Generative AI in Foresight

Futures Methods

With no class topics to cover the week of Thanksgiving, I want to turn to the use of generative AI in Foresight. It’s been almost exactly one year since the release of ChatGPT unleashed consumer-grade, instruction-following Large Language Models (LLMs) into the world, with society still discovering the second- and third-order effects. I thought it would be helpful to share how I’ve been using GenAI as an assistant in my Futures work, including some examples of prompts¹. For the record, I’ve been using both Anthropic’s Claude and OpenAI’s GPT-4 and find them reasonably comparable for most tasks.

General Purpose Writing

Like many people in the knowledge economy, I basically write for a living. In addition to the short stuff (emails and instant messages), I write a lot of long-form content. This includes articles (like this one!), documentation, FAQs, project proposals, etc. Having AI review these pieces won’t get you as far as a good human editor, but it can get you about 60% of the way there (and if you do have a human editor available, they can then focus on higher-level stuff).

Example prompt:

I'm in the Masters of Strategic Foresight program at the University of Houston. As part of this, I'm writing and publishing a weekly newsletter about what I'm learning and my growth as a futurist. I'd like feedback on the following essay for my newsletter. The tone is intended to be intellectual but also well-written, fun and engaging. Suggest any big or small edits, point out awkward passages, weak arguments, use of cliches, and other opportunities to improve the writing. I generally like Orwell's rules for writing.

[Article pasted in]

Drivers and Scenarios

Part of the Houston method is the development of the major drivers of the future, then their possible direction in various archetypal futures and scenarios for each of these. Because the process involves specific structured input, it’s easy to use AI for a first draft so you’re not starting with a blank page.

Example prompt:

You are a seasoned and creative futurist. You have been tasked with the writing of scenarios to be presented to the leadership of a [client type], giving potential futures for [industry] in the US in 2043² from the perspective of 2023. These scenarios should be the kind of narrative typical of scenarios in futures or strategic foresight engagements: 2 pages or so of narrative text telling a plausible story from the perspective of the future: present tense, at the level of institutions (calling out appropriate real-world institutions such as companies or parts of the US government by name), showing how the drivers interact and how this future may have come to pass. Use a standard framing device for this: a newspaper article from the future, a press release, etc. This is using the Houston method so the scenarios will follow one of four story archetypes: baseline, collapse, new equilibrium, or transformation. Start the scenario with a catchy name - something short and memorable. If anything is unclear, ask follow-up questions before writing the scenario.

One oversimplified example of the narrative structure: In 2043, XXXXXXX is (based on the scenario logic). In 2023, it began to feel the effects of (DRIVER) and (EVENT). But, then (DRIVER) happened, and somewhat amazingly, lots of people were surprised by (WEAK SIGNAL). This had a major impact on XXXXXX, specifically XXXXXXX. However, in 2025, XXXXXX took advantage of (DRIVER) and kept an eye on (EVENT / WEAK SIGNAL). And by 2026, XXXXXXX was doing things that would have seemed unimaginable before, including (DRIVER / EVENT / WEAK SIGNAL). In 2027, it became clear that XXXXXXXX. And, in this future, XXXXXX shows that XXXXXXXX is XXXXXXXXX.³

Here are the driver values for the [whichever] scenario, give me your best work:

[Drivers and baseline values]

What came out was serviceable but needed further prompting: I had to ask it not to use to 2023 as a pivotal year, to identify an event the story was framed around, and to work harder to show how the drivers interacted with one another. One additional trick was to ask it to self-critique:

How well do you think this scenario tells the story of all the drivers? Anything you think should be added or changed?

And then, satisfied with its suggestions, asking it to incorporate them (and some others I wanted). This still needed a fair amount of editing work, but I was impressed with the final result and it probably saved me about 75% of the time it would have taken to write something at that level from scratch.

I asked for names for the scenarios ahead of time, and then again at the end:

Can you think of catchy names for these four final scenarios (baseline, collapse, new equilibrium, and transformation)? Maybe something where they are all tied together by a common metaphor, cultural reference, or trick of language. They should be short, clever, and memorable.

The results were not impressive, but after a few iterations I found a good unifying theme and a couple of decent names, and then asked it for more rounds of suggestions until I found something adequate. Clever naming still seems like a weakness of these tools - I never felt like I saw a spark of true inspiration.

Game Development

I’m still very proud of the game I built for my final project this semester. Designing a decent game is not trivial, and though I have thought a lot about what makes a game successful, I would not have been able to develop it so quickly⁴ without AI as an assistant. I tried feeding it the assignment and generating ideas for the format, but none of the ideas were all that exciting - the idea to make a game came from my wife, who knows me better than any AI. I then thought about a game that might be a good mechanical inspiration; Inscryption is a great single-player experience that gives four different views of the same underlying game, which connected well in my brain to trying to deal with similar challenges under different scenarios. I used Claude to help me work through how I might implement this. An early prompt:

Here's what I'm thinking: once I develop a core set of mechanics, I could make 4 different skins/themes, using aesthetics to represent the 4 different potential futures. Each one could maybe have the same underlying cards but different distributions, or different starting conditions, environmental factors, etc. what do you think?

It gave a bunch of potential mechanics I could use to implement this idea. I then decided what direction I wanted to take and asked for more feedback:

OK so I'm thinking of it as a one-player game where the opposition is from a passive event deck. There's a set of trackers for problems like social discord, loneliness, political dysfunction, and suffering of the poor. Each event has a set of requirements about what kinds/combination of cards can resolve it, and for every turn it isn't resolved it moves the trackers some amount. Presumably resolving might give some benefit. For example, there's a natural disaster event. Every turn it stays on the board it increases suffering of the poor by 1. To resolve, you need to have either 3 religious citizens or 3 non-religious citizens and $5 (an in-game resource). Once resolved, it reduces social discord by one.

How well do you think this would work?

You can see that this is basically the form of the final game, but the conversation went on for pages, with me putting in my inputs from the class (like the driver values under the various scenarios), going back-and-forth about implementation details, brainstorming about the resources being tracked throughout the game (and appropriate icons), feeding it six ideas for events and having it suggest more ideas to get to 30, etc. There were some things it did a bad job with (like tracking attributes of the cards in a table), so there was a fair amount of work in spreadsheets etc, but having a conversational AI around is invaluable to speed up the generation and refinement of ideas. I did find that it kept trying to make my game more complicated, and so I had to restrain myself and prune away most of its suggestions⁵.

Art

For now, this isn’t something that can be done with Claude, but GPT-4 with DALL-E is pretty competent.

I’ve used AI to help me imagine images to illustrate my scenarios once I wrote them:

for each scenario, please suggest a striking image to illustrate the world of the scenario, including an appropriate subject and tone. Please write it in a way that could be used as a great prompt for a generative AI image generator.

After generating a few ideas, and a few attempts for each, something pretty good usually emerged.

When doing the art for the cards in my game, it was important to have a single unifying style, but AI can easily produce art that’s all over the place. I worked on defining a useful style:

I'm making a card game for a student project about the future of religion in the US in 2050. The basic idea is to have Events that occur each round, and the player has resources representing people with different beliefs that can come together to resolve the issues. I ended up using creative commons photographs for the resources because AI pictures struck me as too stereotypical. However, I would like to use AI-generated art for the event cards (pandemic, misinformation campaign, students needing tutoring, etc). What do you think are some good ideas of styles for this art that I can use in my prompts? I want a few ideas so I can pick one unifying style.

Its suggestions weren’t quite what I was looking for - cyberpunk, retrofuturism, biopunk, minimalist vector art, solarpunk, digital collage, and mystical futurism - so I got a little more specific about guidance, basically reverse-engineering the style I was looking for:

OK I like these, but most of them are almost like a setting. What's a good way to describe the kind of rounded cartoon style that's becoming popular in board games? I'm thinking about Katie O'Neill in Aquacorn Cove, Keith Pinkston for Mystery at Magnolia Garden, or to a lesser degree Stephen Universe.

It gave me an elaborate six-point list, so I had to shrink it down:

can you please condense these ideas into a single generic prompt that combines these concepts and I can use alongside the topic?

Then I was able to use its output as the core of the prompt, requesting images using variations on the theme by just changing the last sentence⁶:

Please draw an illustration in a modern cartoonish style that depicts a scenario in the modern-day United States. Emphasize soft and rounded shapes, a bright and pastel color palette, whimsical charm, clear line work, simplified details, inclusive representation, and emotionally expressive characters. Specific topic: students struggling with schoolwork and being tutored by community members.

This resulted in a great picture that really set the tone for my Event deck (check out the alt text on the image to see the prompt it created from my description):

Final Thoughts

I know when ChatGPT came out a year ago there was deep concern in education, fearing that students would use it constantly for cheating, but it’s really making us reconsider what “creation” means in the first place. This is an area where norms are changing fast: I still write all my articles and thought pieces myself and then use AI to edit, in order to maintain my voice and make sure the thoughts are my own (and I don’t waste the audience’s trust and time with the kind of vapid non-thoughts AI tends to write), but that distinction may soon seem as arbitrary and pointless as refusing assistance from spellcheck.

You’ve probably noticed that prompting like this may take some practice, but there’s not really any alchemy here: I give context, ask for what I want, and keep asking and tweaking until I get something good. The answers aren’t always incredible, but it’s hard to imagine doing creative work without it now; having an eager assistant that can generate infinite ideas and never gets tired or frustrated is really a remarkable thing. I imagine we’ve only seen the beginning of how Generative AI will affect creative work. I’d love to hear how these tools have changed your work - please share in the comments below.

1. I’m not sold on prompt engineering as the intricate art it’s often made out to be. I agree with Ethan Mollick’s points that precise prompts are less important the further these models get, and practice with LLMs for your own use cases is a better way to learn than reading tips from someone else. I’m also convinced that many of the “rules” people learn through playing around with prompts are just the product of superstition brought about by small sample sizes and confirmation bias.

2. Insert your own geographic scope and timeframe here.

3. I borrowed this example structure from John Sweeney’s Intro to Futures.

4. In this case, “quickly” means something like 60 hours over 4 weeks.

5. It’s always enthusiastic - at one point I literally told it “I have 4 weeks to develop this and am just one person, so I don't want to get too ambitious”.

6. I ended up removing any mention of the events occurring in the future, because it used too many future tropes that detracted from my goals. I think the USA setting was important to keep it grounded, but it also tended to fill my images with American flags.