"If the tech industry sought to create a method of seducing users into sending their deepest, most sensitive secrets to a server in a faraway data center, it would be hard-pressed to create a better honeypot than an AI chatbot." That is how WIRED's Andy Greenberg opens a guide, published 22 September 2026, to using AI without giving up personal privacy. The guide ranks the available options from weakest to strongest, and this piece follows that ladder and what it means on a Mac.
Background
Greenberg's baseline for the big three consumer chatbots, ChatGPT, Claude and Gemini, is blunt: expect "approximately zero real privacy" from anyone determined to get conversation records who has a legal path to them, including the service itself, its partners, contractors who help fine-tune it, law enforcement, or a party that subpoenas the records in a civil lawsuit. Matt Green, a privacy- and security-focused computer science professor at Johns Hopkins University, describes the underlying problem: "You have this intelligent thing staring back at you, and you're basically telling it, one question at a time, every possible thing there is to know about your life" [1].
Findings
The strongest cloud option WIRED identifies is a contract: "zero data retention" (ZDR), which OpenAI, Anthropic and Google offer for enterprise versions, generally requiring them to delete interaction records as soon as they are processed. It is not available to ordinary consumer accounts, and even it has exceptions: Anthropic does not offer ZDR for its most advanced models, citing risks such as scams, hacking and "autonomous misbehaviour", and OpenAI says its own ZDR systems will analyse activity for abuse before deletion and can flag it to a customer organisation [1].
Below contracts sit proxies: services such as Duck.ai and Venice.ai relay requests to models like Claude or ChatGPT, limiting what the underlying model learns directly about the user, though the messages themselves still reach that model. Even stripped of identifying metadata, Green says, the content can reveal who someone is: asking about the best coffee shops in a specific neighbourhood reveals where that person lives. "All that stuff tied together is basically a fingerprint for you," he told WIRED [1].
Above proxies, WIRED places services that run inside a trusted execution environment (TEE), hardware that processes a query in isolation and can cryptographically prove the rest of the system cannot access it. WIRED describes Confer, the chatbot launched by Signal creator Moxie Marlinspike, as offering "perhaps the strongest consumer-level privacy protections currently available in an AI chatbot." Meta uses a TEE for a more private mode of Meta AI inside WhatsApp, and Apple's Private Cloud Compute has TEE-like properties for Apple Intelligence requests handled on Apple's own servers [1].
At the top of the ladder are models that never leave the user's computer. Tools such as Ollama, LMStudio and LocalAI guarantee a level of privacy WIRED calls "tough to obtain with any cloud-based AI tool… by the simple fact that no data ever leaves your computer." Apple also handles many simple requests with an on-device model. The trade-off, Green warns, is capability: local models "get stumped and hallucinate far more often" than large cloud models. "You can run these on any laptop," he says. "They're just dumb as rocks." [1]
Implications for Mac users
None of these four levels changes what a Mac user's browser or apps send unless the user also controls the connection itself. A locally run model keeps a conversation on the machine, but a Mac still runs cloud chatbots for most everyday use, and those send data by design, however carefully the provider says it protects that data in transit. The privacy of an AI conversation and the visibility of a Mac's network connections are two separate questions, and WIRED's guide addresses only the first one [1].
Recommendations
- Turn on "zero data retention" or an equivalent business-tier privacy setting when a chatbot you use for work offers it.
- Avoid discussing anything you would not want subpoenaed, sold to a data partner, or reviewed by a contractor, in a consumer chatbot without such a setting.
- For sensitive personal use, consider a service built around a trusted execution environment, or a model that runs entirely on your own device.
- Check which AI apps on your Mac can reach the internet, and remove the ones you no longer use.
Relevance to FireAI
FireAI made its own choice at the top of WIRED's ladder. Its AI runs entirely on the Mac and never sends what it analyses to any server, including FireAI's own. That trade-off works because the job is narrow rather than open-ended: looking at a connection an app is making and explaining, in plain language, what it probably is. See the on-device AI model.
For the cloud AI apps a Mac user does run, FireAI adds the layer WIRED's guide does not cover: control over the connections themselves. The world map shows where each app connects, per-app rules decide which apps may reach the internet at all, and Paranoid mode blocks known telemetry. None of that makes a cloud chatbot itself more private; it makes sure nothing leaves the Mac that the user did not allow.
Limitations
WIRED's ranking reflects each company's own description of its systems, and does not independently verify that any TEE implementation behaves exactly as its vendor describes, or say how a user could confirm a "zero data retention" promise is being kept in practice. WIRED also notes that Venice.ai staff did not answer questions about its privacy claims beyond an automated response. As Greenberg concludes: "If you want real privacy, you might have to pay for it"; Confer, for one, charges $34.99 a month after a daily allowance of free queries, more than Claude's or ChatGPT's cheapest paid tier [1].
Try FireAI, by HisnLabs free for 17 days.