A new AI tool has recently gone online that is supposed to help with the creation of screenplays - starting from a simple plot prompt that roughly outlines the plot, it spits out possible characters, locations and scene dialogs. Dramatron, as the system is called, is also capable of automatically generating entire screenplays or plays, but these - despite a new method - lack coherence in content, as is so often the case, which is why Dramatron is initially being launched primarily as an interactive writing tool.
Coherence - that is, an inner logic, a coherence - is so difficult to establish for Language Model Systems because they ultimately have no understanding of the content, since automatically generated texts are created as the result of probability calculations. Using a text corpus that is as large and varied as possible, a system learns when which words or phrases appear together where and with what frequency. If it is then to create a new text, it simply combines words according to the patterns found and adopted in the training texts. The longer a text is to be, the more noticeable the weaknesses become and the more memory a system needs.
To get around this problem when creating scripts, a new hierarchical approach was taken for Dramatron in combination with prompt chaining. From the plot prompt entered at the beginning, it first creates the characters involved, then describes the scenes needed and the locations where they take place. Only in a final step, based on these already created defaults, dialogs for each scene are then created (in parallel, they do not build on each other further than specified in the plotpoints).
Narratively, the system also has tight defaults. It is presented with an introductory prompt set virtually as a sample (in a arxiv.org/abs/2209. 14958 (accompanying study), the ancient tragedy Medea by Euripides was used as a prompt prefix, as well as one based on science fiction films) and must follow a sequence of plot points in the scenes, such as the dramaturgical pyramidal structure according to Gustav Freytag (exposition, triggering event, rising action composed of a series of conflicts, complications, and dilemmas, climax, descending action, resolution, and dénouement).
During the interactive writing process, all text components output by Dramatron can be edited, deleted, or regenerated, so that a play or screenplay is produced virtually as if by a virtual team. To test how well this works, 15 professional film/TV/theater writers were commissioned to create plays using the system; some of these were even later performed in a performance (Plays By Bots). According to the evaluation that is part of the study, almost all of the writers were positively surprised by the quality of the texts and found the writing process interesting to very entertaining.
Reservations were more likely to be expressed by authors from the theater sector, while screenwriters were more attracted to the hierarchically structured way of working - after all, simple TV series and typical Hollywood films are often created in a similarly standardized way. However, everyone agreed that the tool should only be used as a companion. Basic points of criticism were a lack of coherence, a partly high redundancy and a tendency to cliché. Likewise, character development and subtext are naturally missing, among other things. All this is not surprising from our point of view, but on the other hand also points that one can reproach many a film written by humans.
It may well be that it won&t be long before such scripts can be compiled by AIs like Dramatron itself (and that this won&t even be noticed...). Until then, the system can be used as an idea generator or for entertainment, with sentences like "You&re all in love with me, so I&m in love with myself."
Dramatron was developed by AI researchers from Google&s DeepMind and Stanford University and is available on the web, but not for free; a GPT-3 API license from OpenAI is required for use (the current underlying language model; Chinchilla was still used for the study). By the way, the study explicitly warns against taking over any text modules without checking them, since it is not possible to know whether the output sentences were taken over by the AI unchanged from the training corpus (and can thus represent a copyright infringement). more infos at bei deepmind.github.io