Executive Summary

Key Numbers

The first two are what users changed by handing the job to AI. The last two are numbers being fought over in US courts right now.

Sources: TechCrunch (October 19, 2026), Texas Attorney General announcements (February 26 and May 11, 2026), class action complaint in the Northern District of California (October 6, 2026), ACR measurement study at the ACM Internet Measurement Conference (November 2024).

Smart TV ACR is short for automatic content recognition. What it does fits in one sentence. The TV works out for itself what is showing on it. The principle is the same as Shazam, which names a song you play to it. Features of the sound or the picture are compressed into a short value to make a fingerprint, and that fingerprint is matched against a reference list held on a server to find the title.

There are two approaches. One lifts the fingerprint from the content itself. In the other, whoever transmits the content plants a marker invisible to the eye every few seconds and the TV reads it. The first approach does not care about format or codec, so video that arrived by any route can be measured with the same ruler.

The difference that matters is which layer the feature runs in. It is not a permission that each app is granted but a function of the TV itself, so whatever you put on the screen can become its subject. Live broadcast, certainly, but also the picture from a games console or a set-top box connected over HDMI falls inside the identification range. This is not a structure you can get out of by deleting one app.

That this is not guesswork has been confirmed by measurement. Researchers at University College London, UC Davis and Universidad Carlos III de Madrid put new Samsung and LG sets on a controlled network, recorded every packet leaving them, and presented the results at the ACM Internet Measurement Conference in November 2024. ACR traffic continued even when the TV was used only as an external HDMI monitor. The TV was reading the screen of a laptop or a games console plugged into it and sending that out. The upload cadence was recorded as well. LG reached its own ACR servers every 15 seconds, Samsung every minute. The rate at which the screen is grabbed is much shorter than that. An LG document the paper cites puts it at 10 milliseconds, and Samsung's advertiser guide at a frame every 500 milliseconds. What it takes is buffered and sent in batches.

Why the feature is in the TV at all becomes clear once you ask where the money comes from. Vizio, an American TV maker, earned more profit in 2021 from selling user data than from selling TVs. A TV sells once, but the viewing record from the living room arrives every month. How big screens got cheap and how this feature came to ship by default are the front and back of the same story.

On October 19, TechCrunch gathered the people making their living-room TVs dumb again. What they share is the method. They hand the TV's model number to an AI agent and tell it in plain words to strip out the ads and the preloaded apps. The agent then works out which packages that model ships with and walks them step by step through putting a computer and the TV on the same network and sending commands.

The cases in the report run like this. Mihir Gupta spent an afternoon switching off 25 preloaded apps, then replaced the 192MB stock launcher with an 83MB custom one and set the wallpaper to rotate. Ryan Doyle asked Claude Code to hack his own TV, cleared out the ads and the bloatware, and reworked the screensaver while he was at it. Gathin T. disabled 25 Xiaomi packages on his parents' TV and put on a launcher that resembles Apple TV, publishing the whole process under the name cleanyourtv.

What set Gathin T. off was small. It bothered him, he wrote, that his parents' TV was showing a sponsored ad before you could even get to Netflix. He put the prompt on GitHub and posted the before and after, and over the next two days, he says, a few thousand people did the same to their own TVs. The sets they followed along on were Xiaomi, TCL, Sony, Fire TV, LG and a projector or two. That both of them switched off exactly 25 apps is a coincidence. They were working on different TVs.

dumbtv, built by Muzzammil Yasin Sabuwala, reaches wider. It takes out ads, tracking and updates that arrive unasked, all in one pass, and it was tested end to end on a Fire TV set and on the Android and Google TV platforms. Sony, TCL, Philips, Hisense, Xiaomi and the Nvidia Shield are on that list. The work mixed Deepseek V4.1 Flash with Claude Opus 5.5 and Claude Sonnet 5.5. The goal posted on the repository is short. A TV should do what you paid for and nothing else: your apps and your HDMI inputs, with no ads from the TV itself, no watch tracking and no update you did not ask for. That watch tracking sits on that list of targets is what catches later.

The same work was possible before. You just had to dig up package names, attach a shell over the Android Debug Bridge, and find and install a launcher yourself. Reading through community threads took its own kind of patience. What has changed is that the side holding that knowledge has moved from people to agents. What the user has to hold is the model number and enough will to follow it through.

While users were working on their own TVs, the regulators were looking at the same feature. In December 2025, Texas Attorney General Ken Paxton sued five companies, Samsung Electronics, LG Electronics, Sony, TCL and Hisense, saying they collected viewing data through ACR without disclosing it properly. Samsung reached a settlement on February 26, 2026, and the suit against it was dropped. Under the terms, it will not collect or process ACR data without explicit consent, and will use a software update to make the notice and consent screens conspicuous. On May 11, 2026, LG Electronics USA settled on similar terms. It agreed to obtain consent before collecting viewing data through ACR, and to run pop-ups on the TV and on its website setting out what is collected and how to refuse. The settlement records that this is not an admission of liability. The cases against Sony, TCL and Hisense are still going.

Into the space the settlements left, another suit arrived this October. On October 6, three California residents filed a class action against Samsung's US arm in the Northern District of California. The 88-page complaint alleges the TV captured and analyzed the screen twice a second and attached the result to IP addresses and device identifiers for targeted advertising. It also alleges the analysis covered not only broadcast and streaming but the screens of external devices such as games consoles. Samsung maintains that ACR is an optional feature that works only with consent and can be turned off at any time, and that it is not a technology that records or looks at the screen. Still, twice a second is not a figure the complaint produced on its own. Samsung's advertiser guide, cited by the researchers above, puts the capture interval at 500 milliseconds. The value explained to advertisers as a specification and the value now in dispute are the same. Capturing at that rate does not mean the captures are kept. The manufacturers' account is that the frame is compressed into a fingerprint and the original is discarded.

Four months after settling, LG Electronics stood in a similar place again in September. When the YouTube channel Gamers Nexus raised the possibility that TVs were surveilling viewers, the company denied it on the 9th and issued a detailed statement on the 12th. The thrust of the statement is that its ACR lifts a fingerprint of the sound using an audio processor inside the TV, and is not a method that captures the screen or records and stores video or audio. Voice recognition, it said, works only after a wake word or a button on the remote, and ACR and interest-based advertising are optional features that turn on only with consent. That viewing information may be shared with its advertising subsidiary is written in the same document.

The layer the reported work touches and the layer ACR lives in are not the same. Switch off the preloaded apps and change the launcher, and the ads and recommendation tiles that appeared the moment the TV came on are gone. The feature that identifies what you are watching, by contrast, is tied to a consent item inside the settings, and withdrawing that item is the proper route. Turning off Live Plus on an LG TV, or withdrawing consent to the viewing information terms, is that kind of action. Expecting the two jobs to produce the same result goes wrong.

The complaint that the items are not gathered in one place is an old one. You turn off the viewing information terms, and the same kind of collection carries on under the advertising identifier or a policy with another name. You have to open settings in more than one place, and the path changes with the model and the firmware. Refusal is designed as several errands rather than one choice.

Does the errand pay? The researchers measured that too. Once every advertising and tracking item was switched off, traffic to the ACR servers disappeared completely, and no replacement address came up in its place. The refusal is not a feature in name only. Whether the TV account was logged in, on the other hand, made no difference to the traffic, because the grouping is keyed to IP address and device rather than to an account. Logging out is not an escape route. The door out is in the consent items.

Leaving the work to an agent has a price of its own. TechCrunch also reported a case where Claude refused a request that read as hacking, a case where an Android TV was bricked because a necessary package went off with the rest, and a case where a Fire TV halted partway through. It remains a modification outside the manufacturer's warranty. The report set out the route that does not use AI as well: changing only the screen, with launchers that already exist such as ATV Launcher, Projectivy Launcher and FLauncher.

Put together, what AI is doing for people right now is closer to making the TV quiet. To stop the collection itself you have to go into the settings once more afterwards, and that job too can be walked through if you hand over the model number and ask. Being walked through it and confirming that it actually went off are different things.

The collecting side runs automatically and the refusing side is left to errands. That asymmetry is what consent looks like in practice. Reading the terms to the end, going into more than one settings menu to withdraw, and checking after an update that the withdrawal held, are things only people with time and knowledge do. Consent is written as everyone's right, and in practice it has been the portion of those who know.

When AI agents start taking on those errands, the balance moves a little. Work that belonged to people who know the technology comes down to people who know how to ask. When refusal gets easier, you also start asking what the consent rate has been measuring all this time. Whether so many people consented because they agreed, or because they did not know how to turn it off, is starting to come apart.

From where data work sits, one more thing catches. If the people stripping out the collection cluster among the technically fluent, that group drops out first from the viewing panels ad measurement leans on. The problem is that the gap they leave is not random. The measurements look fine while the population underneath tilts, and the tilt is invisible from the aggregate numbers alone. Data that never recorded where its missingness came from always behaves this way.

Three things to check in your own living room.

• • The defaults. Whether the consent screen at setup went through with everything already on, or whether you picked the items one at a time. If you cannot remember, it is safer to assume they are on
• • How many items there are. Whether collection of a similar kind is split across several names, such as viewing information, interest-based advertising and voice information. If you switched one off and stopped there, the rest may still be running
• • The state after an update. Whether the same item is still off when you open it again after a firmware update. If a withdrawal does not hold, one withdrawal is only a record

Thank you for reading this far. The facts here were checked against TechCrunch's reporting, the Texas Attorney General's announcements, reporting on the California class action complaint and LG Electronics' statement, and how ACR works comes from technical descriptions of the method. If you have been through your own TV settings, or have a record of which item was hidden where, we would be glad to hear it.

Pebblous Data Communication Team
October 20, 2026

Academic Paper

• 1.Anselmi, G., Vekaria, Y., D'Souza, A., Callejo, P., Mandalari, A. M., Shafiq, Z. (2024). "Watching TV with the Second-Party: A First Look at Automatic Content Recognition Tracking in Smart TVs." ACM Internet Measurement Conference (IMC '24), pp.622–634. DOI: 10.1145/3646547.3689013.

Official Announcements & Court Filings

• 2.Office of the Texas Attorney General. (2026). "Attorney General Paxton Secures Major Agreement with Samsung to Ensure that Texans are Protected from Smart TVs Collecting Their Data Without Their Knowledge." February 26, 2026.
• 3.Office of the Texas Attorney General. (2026). "Attorney General Ken Paxton Secures Major Agreement With LG to Protect Texans' Privacy and Stop Data from Being Collected Without Their Knowledge." May 11, 2026.
• 4.Courthouse News Service. (2026). "Samsung faces class action over privacy concerns with smart TVs." October 7, 2026.
• 5.LG Electronics. (2026). "Statement: Understanding Privacy on LG Smart TVs." LG Newsroom, September 12, 2026.

Industry & News Coverage

• 6.Perez, S. (2026). "People are using AI to make their smart TVs dumb again." TechCrunch, October 19, 2026.
• 7.The Korea Herald. (2026). "LG Electronics' US unit settles Texas smart TV privacy lawsuit." May 12, 2026.
• 8.Wikipedia contributors. "Automatic content recognition." Wikipedia.