“The claims made in the recently published video are not true.” That’s LG Electronics, in a statement to Tom’s Hardware on 9 September 2026, pushing back on an investigation from Gamers Nexus, Level1Techs, and independent security researchers who said LG smart TVs log and upload user data, and record audio while sitting in standby.
My first reaction, as someone who spends most of his week wiring up bots that talk to APIs and sniff around local networks on purpose, was not outrage. It was recognition. Because the part LG did not deny is the part I would have flagged first in a code review: the TVs scan local networks. LG confirmed that. The denial covers the constant logging, the uploading, the standby audio recording. The network scanning is just… a feature.
The confirmed thing is more interesting than the denied thing
Every builder reading this has written something that enumerates devices on a LAN. It’s how you find a Chromecast, a printer, a Hue bridge, a Sonos speaker. mDNS, SSDP, an ARP sweep, maybe a port probe to confirm what you found. Totally normal engineering. I have shipped bots that do exactly this and I did not think twice about it.
The difference is scope and consent. When I write a discovery routine for a home automation bot, that bot lives on one network, owned by one person who installed it deliberately. When a TV manufacturer ships the same routine, the online investigation puts the count at 216,000,000 TVs. That’s the same code with a wildly different blast radius, and the ethical weight scales with the fleet size, not with the elegance of the implementation.
So the question I’d ask in a design meeting is not “does it scan?” It’s these:
- What leaves the device after the scan finishes?
- Who sees the result, and for how long is it retained?
- Can the user say no without losing the thing they paid for?
- Is the scan disclosed in language a normal person would actually read?
LG’s statement addresses intent. It doesn’t answer the four questions above in a way an engineer could verify. That gap is where all the distrust lives.
Opt-in is doing a lot of work in that sentence
LG also says voice data collection is opt-in and requires activation. Fine. But “opt-in” in consumer hardware and “opt-in” in a well-designed bot are different animals. In a bot I ship, opt-in means a config flag that defaults to off, a log line when it flips on, and a hard failure if the credential is missing. In a TV setup flow, opt-in can mean a screen with a big blue Continue button and a smaller Learn More link, presented while you’re standing there holding a remote and just wanting to watch something.
Both are technically opt-in. Only one is honest about it. If you build anything that touches user data, that distinction is the whole job.
The rooting detour, and why I’m not doing it
One of the follow-on discussions on Hacker News, with 133 comments, went after why rooting your TV is a bad idea. I’ll take that as read. The instinct is understandable, you want to see what the box is doing, so you crack it open. But a rooted TV is a device with an unclear update path and unclear security posture, sitting on the same subnet as everything else you own. You’ve traded an opaque appliance for a fragile one.
The better move for builders is observation without modification. Put the TV on its own VLAN or guest network. Point DNS at something you control and read the query log. Watch egress at the router. You learn what the device talks to, how often, and how much, and you keep the warranty and the security updates. That’s the same approach I use when reverse-engineering an undocumented API for a bot, watch the traffic, don’t patch the binary.
What I’m taking back to my own code
The uncomfortable lesson here isn’t about LG. It’s that the techniques in this story are the techniques in my toolbox. Device discovery, telemetry, always-on listeners, background sync. None of it is inherently hostile. All of it becomes hostile when the person on the other end doesn’t know it’s happening.
So a few habits I’m tightening up. Default every data-collecting path to off. Log what the bot collects in a place the user can read. Make the disable switch a real switch, not a preference that quietly resets on update. And write the disclosure before writing the feature, because if the disclosure sounds bad, the feature probably is.
LG says the concerns aren’t true. Maybe. The 359-comment thread suggests plenty of people want more than a statement, they want something they can measure. That’s a reasonable ask, and it’s the standard I’d want applied to anything I ship too.
🕒 Published: