LG says its smart TVs do not record your conversations. LG also confirms its smart TVs scan your local network. Both of those things are in the same statement, and the gap between them is where this whole story lives.
On 9 September 2026, LG Electronics told Tom’s Hardware that claims from a recently published video were “not true.” The video in question came out of work by GamersNexus, Level1Techs, and independent security researchers, who described LG TVs logging and uploading user data, with allegations of audio recording during standby. Tom’s Hardware framed the investigation’s reach at 216,000,000 TVs. LG’s response denied the constant logging and the standby audio recording, said voice data collection is opt-in and requires activation, and confirmed the network scanning behavior.
That’s the full set of facts I’m working with. No court filings, no independent replication of the findings yet, no regulator statement. So let me be honest about what this piece is: not a verdict, but a review of how the denial is built.
Denials are a product, and this one has a spec sheet
I spend most of my time evaluating AI tools, and one habit transfers directly here. When a vendor publishes a claim, I read for what the claim does not cover. Marketing copy and corporate denials are written by the same instinct: say the true thing that sounds like the bigger thing.
Look at the structure. “Not true” is aimed at the video’s most dramatic allegation, the standby audio recording. That’s the accusation most likely to make someone unplug their TV, and it’s the one LG addresses most forcefully. Voice data being opt-in is a real fact about a real setting, and it’s a fair thing to point out. But “opt-in” describes a mechanism, not an outcome. It tells you consent exists somewhere in the flow. It does not tell you how many people ended up consenting, what the prompt looked like, or whether the default state survives a firmware update.
Then there’s the confirmation. LG did not have to volunteer that its TVs scan local networks, and the fact that it did suggests the company considers that behavior defensible. It probably is defensible in engineering terms. Device discovery, casting, connected speakers, second-screen features — a TV that wants to talk to your phone has to find your phone. The scan has an obvious purpose.
But a scan with an obvious purpose still produces an inventory of every device on your home network. What happens to that inventory is a separate question from why it was collected, and the denial does not touch it.
Why this pattern matters to anyone reviewing AI tools
The parallel to the AI tool space is close enough to be uncomfortable. Every assistant, note-taker, and browser extension I test makes claims in this exact shape. “We don’t train on your data” often means “we don’t train on your data by default.” “Processing happens locally” often means the inference is local and the telemetry is not. “Opt-in” often means a checkbox on a screen most users clicked through in four seconds during setup.
None of that makes the vendor a liar. It makes the vendor a company writing carefully. The useful skill is separating the denial from the disclosure and asking what sits in between:
- What specific behavior was denied, in the vendor’s own words, and how narrow is that wording?
- What behavior was confirmed, and did the vendor volunteer it or get cornered into it?
- Which claims are about mechanisms, like consent flows and settings, and which are about outcomes, like what data actually left the device?
- Who can check any of this independently, and has anyone done it yet?
On that last point, the LG situation is still open. Researchers made claims. LG rejected some of them and confirmed one. Nobody outside those two camps has published a resolution I can point to.
What I’d actually do about it
Nothing dramatic. Network scanning by a TV is not the same category of problem as a microphone in standby mode, and treating them identically wastes the alarm you might need later.
Go look at your TV’s privacy settings and see what’s on. If voice features are enabled and you never use them, turn them off — that’s a thirty-second action with no downside. If you care about the network scan, a separate VLAN for media devices handles it, though I’ll admit that’s a project, not a tip.
The broader habit is the point. Read the denial. Note what it declines to mention. That skill costs you nothing and applies to every AI tool that lands in your browser this year, including the ones I’ll be reviewing here.
đź•’ Published: