Samsung brings Galaxy AI to the Galaxy S24

Samsung unveiled the Galaxy S24 family on January 17, 2024, placing “Galaxy AI” at the centre of the launch. The label covered call translation, face-to-face interpreting, writing and summarisation assistance, visual search and generative photo editing. The durable lesson is not the list itself. It is that mobile AI features can follow very different technical paths even when they share one commercial name.

The useful question, then, is not whether a phone “has AI.” It is what each tool receives, where it processes that input, who receives the data and what happens when connectivity disappears. Those four questions distinguish a local capability from a remote service and help a person decide before exposing a private call, a work recording or a personal photograph.

A brand is not an architecture

Samsung’s original Galaxy S24 announcement described experiences with different routes. Live Translate provides two-way voice and text translation inside the native phone app and, according to the company, uses on-device AI. Interpreter translates in-person conversations and can work without mobile data or Wi-Fi. Circle to Search, by contrast, is a Google collaboration that turns a screen selection into a search. Generative Edit requires a network connection and a Samsung account to fill or reposition parts of an image.

“On-device” means that a substantial part of the computation happens on the phone. It can reduce data exposure, latency and dependence on the internet, but it does not automatically guarantee accuracy or complete confidentiality. The operating system, the application, backups and other services remain part of the environment. “Cloud” does not automatically mean unsafe either: it describes a processing location, not the rules for retention, access or reuse. Those rules require a separate check.

Smaller models make some of this division possible. Google had explained in December 2023 that Gemini 1.0 came in Ultra, Pro and Nano sizes, with Nano optimised for on-device tasks. That does not establish which specific model runs every S24 feature. It does illustrate the engineering choice: a mobile-sized model can improve speed, privacy and availability, while a remote system normally has more compute and can be updated without replacing the handset.

Three features, three data paths

Live Translate and Interpreter are the clearest cases of announced local processing. Interpreter’s ability to work offline creates a practical test: download what is required, enable airplane mode and see whether translation remains available. This identifies technical dependency, not quality. Names, jargon, accents, noise and ambiguous sentences can still produce mistakes. The cost may be modest for a restaurant booking; in a medical, legal or financial conversation, automated translation should not replace human confirmation.

Circle to Search follows another route. A user circles, taps or highlights an item on screen and receives results without leaving the visible app. Google said on January 17 that the feature would launch on January 31 on the Galaxy S24 and Pixel 8 families, and that relevant ads would continue to appear in dedicated result-page slots. The uninterrupted interface can feel local, but this is a search: the selection and the context needed to answer travel to Google’s service. Staying inside one app does not mean the data stays on the phone.

Generative Edit introduces a third pattern. Samsung said it could straighten an image, move objects and fill the resulting space, but required a connection and an account. Resulting images carry a visible watermark and metadata to indicate generative editing. Those signals do not prove an image is truthful, and cropping can remove a visible mark; they are provenance evidence that can be lost. Before sharing an informative photograph, a careful user should preserve the original, the file containing metadata and a record of what changed.

The seven-question card

Any mobile AI feature can be audited with a short card. First, what exactly is the input: voice, text, image, the entire screen, contacts or location? Second, which part is processed on the handset? Third, if data leaves, which company and service receive it? Fourth, how long is it retained and can it improve models? Fifth, which permission or setting limits transmission? Sixth, what stops working offline? Seventh, how is the output labelled and what human review does it require?

This card avoids two common shortcuts. The first is treating “private” as a property of an entire feature family when the promise may cover only one tool. The second is assuming that everything integrated into an operating system remains on the device. Samsung provides an Advanced Intelligence setting to process data only on the handset, but its own footnote warns that restricting online processing may limit some features. Enabling it and observing what disappears turns a general claim into a testable map.

Samsung’s statement that it does not save user input or output, regardless of server access, must also be read precisely: it is the manufacturer’s claim about its services, not an automatic description of every participating third party or a promise about future policies. Google also participates in Circle to Search, so each recipient needs its own answer. For sensitive information, the account in use, synchronised backups and an organisation’s terms matter alongside the public policy.

Decide according to the cost of error

The same data path can be suitable for one task and wrong for another. Identifying a lamp from a public photograph does not expose the same information as circling a name in a medical record. Translating a tourist greeting does not require the certainty of interpreting a medicine dose. Before enabling a feature, classify both the sensitivity of the input and the harm from a wrong output as low, recoverable or severe. When either axis is high, minimise the data, anonymise the sample where possible and require an independent check.

Consent from the person on the other side matters as well. Local translation reduces one exposure route, but it does not remove the value of telling someone that a call is being transcribed or translated. In a meeting, whoever turns on an assistant should be able to explain what it captures, where it operates and which file remains afterwards. Technical architecture helps; it does not replace the social, professional or legal rules of the setting.

Privacy, availability and accuracy are separate axes

A tool can be local and wrong. It can be remote and have strong controls. It can protect content yet become unavailable exactly when coverage fails. A single label hides those differences. A better assessment scores data exposure, offline availability, quality in the user’s own language and context, the ability to correct the output and clarity about responsibility when something fails.

A useful test takes fifteen minutes. Choose a non-sensitive sample, run it while connected, repeat in airplane mode, inspect permissions and settings, and compare the result with a known reference. Then repeat with a difficult case: a name, background noise or a partly obscured image. This small protocol reveals more than a carefully staged demonstration and transfers to any future phone, assistant or application.

The Galaxy S24 therefore marked an important transition: AI stopped being a separate app and became entangled with calls, keyboards, cameras and search. That convenience makes the boundary between device and server less visible. The enduring skill is learning to draw it. Whenever one brand groups many features under one word, follow each piece of data from input to output and make a decision about that specific route, not the slogan.