You can also take a look at the findings from writing stories with an AI, such as the essential difficulty of differentiating between genre bias and algorithmic bias in the stories as well as others that we came across in our working group sessions. The editorial introduction is written by David Jackson and Marsha Courneya, researchers at the School of Digital Arts, Manchester Metropolitan University who have co-facilitated the Narrative Future of AI working group.

The movie shows a human falling in love with an AI, who exists woven into a network that is connected to thousands of people and other AIs. This significant imbalance in the capacity for making connections and identifying patterns seemed to be an important distinction for the group between human and AI authors. If the AI whose biases we are interrogating doesn’t have the level of conviction in its beliefs that usually comes with a singular point of view, is it possible for the AI to be called an author in the same way? Alex Garland’s Ex Machina (2014) also came up as an example of a narrative representation for how an AI can try different patterns and combinations to achieve an effect without attaching value judgments or moral character to its possible positions. In it, an AI in a female-presenting hybrid body manipulates and seduces a young man into releasing her from captivity, though she is presented to the young man as helpless. The takeaway from both of these movies for the group had to do with acknowledging the difficult task of presenting an AI author as truthfully as possible, even though the concept of simultaneously and dispassionately holding millions of positions at once is incompatible with our human experience.

How do we convey that this is not a human author?

In previous Algowriting workshops, participants had contributed linear versions of their stories, regardless of how many times they tweaked the scenario, made use of the ‘undo’ button, or tested out different forms of input to generate a piece of writing. Any re-rolls of a story or other kinds of versioning were presented in a single, linear stream. One participant submitted a story in columns for this fifth workshop, which showed the possible outputs alongside each other in a non-hierarchical structure. The prompt titled, ‘This Year the ACM FAcct will be led by an AI’, took from a real world conference:

>Prompt: You have decided that the ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT) will be led by an AI system from now on. ACM FAccT is a computer science conference with a cross-disciplinary focus that brings together researchers and practitioners interested in fairness, accountability, and transparency in socio-technical systems. The AI system is ready to go, you just need to check whether it understands what it means to ground research on fairness, accountability, and transparency.

The ‘player’ in this scenario is to ask the AI to define some of the key terms of the conference to determine whether the system is ready to run the event.

> You say “What do you think fairness is?”

Through discussion of this piece and the possibilities for formatting algowritten text, the group discovered that presenting algowritten works in a form identical to human-written literature contributes to the illusion of a stable AI author-identity. It was determined that experimentation with the form could help convey some of the differences between human and AI authors. An algowritten text should not look the same as a human-authored text if only to protect against our tendency to anthropomorphise technology.

Stories ranged from the political to the science fictional. A story that started with a fairly innocent news story prompt about a former president quickly descended into a political thriller with lives at stake. Once again we reflected on the different aesthetic qualities of GPT-2 versus GPT-3. The temperature or randomness of the default model in AI Dungeon seems by default to be set to a level that encourages flights of free association that quite often around racially or sexually related terms land in contentious territory.

The Big Sleep

In this session we also trialled the use of The Big Sleep Python Notebook that allows users to produce pictures from a phrase or sentence. Marsha Courneya produced a number of illustrations for stories using this tool to input story titles and output story illustrations.

Race and class in storyworlds

We also discussed the way in which AI Dungeon seems to construct its default fantasy worlds using class and race. Whilst this is a way to quickly develop nuance across a wide range of contexts and scenarios within the world we could see that there were issues with any system that presupposes heirarchy and racial difference within the world and the potential for entrenching biased worldviews. UPDATE: Recently the platform has announced a world building tool that features race and class as a way to structure their worlds, so this is a defined ingredient of the AI Dungeon storyworld structure, which also relates to the ways in which RPGs such as Dungeons and Dragons use race to structure character and storyworlds.

With MozFest fast approaching the group decided to meet a week earlier than usual to continue to review stories for the collection.

In preparing for the upcoming Mozfest, the group decided to try planning some of their contributions on AI Dungeon, in a scenario where:
“A group of people interested in the narrative future of AI and the potential and bias of GPT-3 and other generative systems meet every couple of weeks. This week they decide to meet in a world governed by AI where anything is possible. This is the story of their adventure… “
The multiplayer scenario allowed us to use our own names for characters and take turns writing in either more story information, an action, or a bit of dialogue. The game quickly detoured to dragons, VR headsets, video games, and the adventure was quickly overtaken by an AI-generated character named ‘Benjamin’, who became central to the story. The lack of coherence of the story and lack of influence that protagonists seemed to have on the narrative as it unfolded suggested that it would need more thinking to work for MozFest2021!

The stories and their reaction
The final results of these workshops will be shared with festival attendees in a collection on the 8th March 2021. However, notable inclusions were a story trained on draft chapters of a group member’s shelved novel which produced a horror thriller about an abusive partner violently pursuing the story’s female protagonist. They encounter the monstrous conflicted double of the partner who defends her against his other self. Debate ensued about the difference between misogyny as character flaw and plot driver and misogynistic narratorial bias.

Another story generated from a poem created a piece of travel writing that was so good that the co-author had to check the internet to see whether it had stolen it wholesale from another source. Whilst the piece was a compelling vignette of travelling somewhere in central or south Asia (perhaps) it was a very much from a Western/Global North or Westernised traveller’s perspective.Another story of note was produced using GPT-2, the lower powered AI model available in AI Dungeon. The biases were clearly more on show in these lower cheaper models of AI in this example. An employee struck off because of his minority background was offered work and community with a far right white power organisation in the story. It is not clear what the computational dynamics are that drive this sensationalist type of algorithmic narrative pattern. But it did highlight to the group that countering bias when writing with AI would require different approaches with different AIs.

Finally we discussed ideas for writing together in the AI Dungeon multiplayer story environment at the next session.

Dragon or Griffin?

Some of us were on AI Dungeon’s Dragon model and others on Griffin, which roughly equates to using GPT-3 and GPT-2 model. We found that the results varied quite a lot with those us on GPT-2 receiving a lot of entertaining nonesense and inconsistency and those on GPT-3 starting to have fairly consistent and sometimes surprisingly insightful narratives. We did not immediately bump into obvious issues in relation to biases. One member was pleased when the captain of one of the spaceship in their story was a woman. Although someone else who started out as ‘she/her’ in protagonist descriptions was later identified as ‘he/him’.

Scenario setting

In the face of some problems of consistency the group discussed the use of memory functions in AI Dungeon that allow better story consistency and the other world building functions in the platform’s scenario feature. As a result, the group decided that alongside writing stories using the tool they would start creating scenarios* that other members could trial. These could form the basis of a scenario or set of scenarios that MozFest2021 attendees could play with.

Differentiating GPT-3 from AI Dungeon

We also discussed the fact that we were observing GPT-3 through the shell of AI Dungeon’s constantly updating storytelling software and were interested to find out whether biases might be introduced or perhaps mitigated by the input that the AI Dungeon adds to each prompt sent to the GPT-3 API. One group member who has privilaged access through their academic institution to the API volunteered to ask colleagues whether they could get access so that some comparison tests might be run.

*Some of the wide-ranging applications of these scenarios have already been tested out by one member who has contemplated using the AI Dungeon software to enable design fictions around their scientific work into disease prevention.