TVOS players and streamers are looking to AI to help power new content search and discovery experiences, including more two-way conversational interactions with apps and TV screens.
Although technology is quickly evolving, it may take some time for new experiences to be fully envisioned and rolled out. However, some in the industry – namely Fubo CEO David Gandler - thinks owning the so-called conversation layer could mean owning the TV in the future (more on that farther down).
But first, announced last week, Samsung is the second to tap metadata from Nielsen-owned Gracenote to help ground and validate a range of AI-powered initiatives and entertainment use cases on the company’s smart TVs worldwide.
With a separate but similar multi-year strategic partnership extended with Google last month, the latest shows traction for the use of the vendor’s metadata on this front and provides another signal of where streaming platforms are headed as they seek to develop and introduce more conversational LLM-based and ChatGPT-like content discovery and other experiences on the TV screen.
Gracenote’s structured and human-verified metadata across TV, movies and sports content can be used to help ground responses generated by AI chatbots in factual and up-to-date information, which helps validate LLM-based, AI-powered advanced search and discovery capabilities to ensure accuracy and enable more reliable conversational interactions by users.
In addition, the new agreement allows Samsung to use Gracenote to develop new offerings and realize AI-driven operational efficiencies, such as data ingestion and harmonization that happens on the backend.
Earlier moves by others have shown how Large Language Models (LLMs), coupled with enhanced voice capabilities and AI chat-like experiences are being developed and integrated by some streaming platforms – such as Roku with AI-powered voice content queries and responses, Google TV introducing Gemini on smart TVs including TCL, and Amazon incorporating Alex+.
Google TV’s Shalini Govil-Pai previously told StreamTV Insider about Google TV bets including AI that enables natural conversations for users to have with their TV about content as they decide what to watch.
It also sees applications for AI-powered voice queries to answer user questions around use cases like education, travel and shopping – with genAI providing voice-based responses back from the TV as well as text, while also surfacing related videos. For more on Google TV’s bets read here.
Mitigating AI hallucinations, knowledge lock
But as players pursue new AI-enabled experiences, Gracenote SVP of Product Tyler Bell, in an interview with StreamTV Insider described how there are two known limitations with LLM-based AI-generated responses when it comes to content discovery and other CTV applications.
One is that they hallucinate or provide responses that might sound correct but are actually made up as they’re probabilistic in nature – meaning the output may not be accurate or grounded in verified, factual information.
The other is that AI large language models are knowledge locked, meaning they can’t see information outside of their training date, as they’re not continuously updated. So the information can be static, in part because model training is expensive and time consuming. As such, they’re often only created or trained once or twice a year – without access to new information after being built.
A key aspect of utilizing Gracenote’s metadata is grounding AI models in the entertainment realm with trusted, real-world entertainment information to enable accurate content discovery and user experiences through these new mechanisms.
And here we’ll just take a minute to note a distinction between AI chatbots (or the consumer-facing mechanism or client - a la a ChatGPT) and the underlying LLM.
As Bell explained, chatbots or clients like ChatGPT or Gemini are grounded in that they use both their LLMs and then a lot of additional data sources to provide strong results.
But if one is a vMVPD or other type of streamer, they can’t use the Gemini client and need to rely on an underlying model, such as Gemini 3 Pro, for example.
“And if you’re using this model, there’s always a gap between what the model knows and the rest of the world,” Bell described.
For example, Gemini 3 Pro was built in November 2025 and as such doesn’t natively have visibility into entertainment data such as content releases, sports scores, or awards like the Emmys or BAFTA’s that have happened since. So in an entertainment context, real-world data like that from Gracenote needs to be incorporated into the model context to ground it and provide accurate results.
“Grounding really helps all of these companies fundamentally ensure that the model that they have, and it's strong inference capabilities, can be combined with real world, authoritative, factual data that's also up to date,” Bell explained. “And therefore it basically mitigates that grounding process of bringing this data in, mitigates the deficiencies with the probabilistic technology that AI introduces.”
While Samsung and Google aren’t training LLMs on Gracenote’s content metadata, it provides a source of reference and truth to check against and abate those limitations of hallucinations and knowledge lock.
“Samsung is committed to delivering the most useful and engaging entertainment experiences to our users,” said Bongjun Ko, Corporate EVP at Samsung Electronics, in a statement. “By combining our AI technology with Gracenote’s industry-leading metadata, we aim to push content search and discovery to new heights, delighting viewers by empowering them to find the entertainment they love intuitively and naturally.”
The vendor previously debuted its Gracenote Video Model Context Protocol (MCP) Server, which is one mechanisms platforms can opt to connect data through or they can choose to use their own technology and tech stacks to integrate the Gracenote metadata with AI models – the latter the route that Samsung’s implementing.
Three uses for Gracenote metadata combined with AI
While unable to comment specifically about the types of AI-powered experiences Samsung will debut, Bell cited three main buckets where the vendor sees customers using or interested in using Gracenote metadata combined with AI.
Up first is search, which is also enabled by voice and already in the works as seen by previous announcements. This is “less about voice as navigation” and more about providing ancillary information, per Bell.
Here again, this could be seen with Google’s vision for TV, as well as efforts by Roku where users can ask follow-up questions to content queries. Or in an example scenario shared by Bell, querying a list of horror movies around Halloween, but then diving deeper and being able to ask conversationally for ones that aren’t so scary that they’d freak out a 12-year-old.
This type of use also has the benefit for streamers in that consumers don’t need to go on their phone to check Google but can stay within the app itself.
Second: Recommendations based on LLMs and behavioral data, like watch history, or demographic data like age. Bell noted AI can be used to assist editors or recommendations systems in creating “highly customized and personalized rails that exist on the home screen before the user logs in.”
The Samsung announcement also noted how the smart TV-maker can offer curated carousels and recommendations of programs to viewers in a lean-back mode.
A standardized and structured metadata taxonomy working with AI also provides value for infrastructure behind the scenes by providing a more common language.
To that point, the third way customers are using or looking to use Gracenote metadata on the data ingestion and harmonization side with AI to normalize and match information and content descriptions so that it’s standardized across a wide variety of content.
Slowly at first, then all at once
Content discovery has been a persistent struggle for consumers amid a vast array of content and not everyone feels platforms are incentivized to surface the content users actually want as commercial interests supersede (as is part of the underlying thesis for The Trade Desk’s Ventura effort).
Bell believes metadata can solve for content discovery but acknowledged “very often it remains a zero-sum game” as TVOS platforms have their own channels, which compete with apps, which are also partners.
Still, he thinks it’s possible for both consumer and commercial interests to co-exist when there’s a more neutral business model at play.
“You can have better consumer experiences and also satisfy your commercial needs, usually through better personalization,” said Bell (whose previous role included head of product for the Ventura OS).
As for continued implementation of LLM-based and AI-powered CTV experiences, Gracenote expects a lot more.
“Our bet is that all players within the CTV space will be using LLMs as the primary mechanism for them and their consumers to interface with media metadata,” Bell said. “We think this is going to happen very slowly at first and then all at once.”
One reason for that is because much of the industry is still figuring this out, and he noted most are taking an augmentative approach in terms of seeing what AI can do now, but recognizing limitations and not completely getting rid of long-running systems.
“What they are doing is they’re looking to see how they can use AI in combination with Gracenote data to augment and make better what they already do well,” he said.
Owning the conversation layer?
As noted, AI-powered two-way conversations or ChatGPT-like interactions with TVs or streaming apps is something various players are exploring - and streamers beyond TVOS providers are also paying attention to.
Last month the The Paley Center for Media held an event in New York City featuring a conversation between Optimum CEO Dennis Matthew and Fubo CEO David Gandler, where the executives emphasized the need for a simpler and more consumer-centric approach to streaming, and where the topic of AI inevitably came up.
During the event, attended by StreamTV Insider, Gandler zeroed in on a need for platforms to embrace technology and shift from a reactive approach to user experiences and content discovery (where you watch one show that’s a mystery and the algorithm then surfaces a bunch of titles in the similar genre that may or may not be of interest) to a more predictive and anticipatory experience.
Not all of that is necessarily two-way conversation based, where he noted inputs on time spent and data accumulated can help figure out ways to anticipate what someone will watch – although acknowledged it sounds a little bit “hard to get to” but believes it’s where the industry needs to be.
However, down the line, he sees potential for more significant changes around the user interface and experience – assuming business models don’t impede on technological advances.
“My sense is that the UI we all see today is going to change dramatically. The question is, how dramatic is that going to be?” Gandler posited, contending there are rules that “preclude us from really innovating.”
Pointing to conversational back-and-forth with genAI platforms like ChatGPT or Gemini, he questioned whether one could envision “a world where there is no UI” at all.
In that world, he used an example of a child that doesn’t like to watch TV with their parent, then relaying that fact to the AI-based TV chat that is already well familiar with the user and asking, “What are the things we could watch together?” – where it starts to be a conversation.
“That’s where we’re talking about the anticipation problem,” he added.
And going a little farther, AI moves beyond the question of who holds a billing relationship for subscriptions, where for Gandler, “this now becomes the conversation layer” and who owns that.
The Fubo executive acknowledged he doesn’t know the answer but believes the position will be key over the next decade.
“The company or the group that’s going to figure out or own that layer, those are the groups that are going to really own TV like ten years from now,” Gandler predicted.
Matthew, meanwhile, noted that in his own life, most of the time what he watches at home is “massively predictable” given the day and time – be it watching something with his family on a weeknight, tuning into an NBA game for his favorite team during the season, or playing Chopped or a show like The Office in the background as he goes through emails on a laptop.
“I imagine AI is going to completely transform this experience,” he said.
But like Gandler, believes stakeholders have to be willing to work together “to unlock all the mechanisms to deliver that to the customer,” noting it won’t just show up.
“We've got to work together to be able to deliver this experience that's going to deliver tremendous value for the customer, which I think customers would be happy to pay for, and we can monetize in a way that's much different than what it is today,” Matthew said.