Comparison of audio description by ChatGPT4o versus ChatGPT5.2

Rintje

Fragment 1: Greeting Grandma

ChatGPT4o

• Does not recognize the mother
• Includes unnecessary details (apples)
• Interpretation instead of description ('cozy kitchen')

ChatGPT5.2

• Very limited audio description, with (too) little context
• Mentions the dropped egg afterwards, requiring the viewer to infer what happened

Rintje

Fragment 2: Boiling Eggs

ChatGPT4o

• Hallucinations (Henriette, a toy, conversation with a character)
• Misses the moment where Rintje falls asleep
• Repeats nearly the same description three times

ChatGPT5.2

• Very limited audio description, with (too) little context
• Misinterpretation ('tense' instead of falling asleep)
• Descriptions of actions are more accurate

Rintje

Fragment 3: Rintje pulls a prank

ChatGPT4o

• Misinterpretation (describes boiling eggs instead of transferring them to an egg cup)
• Mentions that the egg is secretly eaten
• Misses the detail that the empty shell is placed upside down in the egg cup

ChatGPT5.2

• Very limited audio description, with (too) little context
• Mentions the eaten egg only as a result, rather than describing the action itself, even though the action is more important (this may cause confusion and misunderstanding)
• Also misses the detail that the empty shell is placed upside down in the egg cup

The full episode of Rintje by ChatGPT4o

Weaknesses

• A great deal of repetition ('dogs boiling eggs in a kitchen' repeated many times)
• Correctly identifies the main character Rintje, but not always the other two characters
  (sometimes simply 'dog wearing an apron')
• Often describes the overall setting or scene rather than what is actually happening at that moment
• This creates confusion about the mise-en-scène (everything visible within the scene) and the events themselves
• Sometimes includes misinterpretations or hallucinations (a character who is not there, a conversation that is not actually taking place)

Strengths

• Recognizes and names Rintje
• Recognizes family relationships
• Correctly mentions that Rintje secretly eats the egg

The full episode of Rintje by ChatGPT5.2

Weaknesses

• Much emptier audio description, sometimes too brief because of insufficient context
• More frequently mentions incorrectly interpreted details ('mother looks out the window')
• Separate actions from different shots are sometimes merged into one long sentence (meaning sequence and timing are not always accurate)
• Often describes an event afterwards as a fact, without describing the action itself (Rintje discovers an empty egg, even though he was the one who emptied it)
• This creates confusion or leads to misunderstandings

Strengths

• Less repetition
• Better identification of all characters, although not immediately (initially 'female dogs', later 'mother' or 'grandmother')
• More often describes what is actually happening at that moment
• Better at describing the separate steps or actions within an event
• Better identification and clearer descriptions of actions lead to less repetition, a more logical structure, a more coherent story, and therefore better comprehension

Compared to professional audio description

AI does not provide overarching information and misses important elements beyond setting, characters, and events. This includes information about the animation style, opening titles, on-screen text, and aspects of the fictional world such as the appearance or clothing of characters.

Conclusion

The audio description of Rintje generated by ChatGPT4o results in a product of limited usefulness. The audio description produced by ChatGPT5.2 provides better viewing support overall. It is generally easier to understand in terms of who is doing what, and where. Because character identification is stronger and individual actions are described more clearly (rather than endlessly repeating the general situation), the story has more logic and is easier to follow.

However, the single most important story element — Rintje pulling a prank on his grandmother by secretly serving her an empty egg — is actually described less effectively by ChatGPT5.2 than by ChatGPT4o. ChatGPT5.2 describes Rintje holding up an empty egg that his mother had just been handling, implying that he merely found the egg rather than secretly eating it himself. This may lead to a misunderstanding of the events, for example suggesting that the mother ate the egg.

Please note: ChatGPT5.2 may produce different results in another attempt using the same source material and prompt — either better or worse.

Der Kaiser

Fragment: Kai sits on the couch wearing a mask next to a woman with a baby

ChatGPT4o

• Repeatedly and unnecessarily mentions that Kai is wearing a mask
• As a result, misses more important details (such as the mother entering and leaving)
• Mistakenly identifies another woman twice as his mother
• Fails to recognize his actual mother within the scene
• Describes individual actions fairly accurately
• Also mentions facial expressions

ChatGPT5.2

• Does not recognize the mother and father ('man and woman')
• Still misses important details (such as the mother entering and leaving)
• This creates confusion about who is doing what and where
• Describes individual actions fairly accurately
• Misinterpretation (a buttock is described as a pregnant belly)

The full film Der Kaiser by ChatGPT4o

Weaknesses

• A great deal of repetition
• Frequent misinterpretations (keeping the mask on when it is actually removed, stepping out of a car when Kai is refusing to get in)
• Sometimes incorrect identification or placement (another woman identified as the mother, 'inside the car' while actually outside)
• Main character Kai is not always recognized
• Sentence structure is often repetitive ('Kai does this, Kai does that')
• This results in an audio description that feels artificial

Strengths

• The main characters are usually recognized and named immediately from the start ('a boy named Kai', 'his mother Janice')
• Recognizes family relationships
• The opening scene is described relatively clearly, including important details

The full film Der Kaiser by ChatGPT5.2

Weaknesses

• Much emptier audio description, sometimes too brief because of insufficient context
• Still contains a lot of repetition
• Still contains many misinterpretations (mask usage, incorrectly described objects)
• Identifies Kai correctly, but therefore no longer describes his appearance (he is immediately referred to simply as 'Kai')
• Identification of other characters is weaker (the father is only identified later; the mother remains 'a woman')
• Misses important details (such as the prison setting)
• Separate actions from different shots are sometimes merged into one overly long sentence (meaning sequence and timing are not always accurate)
• Sometimes creates connections between scene transitions that are not actually there
• Interprets more frequently, which adds liveliness, but is not always factually accurate

Strengths

• Forms well-written sentences when viewed in isolation
• Language use is richer, more vivid, and more varied
• As a result, it is more pleasant to listen to

Notable

• Occasionally also describes the positions of characters ('in profile', 'seen from behind')

Compared to professional audio description

Where professional audio description serves the overall storytelling and deliberately chooses what should or should not be described, AI often zooms in unnecessarily on irrelevant details. At the same time, AI misses important details, actions, or events.

Conclusion

Character identification in Der Kaiser has become weaker with the new AI model. This can create more confusion about the characters, their positions, and their actions. The language itself has improved, becoming more vivid and varied. ChatGPT5.2 may not provide clearer or more useful audio description than ChatGPT4o, but it does produce audio description that is more pleasant to listen to.

The most recent AI-generated audio description has improved stylistically, but because character identification has worsened, the story itself is communicated less effectively. In Rintje, by contrast, identification with the new model was stronger and the 'who, what, where' was clearer. In that sense, the results are exactly the opposite.