> ## Documentation Index
> Fetch the complete documentation index at: https://stem-docs.intellectualpoint.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Data privacy
> Understand what data you sent to third-party servers this week, how to protect yourself, and when local models are required.
## Data privacy: what you shared and how to protect yourself
**Time**: 10:10 AM to 10:40 AM
Every API call you made this week sent data to a third-party server. Your voice, your questions, and images from the camera all left your Raspberry Pi and traveled to OpenAI's data centers. This section covers what that means and how to protect yourself.
***
### What you shared this week
Here is a concrete breakdown of every piece of data that left your robot:
| Day | What was sent | Where it went |
| ----- | -------------------------------- | -------------- |
| Day 3 | Text prompts and responses | OpenAI servers |
| Day 3 | Transcribed voice commands | OpenAI servers |
| Day 4 | Camera images (base64 encoded) | OpenAI servers |
| Day 4 | Images + text questions combined | OpenAI servers |
```mermaid theme={null}
flowchart LR
Robot["Your Pi"] -->|"text prompts"| OpenAI["OpenAI Servers\n(US data centers)"]
Robot -->|"voice transcripts"| OpenAI
Robot -->|"camera images"| OpenAI
OpenAI -->|"responses"| Robot
OpenAI -.->|"may be logged"| Logs["Usage Logs"]
```
By default, when you use the OpenAI API:
* Your prompts, completions, and uploaded images **may be logged**
* The account owner can opt out of training data usage, but this is **not the default** for all plan types
* Even "private" API usage still sends data over the internet to OpenAI's servers
* You do not control how long data is retained
***
### What was NOT shared
Not everything left the Pi. Here is what stayed local:
| Component | Where it ran | Data stayed local? |
| --------------------- | ------------ | ------------------------------------- |
| Vosk speech-to-text | On the Pi | Yes — audio never left the robot |
| espeak text-to-speech | On the Pi | Yes — audio generated locally |
| Motor control | On the Pi | Yes — commands never left the robot |
| Camera capture | On the Pi | Images only left when sent to the API |
Vosk is a privacy win — your raw voice audio never leaves the robot. Only the transcribed text gets sent to OpenAI.
***
### Practical protections
API providers have settings to control whether your data is used for model training. Check your provider's dashboard and disable training on your data where possible.
Do not include names, addresses, medical data, passwords, or any personally identifying information in prompts. Treat every API call as a postcard — assume someone can read it.
For private or sensitive use cases, run models locally so nothing ever leaves your machine. Tools like **Ollama** let you run LLMs on your own hardware.
Use `detail: "low"` for vision API calls (as we did). This sends less image data. Consider whether you need to send an image at all — sometimes a text description is enough.
Read the actual terms. Understand what the provider can do with your data, how long they retain it, and where it is processed.
***
### Local vs cloud: the tradeoff
| | Cloud API (OpenAI) | Local model (Ollama) |
| ------------ | --------------------------- | ---------------------------------- |
| **Privacy** | Data leaves your device | Data never leaves your device |
| **Quality** | Best available models | Smaller, less capable models |
| **Speed** | Fast (powerful servers) | Slow on Pi Zero (limited hardware) |
| **Cost** | Pay per token | Free after download |
| **Internet** | Required | Not required |
| **Control** | Provider controls the model | You control everything |
For this camp, cloud APIs make sense — the Pi Zero 2W does not have enough power to run useful LLMs locally. But for real-world applications with sensitive data, local models are often required.
***
### Enterprise and government context
Organizations handling sensitive data face strict requirements:
* **HIPAA** (healthcare) — patient data cannot be sent to unauthorized third parties
* **FERPA** (education) — student records have strict sharing rules
* **ITAR/EAR** (defense) — export-controlled technical data cannot leave US jurisdiction
* **GDPR** (EU) — personal data of EU citizens has strict processing and retention rules
* **FedRAMP** (US government) — cloud services must be certified before government use
In these contexts, sending data to a commercial API like OpenAI's may be prohibited. Organizations use private deployments, on-premise models, or government-certified cloud environments instead.
***
### Metadata is data too
Even if your prompt content is encrypted, the **pattern** of API usage can reveal information:
* **When** you make calls (timestamps reveal your schedule)
* **How often** (frequency reveals intensity of work)
* **How much data** (volume reveals project scope)
* **From where** (IP addresses reveal location)
Metadata patterns can be as revealing as the data itself. Intelligence agencies have long known this — it applies to AI systems too.
***
### Discussion questions
1. You sent camera images to OpenAI this week. What if those images contained someone's face without their consent? Who is responsible?
2. Should AI companies be required to delete your data after a fixed period?
3. The Vosk model ran locally — no data left your Pi. What did you give up in exchange for that privacy? (hint: accuracy)
After this section, take a 10-minute break (10:40 AM to 10:50 AM).