TikTok has pulled back an experimental AI video summaries feature in the US after it produced surreal results, including one that described dancer Charli D’Amelio as a ”collection of various blueberries with different toppings.” The company says the feature will now only suggest products similar to what appears in a clip, a neat way of saying the robot was not ready for prime time.
The overviews had only reached a limited group of users in the US and the Philippines, where they appeared beneath videos and tried to explain what was happening or add extra context. In practice, the summaries ranged from fine to baffling, which is exactly the sort of quality control you do not want on a platform built on quick, visual shorthand.
How TikTok’s AI video summaries worked
The feature behaved a bit like the AI Overviews now sitting above many Google search results. When users tapped to see more of a video’s caption, TikTok would generate a short written summary of the clip. Some were accurate, but others drifted into the kind of nonsense that turns a launch demo into a screenshot collection.
That matters because short-form video depends on trust in the first glance. If the platform cannot reliably describe what is on screen, it is hard to imagine it confidently selling the same system as useful context rather than flashy autocomplete with a camera problem.
Why TikTok dialled back the rollout
Business Insider first reported the changes, and the odd summaries quickly spread after users shared screenshots online. TikTok has now adjusted the experiment, narrowing what the AI is allowed to say. That is a pretty standard tech-company retreat: launch first, wince later, then quietly reframe the feature as something smaller and safer.
The caution is familiar. Meta, Google, and OpenAI have all had to tighten or reposition generative AI features after early misfires, because hallucinations are tolerable in a chatbot window and a lot less charming when they are attached to real people and real content.
What users are likely to see next
Expect TikTok to keep testing AI-assisted features, but with far stricter guardrails and far less ambition in the copy. The smarter play is probably product suggestions and light metadata, not free-form video interpretation. Anything broader risks another round of viral screenshots and another reminder that ”experimental” often means ”please don’t quote us yet.”