A year ago, someone ran a quietly brutal little test on every AI video tool they could get their hands on. Same photo, same instruction to each one: cry hard. Real, messy, tears-down-the-face crying.
Not a single model could produce one tear.
That test comes from an AI creator I follow on LinkedIn, and emotion has always been their favorite way to judge these tools. I love that approach, because emotion is exactly where AI used to fall apart. Think back to 2023. Every generated face had that frozen, glassy, uncanny stare. Something was always a little bit off, and you could feel it instantly.
What changed this year
This time, the original poster ran the emotion test again across a full lineup: Seedance 2.0, Gemini Omni, Kling 3, Happy Horse, WAN 2.7, and Grok. The difference from last year was night and day.
According to the expert, Seedance 2.0 reshaped the whole landscape, and Gemini Omni is now catching up fast. The faces actually emote. The tears actually fall. The thing AI couldn’t fake twelve months ago is suddenly on the table.
And this isn’t just for AI drama clips and short videos. The creator points out it unlocks a much bigger use case: storytelling that needs believable AI avatars, the so-called digital humans.
The part that blew me away
For a long time, making one of these talking-avatar videos meant stitching together separate pieces:
- the video
- the audio (voiceover, music, and sound effects)
- a lipsync pass on top that still looked a bit fake
The savvy professional behind this post highlights the real leap: you don’t generate those elements separately anymore. You generate all of them in one go, inside the same video. No more gluing a fake-looking lipsync layer onto a clip and hoping nobody notices.
Pair that with workflows getting more agentic, and suddenly this stuff becomes realistic to scale, not just a one-off party trick.
Why this actually matters
The author frames the shift in a way I keep thinking about. Compared to this time last year, the cost has dropped by a full order of magnitude, while the quality has jumped up another. Cheaper and better at the same time.
So where does that leave us? Their take is refreshing. When the tools get this good, your role, your creativity, your taste, and your ideas matter more, not less.
We don’t celebrate, “Wow, AI can do this.” We celebrate, “Wow, you can do this with AI.”
That line stuck with me. The bar has risen, and stakeholder expectations rise right along with it. The tools won’t carry a weak idea. They’ll just make a strong one move faster.
How you can use this insight
A few practical ways to put the creator’s findings to work:
- If you tested AI video a year ago and gave up, test it again. The emotion gap that broke immersion is closing fast.
- Run your own “hard” benchmark. Pick the one thing models used to fail at (crying, laughing, subtle micro-expressions) and judge new tools on that, not on easy shots.
- Try generating video and audio together in a single pass instead of building separate layers, and compare how natural the lipsync feels.
- Spend your saved time on the story and the idea, since that’s the part the AI still can’t do for you.
What I find exciting here is the reframing. The win isn’t that a machine learned to cry. It’s that the gap between what you imagine and what you can actually produce keeps shrinking.
The full breakdown, including how each model handled the emotion test, is worth a look. Check out the original LinkedIn post for the side-by-side details.