One in Five AI-Generated Movie Descriptions Contains Fabricated Facts, Gracenote Study Finds

Gracenote, a company owned by Nielsen that supplies program metadata to streaming services worldwide, has published a report revealing the scale of the AI reliability problem in describing movies and TV shows. A language model without access to a verified database completely fabricated information for nearly one in five tested titles.

How the Study Worked

Gracenote's team put a language model through a series of questions about 2,600 popular movies and TV series. The questions covered basic, objective production attributes: title, plot description, cast, genre, release year and runtime. Each answer was then compared against metadata from Gracenote's own database, considered one of the industry's reference sources for streaming platforms.

The researchers tested the model in two configurations: one without access to external data, relying solely on training knowledge, and one connected to the Gracenote database via an MCP server. The difference in answer quality between the two versions was meant to show how much simply scaling up models solves the hallucination problem, and how much an external, continuously updated source of facts is still needed.

Where the Model Erred Most

The biggest problems involved productions released within the past two years, for which the model often lacked sufficient training data or confused them with earlier titles bearing similar names. The report's authors cite the example of the 2025 thriller Heel, which the model consistently confused with the series Heels, aired from 2021 to 2023, attributing to it the wrong cast and genre.

A similar problem affected the 2024 movie Trucker, whose data the model merged with an identically titled production from 2008. For the 2026 movie GOAT, which grossed around $200 million worldwide, the model was unable to provide even basic information at all. The report's authors stress that the errors did not stem from a lack of data available online, but from the model matching answers based on similarity of words in the title rather than actual knowledge of the production.

Implications for the Streaming Industry

The findings have a direct bearing on how search and recommendation systems work in streaming services, which increasingly rely on generative AI to handle viewers' natural-language queries. An incorrect plot description, a mixed-up cast or a nonexistent genre could lead a viewer to skip a title altogether or, conversely, to feel disappointed after starting a production that doesn't match the description.

Viewers don't care where a bad answer came from. If it's wrong, they blame the service - Tyler Bell, Senior Vice President of Product at Gracenote

Gracenote argues that the solution is not to abandon generative AI in content search, but to combine language models with verified, continuously updated metadata databases. The company presented the study's findings on June 18 at the StreamTV Show industry conference, addressing the message primarily to VOD platforms and smart TV manufacturers rolling out their own AI assistants for content search.

The Broader Hallucination Problem

Gracenote's report adds to a growing body of research showing that even the latest language models remain prone to generating convincing-sounding but false information when they lack access to a verified data source. The report's authors state plainly that no language model currently available is free of hallucinations, which poses a particular risk for systems expected to deliver precise answers at massive scale.

For Polish streaming platforms and smart TV providers, the report's conclusions matter in the context of planned rollouts of AI-based assistants for content search and recommendations. The choice between a model relying solely on training knowledge and one connected to a current metadata database could directly affect viewers' trust in an app's search function.