." Battery life depends on the workload and settings, so ask which test produced the figure and whether the AI features were switched on during it.
A short check before you pay or switch on
Use this list with the device in hand, or with the vendor's documentation open.
- Name the feature in one plain sentencewhat goes in, what comes out.
- Find out where it runson the device, in the vendor's cloud, or both. The vendor's privacy or support page should say.
- Check whether it needs an account, a subscription or an internet connection, and what happens to it if the service changes.
- Check availability for your language, region and exact model, since a headline can cover more devices than the feature does.
- Look for the condition behind any numbertest method, settings, date.
- Decide whether you would use it weekly. If not, do not pay extra for it.
- Test it with something harmless first, and check the result against what you know.
Privacy settings and checking what leaves the device
Once a feature is on, you can often look at what it actually does. Apple's support page describes an Apple Intelligence Report on the Mac: in System Settings, under Privacy & Security, you choose Apple Intelligence Report and then a duration, the last 15 minutes (the default) or the last 7 days, or switch the report off. It lists requests your Mac has sent to Private Cloud Compute, and Apple notes it may be empty if there have been no such requests since you changed the duration. Settings differ by vendor and software version, so look for the equivalent in your own device's privacy section, and read the vendor's description of each AI feature before enabling it.
General device hygiene still matters more than any single AI setting. Keep the system updated, as the software updates guide explains, and keep the basics of consumer device security in place, because an AI feature that can read your messages or screen is only as safe as the account and device around it. If you are setting up a new phone, the smartphone basics guide covers the first-day choices where these toggles usually appear.
