Synthetic Media — What it is and why it’s important
“Synthetic Media” functions as a collective term for media such as video, images, and audio that are manipulated or even created using technologies from the field of Artificial Intelligence (AI).
They are only just finding their way into public consciousness. One type of application attracts particular interest: “deep fakes,” video in which, for example, faces or texts are exchanged in videos.
However, synthetic media possibilities go far beyond deep fakes: In the coming years and as they mature, they will influence and change many media production areas. As the example of deep fakes shows, this will is accompanied by far-reaching and profound discussions for society, politics, and the economy.
Synthetic media are still at the beginning of their development and use. Their future role and reception are open. The idea that Donald Trump will create his own media network and cater to the needs of an already fragmented media world with personalized, synthesized avatars is not out of the question. But synthesizing people, be that popular or not, is not the only way how synthetic media — as a market of a broad array of startups and media companies — will manifest itself.
It is also noteworthy how the companies operating in this field are vividly aware of their products’ relationship with “deep fakes.” For example, Descript addresses their product’s issues being used in a “deep fakey”-way right in the promotional video (it’s exceptional, and you should watch it). Or Reface that received a substantial investment sum from a16z last year and dedicates a portion of it to develop a tool that will detect their deep fakes, so to say.
Example: Video
The principle of Synthesia is simple but powerful: You give the platform a text and get back a video in which a digital human speaks the text in lip-sync. With Reuters, Synthesia has developed a prototype in which a generated sports presenter automatically guides you through a sports broadcast, based purely on data.
No other medium represents the potential of synthetic media like video. This is mainly because the artificial generation of people in videos has been enormously costly for a long time. Film productions used huge budgets, 3D designers, and server farms to create scenes that usually lasted only a few seconds.
The principle of Synthesia is simple but powerful: You give the platform a text and get back a video in which a digital human speaks the text in lip-sync. With Reuters, Synthesia has developed a prototype in which a generated sports presenter automatically guides you through a sports broadcast, based purely on data.
No other medium represents the potential of synthetic media like video. This is mainly because the artificial generation of people in videos has been enormously costly for a long time. Film productions used huge budgets, 3D designers, and server farms to create scenes that usually lasted only a few seconds.
The principle of Synthesia is simple but powerful: You give the platform a text and get back a video in which a digital human speaks the text in lip-sync. With Reuters, Synthesia has developed a prototype in which a generated sports presenter automatically guides you through a sports broadcast, based purely on data.
No other medium represents the potential of synthetic media like video. This is mainly because the artificial generation of people in videos has been enormously costly for a long time. Film productions used huge budgets, 3D designers, and server farms to create scenes that usually lasted only a few seconds.
Example: Audio
Descript not only handles transcribing speech from a podcast recording, for example. It also makes it possible to make changes in the text transcript, which are then generated again as speech. Speech no longer needs to be edited in an audio tool. Still, it can be edited as easily as a Word document.
Example: Text
Virtual Ghost Writer is a simple web tool that works as an interface for the GTP-3 algorithm to generate whole texts based on keywords.
There are examples of algorithms automatically writing texts for selected topics that the audience can hardly distinguish from texts written by humans. These included texts about sporting events, stock market news, and business reports. These contexts have in common that they usually bring along lots of structured data based on which an algorithm can generate texts.
In the meantime, research on Natural Language Generation has progressed rapidly. The current focus of attention is the Generative Pre-trained Transformer 3 (GPT-3), a statistical model for the computational understanding of speech based on Deep Learning, presented by the AI research lab OpenAI in May 2020. It is by far the largest language model to date, with 175 billion parameters (more parameters lead to better language comprehension) — the next largest language model from Microsoft only has just under 17 billion parameters by a wide margin. The consequence: If you give GPT-3 only a few key points, it generates stringent texts of any length. The effect is impressive — and also a bit scary. The texts do not stand out by their individuality but instead because they are consistent with the usual expressions used for a topic. As a result, this technology could have a lasting impact on areas such as content marketing.
