AI privacy in 2026: what really happens to your cloud prompts

Most people assume their chatbot conversations are ephemeral. The reality is more complicated: prompts are logged, retained for days to years depending on the vendor and tier, reviewed by humans in abuse investigations, discoverable in litigation, and - under some default settings - used to train future models.

The four exposure paths

• Retention - conversations stored server-side persist beyond the session, sometimes indefinitely on consumer tiers.
• Training - opt-out mechanisms exist but defaults vary, and policies change faster than users read them.
• Legal process - anything a vendor stores can be subpoenaed. Your prompts are evidence someone else holds.
• Breach - centralized prompt logs are a high-value target. One incident exposes millions of users' most candid text.

Policy can't fix architecture

Vendors respond with better policies: shorter retention, enterprise no-training guarantees, regional hosting. These are real improvements - but they're promises, revocable and unauditable from the outside. The only privacy guarantee that doesn't require trust is data that never leaves your machine.

Local-first as the default

Gotchi runs inference, memory, and search entirely on-device. Conversations live in an encrypted local database. The app collects no telemetry and works fully offline - no license pings, no heartbeat. Your account exists only for licensing; it never sees a prompt. That's not a privacy policy. It's physics.

Privacy isn't a feature we bolted on. It's a constraint we designed around.

Reading a privacy policy like an engineer

Privacy policies are written to be reassuring; the discipline is reading them for mechanisms rather than sentiment. Four questions cut through: Where is the data processed - which legal jurisdictions, which sub-processors? How long is it retained, and is deletion actually deletion or just de-indexing? Is it used for training by default, and does the opt-out cover past data or only future prompts? Who can access it internally - is human review of conversations possible, and under what triggers? For every major consumer AI product, at least one of those four answers should give a careful reader pause.

The enterprise tier illusion

Vendors correctly point out that enterprise tiers offer stronger terms: zero retention options, no-training guarantees, regional processing. Two caveats. First, these are contract terms, not architecture - the same infrastructure processes your data, and the guarantee is a promise to configure it differently. Second, the strong terms cost enterprise money, which means the default consumer and prosumer tiers - where most actual usage happens - keep the weak ones. A guarantee you can't verify, on a tier you probably aren't on, is thin protection for candid text.

• Contract terms can change at renewal; architecture can't be renegotiated.
• Third-party audits attest to a moment in time, not to every future deployment.
• Even perfect vendor behavior can't protect data from legal process served on the vendor.

What zero-collection looks like technically

It's worth being precise about what 'local-first' means in Gotchi, because the phrase gets abused. Inference: the model runs in-process on your hardware; prompts are never serialized to a network socket. Storage: conversations, memory, and vector indexes live in an encrypted SQLite-based store under your OS user account. Telemetry: none - no crash reporters to third parties, no usage analytics endpoints, no heartbeat. Network: the app functions with the cable unplugged; the only outbound traffic is model downloads and license validation, both user-initiated. Any of this can be confirmed with a packet capture, which we encourage.

The regulatory direction of travel

Every regulatory cycle tightens the constraints on moving personal data through third parties: GDPR enforcement is escalating, the EU AI Act adds obligations for AI systems specifically, and sectoral rules in health and finance increasingly treat prompts as records. Organizations betting on cloud AI are betting that their vendor's compliance keeps pace with all of it, everywhere they operate. Local processing is the position that gets stronger with every new rule - data that never leaves can't fall out of compliance in transit.