Every AI app claims to respect your privacy. The claims are made in near-identical language, they are almost never false in any provable way, and they are almost never specific enough to be useful.
The way out is to stop reading claims and start running tests. Privacy that comes from architecture can be verified by you, on your own device, in about ten minutes. Privacy that comes from policy cannot be verified at all — which is itself the most useful thing to know.
This article sits under Private ChatGPT Alternatives.
Test 1: Airplane mode
The single most informative test, and it takes seconds.
Install the app, download a model if it asks, then enable airplane mode and keep using it.
Still working? Inference is happening on your device. Your conversation cannot be transmitted because there is nothing to transmit to. This is not something an app can fake.
Stopped working? The model runs on a server, and your conversations go there. That may be perfectly acceptable to you — but now you know which category you’re in.
No privacy policy can substitute for this test, and no marketing language can survive it.
Test 2: Does it need an account?
Accounts exist to associate activity with a persistent identity. Cloud AI genuinely needs them — to meter usage, bill customers, and enforce rate limits.
An app that processes everything locally has none of those needs. There is no usage to meter because you supply the compute, and no bill to recover because the marginal cost is zero. AI Chat With No Account works through the economics.
So the requirement is diagnostic. It doesn’t prove anything on its own, but combined with test 1 it is close to conclusive.
Test 3: The App Store privacy label
Apple requires every developer to declare what data an app collects and whether it is linked to the user’s identity. Open the App Store listing, scroll to App Privacy, and read it before installing.
Three tiers, in descending order:
- “Data Not Collected” — the strongest declaration available, and one Apple can act on if false.
- “Data Not Linked to You” — something is collected, but not tied to your identity.
- “Data Linked to You” — collected and associated with you. Note which data.
Labels are self-declared rather than audited, so treat them as a signal rather than proof. They are most useful for what they rule out: an app declaring “Data Not Collected” while requiring an account and a connection is worth examining closely.
Test 4: Look for a specific technical claim
This distinguishes real engineering from reassurance, and it works because vagueness is expensive to fake.
Vague — describes protection, not collection:
- “Your privacy is our priority”
- “Enterprise-grade security”
- “Bank-level encryption”
- “We never sell your data”
Every one of these is compatible with storing all of your conversations indefinitely. “We never sell your data” is particularly slippery: it says nothing about retention, analysis, training, sharing with partners, or disclosure under legal process.
Specific — describes architecture:
- “Runs on-device via Apple MLX”
- “No network requests during inference”
- “Model weights stored locally; conversations never transmitted”
- Named models and named frameworks
Specific claims are checkable and carry consequences if untrue. Vague claims are unfalsifiable, which is usually why they were chosen. The AI Privacy Checklist covers more of this vocabulary.
Test 5: Read the retention section
Open the privacy policy — the real one, not the marketing page — and search for retain, train, third party, and delete.
What you are looking for is volume. An app with no server has very little to say about retention, because there is nothing to retain. A policy with several detailed paragraphs on retention periods, deletion timelines, sub-processors, and training use is describing a system that stores your conversations. That may be handled responsibly, but the storage is happening.
Then check one more thing: whether the policy can be changed unilaterally. Nearly all can. That is the structural weakness of any policy-based guarantee — today’s terms describe today’s intentions, and both can be revised. Cloud AI Data Policies Compared covers how to read these documents.
Test 6: Follow the money
Server-side inference costs real money per request. That cost is continuous, and it must be covered.
If an app runs models remotely, is free, has no subscription, and shows no ads — identify what pays the bill before trusting it. Sometimes the answer is benign: venture funding, a free tier feeding a paid one, a loss leader for a larger product. Sometimes it is your data. The point is to know rather than assume.
On-device apps sidestep this entirely: the compute is yours, so there is no recurring cost to fund.
Test 7: Watch the network
For the technically inclined, direct observation beats every other test.
The accessible version: use the app offline for a long conversation, then check Settings → Cellular → app name on iOS to see data usage. A local model generates no request traffic while answering. Model downloads are large and one-off; conversations should add nothing.
The thorough version: put the device behind a monitoring proxy or use a network inspection tool and watch what leaves during a conversation. A genuinely local app produces nothing.
This is the closest thing to proof available to an end user, and it is why on-device claims are qualitatively different from policy claims. They are testable.
The distinction that all seven tests are measuring
Every test above is a way of asking the same question: is this guarantee architectural or contractual?
A contractual guarantee says the company will not misuse data it has received. It depends on current intentions, current ownership, current jurisdiction, and current security. It can be revised, and it can fail without anyone acting in bad faith.
An architectural guarantee says the data was never received. It doesn’t depend on intentions, and it cannot be revised by a policy update — changing it would require shipping a fundamentally different app, which you would notice.
Both are real. Only one can be verified by the person relying on it.
Cloaked is built for the second kind: 11 open-weight models running on-device via Apple MLX, no account, no server, and a privacy architecture that you can confirm in airplane mode rather than take on faith.
Download Cloaked on the App Store and run test 1 yourself — it takes about thirty seconds.
Frequently asked questions
Do any AI apps genuinely not collect data?
Yes — apps that run models entirely on your device. Because inference happens locally and no conversation is transmitted, there is nothing to collect. The distinguishing test is whether the app still works with no network connection.
Are App Store privacy labels reliable?
They are self-declared by developers, so they are a useful signal rather than an audit. They are most informative when they show something: an app declaring that it collects nothing and links nothing to your identity is making a statement Apple can act on if it proves false.
Does 'end-to-end encrypted' mean an AI app is private?
Not by itself. Encryption protects data in transit from third parties, but the service still decrypts and processes your conversation to generate a response. Encryption is about who else can read it, not about whether the provider does.