AI Compass
Compass

Images, audio and video

What can be generated and edited, which labelling duties apply, and how to tell that something is artificial.

·2 min read·By Redaktion KI-Kompass
DETAIL
4 sections

What is possible

TaskMaturityNote
Generating illustrations and backgroundshighText inside images stays hard
Cutting out, retouching, extending imageshigh
Transcribing audio to texthighMicrophone matters more than model
Reading text aloudhighCheck numbers and abbreviations
Generating subtitleshighCorrect names afterwards
Generating videomediumShort shots, high effort
Imitating a specific voicetechnically easyLegally delicate, see below

Spotting generated content

  • Look at hands, text in the image, reflections and repeating patterns. This is becoming increasingly unreliable.
  • Check provenance: who published it first, and can that be traced?
  • Check context: is anyone else reporting it, and on what sources?
  • For any request to transfer money, always check back through a different channel. That is the only reliable test.

The dials when generating

DialEffect
StepsMore steps, cleaner detail, longer compute
Guidance scaleHow closely the prompt is followed; high values oversaturate
SeedThe same seed reproduces the same image
Negative promptWhat must expressly not appear
Edge or pose conditioningTurns the generator into a tool with fixed geometry

See Generating images.

For business use

  • Do not depict real people, not even by resemblance.
  • Do not generate third-party trade marks.
  • Label generated images as such, at minimum in the metadata.
  • Read the provider's terms before the first campaign, not after.
  • Store results with their prompt and seed, so variants stay reproducible.

The duties under Art. 50 AI Act

WhoDuty
Providers of generative systemsMark outputs machine-readably as artificially generated
Deployers, for deepfakesDisclose that the content is artificially generated or altered
Deployers, for texts on matters of public interestDisclose where published without editorial responsibility
Providers of dialogue systemsMake it recognisable that a system is speaking

Exceptions exist for obviously artistic, satirical and fictional works, subject to the proviso that disclosure must not impair the presentation.

Watermarks and their limits

Methods embedding an unobtrusive marker exist and are in use. They survive light compression but not every kind of post-processing, and they evidence origin from a system rather than the absence of manipulation. A watermark is therefore a duty-of-care measure and not protection.

What works against voice fraud is organisational: a call-back procedure to a registered number for every payment instruction, four eyes above a threshold, and an agreed code word for exceptions.

The labelling sheet

FREE ACCOUNT

Labelling and rights for generated media

When labelling is required, which rights are affected, and what to settle before publishing.

Checklist2 items

No password needed. We send you a sign-in link. An account does not subscribe you to anything. The newsletter is separate.

Was this page helpful?
Images, audio and video