Local Inference Virtue Signaling
Local Inference Virtue Signaling
Every AI product claims to care about privacy. "Your data stays secure." "Enterprise-grade encryption." "SOC 2 compliant." These are policies. Policies change. Local inference is architecture. Architecture is permanent.
The Screenshot Problem
Desktop agents take screenshots. That is how they see your screen. Each screenshot potentially contains: your email inbox, your banking dashboard, your private messages, your medical records, your passwords visible in a form field.
Sending these screenshots to a cloud API for processing means your most sensitive visual data traverses the internet, gets processed on someone else's hardware, and exists - however briefly - outside your control. Encryption in transit does not help if the server needs to decrypt to process.
Screenshots Should Never Leave
The argument for local inference is strongest for visual data. Text can be anonymized before sending. Screenshots cannot. You cannot redact a screenshot without understanding its contents, and understanding its contents is the job you are sending it to the cloud for.
Local inference models running on Apple Silicon can process screenshots without any network call. The image stays in memory on your machine. The model runs on your GPU. The result is text that describes the screen contents. Only that text - not the image - needs to go anywhere.
When Cloud Is Acceptable
Cloud inference is fine for text that does not contain sensitive information. A coding question. A writing suggestion. A data analysis request. The content determines the privacy requirement, not a blanket policy.
The problem is that desktop agents process everything. They cannot filter sensitive from non-sensitive before sending because the filtering itself requires understanding the content. For agents that see your screen, local inference is not virtue signaling. It is the only architecture that actually protects your data.
Fazm is an open source macOS AI agent. Open source on GitHub.