Every company building your AI assistant is now an ad company
Source: Hacker News
Pre‑orders Open for the Juno Pioneer Edition
Reserve your Juno today!
Friday, 20 February 2026 – by Adam Juhasz
On January 16, OpenAI quietly announced that ChatGPT would begin showing advertisements. By February 9 the ads were live. Eight months earlier, OpenAI spent $6.5 billion to acquire Jony Ive’s hardware startup io. The company is building a pocket‑sized, screen‑less device with built‑in cameras and microphones—“contextually aware”—designed to replace your phone.
“Every single company We can quibble about Apple.* building AI assistants is now funded by advertising. And every one of them is building hardware designed to see and hear everything around you, all day, every day. These two facts are on a collision course, and local on‑device inference is the only way off the track.”*
The point isn’t just OpenAI; it’s a structural issue affecting the entire AI‑assistant ecosystem. Advertising‑funded models combined with ubiquitous, always‑on sensors create a privacy dilemma that can only be mitigated by moving inference onto the device itself.
The always‑on future is inevitable
Before we talk about who’s building it, let’s be clear about what’s being built.
Every mainstream voice assistant today works behind a gate. You say a magic word—“Hey Siri,” “OK Google,” “Alexa”—and only then does the system listen. Everything before the wake word is theoretically discarded.
This was a reasonable design in 2014, but it is a dead end for where AI assistance needs to go.
A real‑life kitchen scenario (6:30 am)
Anonymized from one of our test homes. The real version was messier and included a toddler screaming about Cheerios.
“Are we out of eggs again? I’m thinking frittata tonight but we also need to—oh wait, did the school email about Thursday? I think there’s an early release. Anyway, if we don’t have eggs, I’ll get them from Target and also that dish soap, the blue one.”
Nobody is going to preface that with a wake word. The information is woven into natural speech between two flustered parents getting the family ready to leave the house. The moment you require a trigger, you lose the most valuable interactions—the ones that happen while people are living their lives, not thinking about how to give context to an AI assistant.
You cannot build proactive assistance behind a wake word. The AI has to be present in the room continuously, accumulating context over days, weeks, and months to develop the understanding that makes proactive help possible.
Where the industry is heading
- Not just audio—vision, presence detection, wearables, multi‑room awareness.
- The next generation of AI assistants will hear and see everything.
- Some will be on your face or in your ears all day.
- They will be always on, always sensing, always building a model of your life.
The question is not whether always‑on AI will happen. It’s who controls the data it collects. Right now, the answer is: advertising companies.
Policy Is a Promise. Architecture Is a Guarantee.
Here’s where the industry’s response gets predictable:
“We encrypt the data in transit.”
“We delete it after processing.”
“We anonymize everything.”
“Ads don’t influence the AI’s answers.”
“Read our privacy policy.”
With cloud processing, every user is trusting:
- The company’s current privacy policy
- Every employee with production access
- Every third‑party vendor in the processing pipeline
- Every government that can issue a subpoena or national‑security letter
- Every advertiser partnership that hasn’t been announced yet
- The company’s future privacy policy
OpenAI’s own ad announcement includes this language:
“OpenAI keeps conversations with ChatGPT private from advertisers, and never sells data to advertisers.”
It sounds reassuring, but Google scanned every Gmail for ad targeting for thirteen years before quietly stopping in 2017. Policies change. Architectures don’t.
Why Local Processing Matters
When a device processes data locally, the data physically cannot leave the network:
- No API endpoint to call
- No telemetry pipeline
- No “anonymized usage data” that could still be useful for ad targeting
The inference hardware sits inside the device—or in the user’s home—on their own network.
Your email is sensitive. A continuous audio and visual feed of your home is something else entirely. It captures:
- Arguments
- Breakdowns
- Medical conversations
- Financial discussions
- Intimate moments
- Parenting at its worst
- The completely unguarded version of people that exists only when they believe nobody is watching
We wrote a deep dive on our memory system in Building Memory for an Always‑On AI That Listens to Your Kitchen.
What Amazon Showed Us
- They eliminated local voice processing – Entrepreneur article
- They planned to feed Alexa conversations to advertisers – AlexaEchos.com
- They partnered Ring with a surveillance network that had federal law‑enforcement access
What happens when those same economic incentives are applied to devices that capture everything?
The Edge Inference Stack Is Ready
The counterargument is always the same: “Local models aren’t good enough.”
Three years ago, that was true. It is no longer true.
Why Local Inference Works Today
- Real‑time speech‑to‑text, semantic memory, conversational reasoning, text‑to‑speech, etc., can run on a device the size of a cable‑box.
- No fan noise. One‑time hardware purchase, no per‑query fee, and no data leaves the building.
- New model architectures, better compression, and open‑source inference engines have converged to make this possible.
- The silicon roadmap points to more capability per watt every year.
We’ve been running always‑on prototypes in five homes. The complaints we get are about the AI misunderstanding context, not about raw model capability. That’s a memory‑architecture problem, not a model‑size problem.
Capability vs. Cloud
- Local models are not as capable as the best cloud models, but we rarely need a smart speaker to re‑derive the Planck constant.
Hardware & Business Model
- Inference runs on‑device; audio and video are processed locally and never transmitted.
- The business model should sell hardware and software, not the data the hardware collects.
- Architecture where the device manufacturer cannot access the data because there is no connection to retrieve it.
Privacy‑by‑Design
- The most helpful AI will also be the most intimate technology ever built: it hears everything, sees everything, and knows everything about the family.
- The only architecture that keeps that technology safe is one that is structurally incapable of betraying that knowledge— not policy, not promises, not a privacy setting that can be quietly removed in a March software update.
Call to Action
- Choose local.
- Choose edge.
- Build AI that knows everything but phones home nothing.