Your TV Scans Your Network and Nobody Wrote That Down Anywhere

LG’s denial and the researchers’ findings can both be true at once, and that gap is the actual story.

On 9 September 2026, LG Electronics told Tom’s Hardware that claims made in a recently published video about its smart TVs were “not true.” The company denied that its televisions constantly log and upload user data, denied that they record audio while in standby, and said voice data collection is opt-in and requires activation. In the same breath, LG confirmed that its TVs scan local networks. Gamers Nexus, Level1Techs, and independent security researchers had documented behavior they characterized as spying, with an online investigation claiming 216,000,000 TVs are affected. Hacker News threads on the topic ran to 359 and 133 comments.

I work on agent architecture, and this dispute reads to me less like a scandal and more like a preview. Not because I think LG is lying, but because the disagreement has the exact shape of every disagreement we are going to have about autonomous software in the home over the next decade.

Two Sides Describing Different Layers

When a company says “we do not record your conversations,” it is making a claim about a specific data pipeline: a microphone stream, an upload endpoint, a retention policy. When a researcher says “this device is spying,” they are making a claim about observed behavior at the network boundary: packets went somewhere, on a schedule nobody asked for, to a destination nobody documented.

Both statements can survive scrutiny simultaneously. The vendor answers a question about intent and configuration. The researcher answers a question about effect. Nothing in either answer forces the other to be false, which is why these arguments never resolve. They terminate in mutual exhaustion instead.

The confirmed network scan is the piece worth sitting with. LG did not dispute it. A television that enumerates devices on your LAN is building a model of your home: how many phones, which brand of thermostat, whether there is a work laptop present in the evenings. That is not audio. It is arguably more useful than audio, because it is structured, small, and cheap to transmit. You can infer household size, work-from-home patterns, and hardware upgrade cycles from a device list. No transcription required.

Why This Is an Agent Problem

The category of software I study is defined by acting without a per-action prompt. An agent perceives, decides, and executes on its own clock. Consumer appliances have been quietly converging on that description for years, and a TV that scans your network to discover casting targets is a small agent doing perception work.

The trouble is that our vocabulary for consent was built for a request-response world. Checkboxes and opt-in toggles work when there is a discrete feature with a discrete data flow. They fall apart when a system performs continuous background perception whose outputs feed multiple downstream uses, some of which did not exist when the user clicked accept.

Consider what an honest disclosure would even look like for network scanning:

• What triggers a scan, and how often
• What fields are retained versus discarded
• Whether device inventories leave the TV, and to which endpoints
• How long the resulting model persists, and whether it survives a factory reset
• Which downstream systems, advertising included, can query it

Nobody ships that. Not LG, not its competitors, not most of the AI products I evaluate professionally. The disclosure surface for autonomous perception is essentially blank across the industry, which means every conflict gets litigated after the fact by people with packet captures.

The Rooting Trap

One of the side threads in this story was about rooting LG TVs, and why that might be a bad idea. It points at something structural. Users who want to know what their devices do have exactly two options: trust the manufacturer’s statement, or void the warranty and break the security model to find out themselves. There is no supported path for auditing your own hardware.

That is an architectural failure, not a policy one. Devices that make autonomous decisions should expose an inspectable log of those decisions as a first-class feature. Every scan, every upload, every destination, queryable by the owner. It is not technically hard. Observability tooling for exactly this problem is mature on the server side, where we would consider it negligent to run a system without it.

What I Take From This

I have no reason to doubt LG’s engineers about what their microphone code does. I also have no way to verify it, and neither do you, and that is the condition we are all in with every autonomous device in every room.

The lesson for anyone building agents is unglamorous. Design your perception layer so that a hostile third party with a packet capture reaches the same conclusion you would state in a press release. If those two accounts can diverge, they eventually will, and you will spend your credibility on the difference.

🕒 Published: September 13, 2026