Ordinary prompts get blocked more than you think
Safety filters are meant to stop harmful images. In practice they also stop ordinary ones. Some video models refuse everyday words for regulatory reasons, some image models reject professional portraits because they "look too much like a real person", and research on open models (arXiv:2409.17156) has found classifiers that flag some art styles and artists far more often than others.
The result is a lot of guessing. People spend more time working out what not to say than actually creating.
Why the filters exist: real people
Most of the tightening is about protecting real people, and the laws are now explicit:
- EU AI Act (applies from August 2026): AI-generated content must be labeled, with machine-readable marks.
- US Take It Down Act (2025): non-consensual deepfakes are a crime, and platforms must remove reported content within 48 hours.
- South Korea: heavy penalties for deepfakes of real people.
- Japan: an innovation-friendly AI law, with existing laws on defamation and portrait rights still applying.
Protecting real people is the right goal. The problem is only when filters are vague and silent about why something stopped.
How to word prompts so legitimate work goes through
- Describe, don't name. "A woman in her 30s with short black hair" works; a celebrity's name will (rightly) be refused.
- Give ages for people. "A 28-year-old chef" is clearer to a model and to a filter than a vague "girl" or "boy".
- Use art-direction words. Lighting, lens, setting and mood carry a picture further than loaded adjectives.
- For conflict scenes, frame them as fiction. "A fantasy battle between two armies" reads differently from violence aimed at real people.
How EGAKU AI handles it
- Published rules instead of hidden ones — see the content policy.
- A clear message when something stops, saying why, in your language.
- Credits back when a result is blocked after it was made.
- A takedown form without an account for anyone who finds their own likeness: /takedown.
Clear lines, explained refusals and fair refunds make filters something you can work with rather than guess around. → Try it free