I have commented on cues that can make AI-generated content easier to recognize, including video presentation, varying fake quality, watermarks, and writing style. These are observations from brief posts, not controlled evaluations, and they do not establish how reliably any cue identifies AI content.
On May 20, 2025, I said the only believable AI UGC videos I had seen at that time were green-screen videos where the avatar was small on screen. I had yet to see a believable AI UGC video outside that style, apart from deepfakes. I did not explain what makes a video believable or limit the observation to a particular platform or content type.
On August 24, 2025, an unidentified account replied to @PenguinSensei__ that fake items vary in quality: some look low quality, while others may not be recognized as fake. The post does not specify what items it describes or what distinguishes the examples.
On November 12, 2025, replying to @JamusMcKenna and mentioning @FTC, I wrote, “True!” and said there was a little watermark that says it is AI. The captured wording includes “the do,” and does not clarify what “they” refers to, where the watermark appears, or what it marks.
On November 14, 2025, I said dashes were the easiest way to spot AI copy and called it a “Huge dub” that, according to @sama’s quoted post, ChatGPT follows a custom instruction not to use em-dashes. That method was @sama’s claim, not a test result I reported; I gave no evidence or implementation details of my own.
Related: ai image and video generation, big brands adopting ai ads, doubts about ai s creative takeover