They Can’t Make AI Look Real—So Real Now Looks AI

Written on 24 August 2025.

They Can’t Make AI Look Real—So Real Now Looks AI

Overview

YouTube has acknowledged an ongoing experiment that applies machine-learning “image enhancement” (unblur/denoise/sharpen) to some YouTube Shorts after upload. While YouTube says this is not “generative AI,” many creators report a conspicuous AI-upscaled/plastic look. The effect can blur the perceived line between human-shot and AI-generated videos—undermining trust in live, human broadcasting.

What changed (platform-side)

• YouTube confirmed it is running an experiment on select Shorts using “traditional machine learning” to unblur, denoise, and improve clarity.
• Creators (e.g., Rhett Shull, Rick Beato) and visual-texture channels (e.g., “Mr. Bravo” VHS aesthetics) documented sudden smoothing, extra-punchy shadows, sharpened edges, and waxy skin/hair in Shorts they did not process that way.

Confounders: viewer-side upscalers (lookalike artifacts)

Separate from YouTube’s servers, many viewers run browser/GPU features that AI-enhance any web video, producing the same “plastic/halo” look:

• NVIDIA RTX Video Super Resolution (Chrome/Edge; RTX 30/40) removes compression artifacts and upscales in real time.
• Microsoft Edge: “Enhance videos” / Video Super Resolution (VSR) ML upscaler available in Edge builds and some GPUs.

Implication: some viewers will blame YouTube even when the change is happening locally in the browser/GPU path.

Why this matters

• Aesthetic convergence—server-side “clarity” plus client-side upscalers—shrinks the perceptual gap between real footage and AI-made clips. Viewers lose familiar “human” cues (sensor noise, analog grain, fine hair/skin texture), making authentic videos feel synthetic.
• The broader context is a bot-heavier web: independent data shows automated traffic surpassed human traffic (51%), with 37% bad bots—fueling a general “dead internet” perception and distrust of what’s real.

How to verify if a video was altered (practical)

1. A/B across browsers/devices: play the same clip in Firefox (no Edge VSR path) vs. Edge/Chrome. If only Chromium looks plasticky, it’s likely viewer-side upscaling.
2. Toggle Edge VSR off: Settings → System & performance → Enhance videos (turn Off). Recheck the same frame/timecode.
3. Disable NVIDIA RTX VSR (if on RTX 30/40) in NVIDIA Control Panel; retest the same moment in the clip.
4. Check codecs/itags via “Stats for nerds” to see whether different pipelines are being used across devices (AV1 vs. VP9), which often correlates with visible differences.
5. Frame-match vs. your master (for creators):

# extract identical frames from master and YouTube copy
ffmpeg -ss 00:00:08 -i master.mp4   -frames:v 1 master_ref.png
ffmpeg -ss 00:00:08 -i youtube.mp4  -frames:v 1 youtube_ref.png

# quick objective check
ffmpeg -i youtube.mp4 -i master.mp4 -lavfi ssim;psnr -f null -

Mitigations for creators

• Avoid Shorts for texture-critical work until the experiment changes; publish as long-form where possible.
• Provenance + perception:
  • Embed Content Credentials (C2PA) / “Captured with a camera” in your workflow;
  • Link a reference still/mezzanine + SHA-256 hash in the description;
  • Mirror to a second platform (PeerTube/Vimeo/Rumble) and ask viewers to compare the first 10 seconds.
• On-screen test cues: open with 5–10 s of thin-line charts, hair/foliage, and natural motion blur/focus pulls—artifacts that denoisers struggle to fake—so viewers can instantly spot over-smoothing.
• Viewer guidance: pin a comment describing how to disable Edge “Enhance videos” and RTX VSR if the image looks “plastic.”
• Encoding strategy: retain some natural motion blur and micro-texture; if you add grain, do it sparingly (over-aggressive grain often triggers stronger denoisers).

Open questions

• Will YouTube expand image-enhancement beyond Shorts or add a creator-side opt-out/label?
• Will platforms reconcile aesthetic normalization with authenticity signals, so that “real” doesn’t automatically look “AI”?

See also

• YouTube Premium’s 1080p “enhanced bitrate” rollout (separate from ML “clarity,” but often conflated in discussions).

References

AI Disclosure: Parts of this page may have been created, edited, or assisted by artificial intelligence tools (such as ChatGPT or other language models). All AI-assisted content is reviewed by a human before publication. For questions, contact the site administrator.