Most writing about AI use cases is a list of things a model can do. Summarize a document. Write an email. Explain a concept. All true, and all equally true of every assistant, which makes the list useless for deciding which one to use.
The interesting question for private AI is narrower: what do people do with it that they wouldn’t do with a cloud assistant?
The answer turns out not to be about capability at all.
The thing nobody mentions: you edit yourself
Watch how someone actually uses a cloud AI assistant with something personal and you see a small, near-universal behaviour. They start typing. They stop. They delete the company name and write “a company.” They change “my daughter” to “a family member.” They replace real figures with round ones. They describe the situation one layer more abstract than it really is.
Almost nobody notices they’re doing it. It is the same instinct that makes you phrase things carefully in a work email.
But it has a direct cost: the model answers the sanitized question, not the real one. Specificity is what makes an answer useful. “I earn a variable income between £3,200 and £5,800 a month and I’m trying to work out a mortgage I can service in a bad quarter” gets a genuinely useful answer. “Someone with an irregular income wants mortgage advice” gets a Wikipedia summary.
Remove the possibility of transmission and the editing stops. That is the actual shift — not better answers from a better model, but better answers because you asked the real question.
Every use case below is a variation on that.
Personal: the questions you’d rather not file anywhere
Health. Symptoms, test results, medication interactions, what to ask a doctor, what a letter from a specialist actually means. People research health obsessively and are acutely aware that health data is valuable and permanent. On-device vision helps here too — photographing a prescription label or a results page and asking about it, with EXIF and GPS stripped locally. A model is not a doctor and shouldn’t be treated as one, but “help me understand this before my appointment” is legitimate and common.
Money. Actual salary, actual debt, actual rent, actual savings. Budget maths, offer comparisons, understanding a pension statement. This is precisely the information people are most careful about disclosing.
Relationships and family. Drafting a difficult message. Working out how to raise something with a partner. Preparing for a conversation with a parent about care. Understanding a conflict from the other side.
Career. Whether to leave, how to negotiate, what a contract clause means, how to phrase a resignation, whether an offer is good. Frequently involves naming a current employer while they are still your employer.
Asking AI the Questions You’d Never Type Into a Search Bar goes deeper on this category.
Professional: where disclosure itself is the problem
For some jobs this isn’t preference — it is obligation, and the distinction is sharp.
Law. Client facts are privileged. Pasting them into a third-party service can be a disclosure regardless of that service’s policies or intentions. The exposure is the transmission.
Medicine and therapy. Patient information is protected by regulation in essentially every jurisdiction. Practitioners want help drafting notes, restructuring a letter, or thinking through a formulation — with no third party involved.
Accounting and finance. Client financials, valuations, positions.
HR. Grievances, investigations, performance issues, restructuring plans — all involving named individuals who did not consent to disclosure.
Engineering and product. Unreleased plans, proprietary code, security details, pre-announcement material.
The common structure: the question is not “is this provider trustworthy?” but “did disclosure occur?” A contractual guarantee doesn’t answer that. Architecture does — which is the argument in Private ChatGPT Alternatives.
Creative: work that isn’t finished
Unfinished creative work is private in a way finished work isn’t. A first draft is embarrassing by design.
Novelists working on a manuscript years from publication. Screenwriters with a premise they don’t want circulating. Songwriters. Academics with an unpublished argument. Journalists protecting a story and, more seriously, a source.
For all of them the concern isn’t that a provider would misuse the material. It is that unpublished work should exist in exactly one place until the author decides otherwise. Private AI for Writers covers the workflows in detail.
Reflective: thinking out loud
A distinct category, and one people underestimate until they try it.
Journaling with a model that responds. Working through a decision by explaining it. Debriefing a bad day. Preparing for a difficult conversation by rehearsing it.
This only works if the space is genuinely private. A journal you suspect is being read is not a journal — it becomes performance, and the value evaporates. Private AI Journaling covers how this works in practice and where its limits are.
Practical: when there is no connection
The plainest category. No signal, no cloud AI. On-device models don’t care.
Flights. Trains through tunnels. Rural areas. Basements and lifts. Foreign travel without roaming — where offline translation and local guidance are genuinely valuable. Hospitals with poor reception. Countries where a given service is blocked or unavailable.
Also worth counting: outages, rate limits, and capacity messages. A local model has no bad days.
Everyday: the unglamorous majority
Most use isn’t dramatic. It is quick, constant, and low-stakes:
Rewriting a message to sound less blunt. Summarizing a long email thread. Explaining an error. Converting units. Naming things. Drafting a reply. Checking whether a paragraph makes sense. Translating a sign.
None of it is sensitive. It doesn’t need to be — the model is already there, it works instantly, and it costs nothing per use. Privacy just means you never have to decide whether a given question is sensitive enough to be careful about.
Where a local model is the wrong choice
Being straight about this is what makes the rest credible.
Frontier reasoning. Genuinely hard multi-step problems. Use a large model.
Very long documents. A phone’s context window is measured in tens of thousands of tokens, not hundreds of thousands.
Current events. Local models have a training cut-off. Search helps; it isn’t the same.
Specialist depth. Obscure technical, legal, and medical detail. Larger models simply know more.
Anything requiring tools. Code execution, image generation, live data.
ChatGPT vs a Local LLM works through the boundary carefully. The sensible setup for most people is both — local as the default, cloud for the handful of tasks that need it.
The pattern underneath all of it
Every use case here has the same shape. The value doesn’t come from the model being better. It comes from the model being usable without hesitation.
A tool you second-guess before every use is a tool you use at half capacity. Removing the hesitation is worth more than a few benchmark points — because the alternative isn’t asking a cloud model the same question. It’s asking a worse version of the question, or not asking at all.
Download Cloaked on the App Store — 11 models running entirely on your device, free, with no account and nothing transmitted.
Frequently asked questions
What is private AI actually useful for?
Anything you would otherwise rewrite before sending to a cloud service. Personal health and financial questions, confidential work material, unpublished creative writing, journaling, difficult conversations, and offline use where a cloud assistant simply doesn't function.
Is a local model good enough for real work?
For writing, summarizing, editing, translating, explaining, and structured transformation, yes. Modern 2–4B models handle these well. For frontier reasoning, very long documents, and specialist knowledge, a cloud model is still clearly better.
What is the biggest practical difference in daily use?
Self-censorship disappears. People pre-edit what they type into cloud assistants without noticing — removing names, softening details, describing situations abstractly. That editing costs answer quality, and it stops when transmission isn't possible.
Do professionals actually use on-device AI?
Yes, particularly in fields with confidentiality obligations — law, medicine, therapy, accounting, and HR. For these users the question is not whether a provider is trustworthy but whether disclosure occurred at all, and on-device inference means it did not.