Page 12 - Detecting deepfakes and generative AI: Report on standards for AI watermarking and multimedia authenticity workshop

Detecting deepfakes and generative AI: Report on standards for AI
watermarking and multimedia authenticity workshop

Figure 2: You can help stop the spread of deepfakes

Source: DeepMedia.AI

The workshop discussions highlighted that tools to generate deepfakes are improving fast,
making it increasingly difficult for people to distinguish between an original and a deepfake.
In addition, these tools are now widely available in the public domain.

It will become more difficult for journalists and governments to detect deepfakes without easy
access to detection tools, said Sam Gregory, Director of WITNESS. Difficulty in trusting the
provenance of content on the Internet includes the complicated issue of authentic multimedia
content being dismissed as deepfakes. Gregory highlighted that, although the number of
observed deepfakes is not yet that high, analysing whether or not content is real or genuine
still requires a substantial amount of time and that the results of such analysis then need to be
communicated to the public The time required for analysis would become more concerning,
therefore, if deepfake generation were to increase substantially.

In addition, Gregory observed that how the content was created, its context, and how it
is distributed are perhaps more important than who created the content. The context, for
example, could help to determine the motivation for the content's creation and whether it is
disinformation or not. With regard to determining the identity of the content creator, various
aspects of privacy and data protection would also need to be managed.

Tobias Bednarz, Legal Counsellor at WIPO, spoke to the challenges that generative AI poses
in the area of copyright, including questions such as whether or not AI-generated works can
be protected by copyright and under what circumstances the generation of content by AI can
infringe copyright.

There is broad agreement that content created by an AI system without human intervention is
not protected by copyright under current legislative frameworks. AI-assisted works, however,
may be protected by copyright.

Addressing the question of under what circumstances the generation of content by AI can
infringe copyright – the subject of ongoing lawsuits in several jurisdictions – Bednarz made a
distinction between outputs (content generated by AI) and inputs (existing copyright-protected

4