Case studies / ArtEd in 3D
The moment AI stopped improving.
In a roughly 20-second cartoon, a pencil speaks as a building comes to life. The creator spent hours trying to make one line land naturally. The breakthrough was a human decision to change the method.
November 11, 2026. Effort is creator-reported. Acceptance of the corrected dialogue does not establish approval of the whole film.
The Building / A lip-sync rescue
The next version wasn’t the answer.
A different method was.
The person leading the film created and directed a short animation for ArtEd in 3D. A pencil character speaks as a drawing becomes a colorful, dimensional building. Making the words meet that reveal was the creative task.
“And voilà! Presto, perspective!”
The line needed a believable rhythm, not just a moving mouth.
The problem
Every fix moved the target.
AI revisions changed the voice, words, timing and mouth poses together. The creator kept rejecting the result.
The human decision
Perform the line. Keep it fixed.
The creator recorded his own delivery. AI changed the voice, then fitted the existing animation to that fixed audio.
The outcome
The line fitted the cartoon.
The corrected delivery was incorporated into the approximately 20-second film, keeping the scene’s timing and gestures. The creator accepted the dialogue sync, not formally the whole film.
- Unsuccessful revision effort, creator-reported
- 10+ hours
- The Building cartoon, approximate running time
- About 20 seconds
Across two days. Exact active time and attempt count are unknown.
The spoken line was corrected within this film. Formal approval of the whole film is unverified.
More recorded cases
When a small fix changed other parts of a design
A homepage logo repair worked. Comparing it with the saved designs revealed other, unwanted changes. A separate ArtEd design case, not the Building film.
Read the homepage design caseResearch notes and other recorded cases
Separate records and illustrations, with their own dates and limits. These are not additional results from the Building rescue.
When correction goes too far
An everyday example: a diner asks for an undercooked steak to be cooked a little more. It comes back burnt. They wanted it cooked well, not taken to the other extreme. A good chef knows the difference.
Corrections can overshoot, change unrelated details, break working parts or leave the original problem unresolved. These are different failures. We do not know that they share one cause, and each can add work.
A second AI reviewer may catch mistakes if it has a clear standard to use. Otherwise, it may prolong the revisions. A director, writer, editor or specialist may know what to keep without being the person operating the software.
Human sense helps us recognise what the person meant. “A little more” is not an invitation to reverse everything. We can read it wrongly, too. Check what was meant before changing more.
What feels right?Beyond the résumé: who notices what is right?
Someone may know what to change without knowing Photoshop, coding or 3D software. AI can help carry out those decisions. A résumé listing tools may miss that judgment.
A proposed hiring approach: we view common sense as innate. That does not make every judgment correct or diminish experience and AI review. Any hiring exercise based on this idea needs testing for fairness and dependable results.
Ask someone to explain what they notice, not merely call it obvious. Familiarity, culture and personal preference can shape a judgment. Listening to another view is part of the work.
Remembering what worked before
The person leading a separate website setup remembered a previous website address connecting automatically. It suggested an approach for a separate website setup.
This is that person’s recollection, not an independently checked result. We do not know whether the new connection worked. Medical outcomes, money saved and time saved were not established.
A separate illustration review · November 12, 2026
A comparison image missed the human detail
As reported by the person reviewing the illustration: a comparison illustration on a different site appeared to show two left hands. A different main image had already been chosen. This happened separately from the November 11 website design review.
The apparent hand error was not independently checked here. We do not know whether it was repaired. A person noticed it. No money or time saved was measured.
Repeated connection advice · User’s account · November 12, 2026
Five attempts. Same locked screen.
The user’s account included repeated screenshots of the same Claude–Notion “Finish connecting a connector?” dialog, with “Make this window larger to continue” and a disabled Continue button. The user reported five failed attempts. The assistant kept offering unchecked advice. The user stopped the repetition and asked it to think differently.
We do not know what caused the problem or whether the connection ever worked. Time, money and mental effort were not measured.
Lesson: when the same approach keeps failing, stop. Look again, try a different approach and check what changed. Credit the person who stopped the repetition.
Separate animated film review · Creator’s account · November 12, 2026
A full minute is not a finished film
The person directing a different film rejected an unreleased, approximately 60-second version of the film. No final MP4 was confirmed as approved.
“5.2 / 10”
What the film creator wanted
The film creator described the characters, their gestures and the exact ending.
What AI made
10 completed Seedance jobs produced 76 seconds of video. The record also notes 60 seconds of generated music. These files were not approved finished work.
What the film creator rejected
Creator-reported observation: the mouths did not move or match the speech well enough. The characters lacked chemistry. The acting and pace felt flat. Music was missing or inaudible in the reviewed film.
What to check next
One option: ask the film creator to check a short acting and sound sample before another version. This approach has not been tested.
What the numbers cannot tell us
76 seconds generated (4 × 7s + 6 × 8s) does not mean a 60-second film was finished or approved. Generating music does not mean it could be heard in the reviewed version. Some clips may still be useful. Rejecting the film does not mean every file was wasted.
Changes in account balances, replacement quotes, earlier website credits and separate AI usage are different records. They cannot be added up as the cost of this film. We do not know the human hours spent or any income lost.
When a closer shot erased the story
The person directing the later It’s Just Art film saw the line “Look at them, they're in love” while the two-person drawing it referred to was cropped away. Earlier artwork was upside down; an abrupt transition and a weak wedding/bar punchline also missed the intention. The film creator wanted subtle expressions to let the audience imagine attraction, not forced explanations.
The file audit: 31h56m was calendar elapsed time, including breaks, not active labor. It found 43 unique altered named long-form MP4 hashes, excluding picture/delivery intermediates and the canonical source. That is not 43 approved creative decisions or paid generations.
The film remains rejected, with no accepted master. The historical 5.2/10 rating belongs to a different, earlier cut. Full active hours and tool-credit totals are unknown.
Common sense connects the words to what the audience can see. Creative direction protects that connection and leaves room for imagination. Someone still has to decide what the story needs and whether it works.
Read the story at The Art of ReframingEditorial review · November 13, 2026
The page worked. The words repeated themselves.
The person reviewing this website noticed “Human Value in AI” beside “The value is human.” The extra label repeated the idea. The reviewer asked to check meaning across every page, not just matching words.
Earlier technical and browser checks passed but missed the repeated meaning. The reviewer decided each nearby line should add something useful. The correction removes the extra label, keeping the title and brand link.
What is documented: the saved source and views on desktop and phone showed the wording. The reviewer identified it on November 13. This is one quality-control case, not a finding about all AI-assisted writing. Review time, cost, reader understanding and business effects were not measured. No saving is claimed. The reviewer still needs to accept the wording.

