"Our AI Doesn't Learn from You and Your Data is Local" - Blatant Lies from Big Tech

Summary of Main Points and Arguments

The video argues that major tech companies (Microsoft, Apple, Google) use privacy-policy wording to mislead users about how on-device AI functions. It focuses on claims such as:

• “AI doesn’t learn from your data”
• “your data stays local”

“Doesn’t learn” and “local” are framed as misleading

The presenter claims companies intentionally highlight two “feel-good” terms:

• “Learn” (e.g., “we don’t learn from your content/data”)
• “Local” (e.g., “data stays on your device”)

The argument is that privacy risk does not require the AI model to be trained (“learn”) on your data. Instead, the AI can still:

• access
• search
• use your data to generate outputs

Those outputs may then be transmitted back to the company.

Data collection exists even outside user actions

The video stresses that “data is data,” arguing that once personal identifiers exist (accounts, phone numbers, IDs), resulting data can be persistent.

It cites examples to support the idea that surveillance can come from:

• Cameras (e.g., license plate tracking via surveillance infrastructure)
• Location data from apps (e.g., location tracking even without deliberate exports)

Windows Recall / Copilot-style features as the key example

The presenter highlights Windows Copilot + Recall, using periodic screenshots to claim that a dataset of user behavior and preferences is created.

Even if Microsoft says the AI “doesn’t learn,” the video argues the AI still can:

• read screenshots/activity/history
• build searchable representations of that information
• expose personal patterns to parties other than the user

AI explanation: trained model vs. “searching” your data (RAG)

To support its technical claim, the video distinguishes between:

• Trained base models: trained earlier, with weights generally fixed
• Runtime incorporation of new information: done via RAG (Retrieval-Augmented Generation)

In this setup, RAG combines the model with a retrieval/search step.

It further distinguishes retrieval methods such as:

• Web search (public information)
• Vector embeddings (described as a “private search engine”)

The presenter concludes that companies can avoid “learning” from your data while still making your data usable to the system through retrieval/search and embeddings.

The “real risk” is surveillance capability, not model retraining

The central claim is that the combination of:

• data availability
• searchable embeddings
• AI-assisted connections

creates a surveillance threat.

The video extends this analogy to larger systems (e.g., Palantir), suggesting such organizations use:

• an AI model
• plus embedding-style indexing over large databases

to locate and interpret individuals/groups more effectively than humans.

“Local storage is not isolation” (architectural flaw claim)

The presenter argues that “processed locally” does not guarantee privacy because:

• on-device AI can read local data
• it can also communicate results outward (summaries, alerts, matches, tool instructions, updates)

So, local storage may only protect the raw data, while the actionable outputs can still be transmitted.

The video labels this “client-side scanning” / “embedded AI” architecture as the core danger.

Policy and trust portrayed as deliberate wordplay

The presenter concludes that the combination of:

• legal/marketing language (“doesn’t learn,” “local”)
• embedded/search capabilities

is designed to keep users comfortable enough to adopt features.

Stated takeaway: the only “safe” AI is AI the user controls themselves (the presenter claims using Linux and restricting embedded AI).

Presenters or Contributors

• Rob Braxman (primary speaker/presenter; referenced indirectly via branding such as “Rob Braxman Tech Deep Dive,” “brax.me,” and “braxtech.net”)