Key Points
- A new study by Gracenote found ungrounded AI models produced completely incorrect metadata for nearly 20 percent of tested film and TV titles.
- Errors included mixing up similarly named titles, inaccurate cast details and missing information for newer releases.
- Only 53 percent of responses correctly identified lead actors for top U.S. films, underscoring reliability concerns.
When it comes to recommendations on shows and movies to watch, off-the-shelf artificial intelligence (AI) products are leaving a lot to be desired.
Some of the biggest products powered by large language models (LLMs) are providing inaccurate or incomplete information about movies and TV shows when they are not connected to verified entertainment databases, according to a new report released by Nielsen’s metadata business Gracenote.
For its report, Gracenote used AI products — including Claude’s Sonnet 4.0 and Google’s Gemini Pro 3.1 client — to ask questions about more than 2,600 film and TV titles across more than a dozen countries. Those models generated incorrect results across every measured metadata category for 506 titles — nearly one in five of those tested — when it relied solely on its training data rather than verified information from Gracenote.
To conduct the study, Gracenote tested 2,600 popular movie and television titles across Australia, Brazil, Canada, France, Germany, Japan, Mexico, the Netherlands, South Korea, Spain, Sweden, the United Kingdom and the United States. The study evaluated information commonly used by viewers and streaming services, including titles, plot summaries, cast information, genres, release years and runtimes.
Among the errors identified by Gracenote were instances where the AI model confused similarly named titles: In one example, the model correctly identified the title and release year of the 2025 thriller “Heel,” but supplied a description, cast information and genre details from “Heels,” the Starz drama series that aired between 2021 and 2023. In another case, the model merged information from the 2024 horror-thriller “Trucker” with a different film of the same name released in 2008.
The research also found weaknesses involving recently released content: Gracenote said the model was unable to provide information on several newer titles, including the 2026 film “GOAT,” which the company said generated nearly $200 million in worldwide box office revenue before its release on Netflix.
Cast information was another area where accuracy suffered. For the 100 most popular movies in the United States, only 53 percent of the model’s responses identifying the primary actor matched information contained within Gracenote’s database.
“Viewers don’t care where a bad answer comes from. If it’s wrong, they blame the service,” said Tyler Bell, the Senior Vice President of Product at Gracenote. “That’s why grounding matters.”
The research comes as streaming platforms increasingly explore the use of AI datasets and tools to help consumers search for programming and navigate expanding content libraries. Gracenote has long positioned its metadata library as a reliable source for search and discovery portals. The company has also filed at least one lawsuit against an AI developer for purportedly training its LLMs off proprietary data without a license.
Still, Gracenote says it works to a company’s benefit to use its verified and accurate metadata to power AI-based search and discovery systems, especially at a time when streamers are complaining that the various apps and services on the market are making it too hard to find things to watch.
“As the report makes clear, no LLM is hallucination-free in 2026,” Gracenote said, noting that entertainment services increasingly require accurate and up-to-date information to support search, discovery and recommendation features.
The full report is available to download, with registration, by clicking or tapping here.
The company plans to discuss the findings of its study during a panel session at the StreamTV Show in Denver on June 18. Nandita Arora, the Senior Director of Product at Gracenote, will participate in a discussion examining the role of AI, personalization and search technologies in content discovery.
The Desk is an editorial partner of this year’s StreamTV Show, which runs from June 16 to June 19. Readers can save 10 percent off a conference pass by using the code DESK10 at this registration page.
