Use of Force training

Sworn employees of the Department of Taxation's Criminal Investigation Division (CID) are required to take an annual training on Use of Force. They carry firearms as well as less lethal weapons (e.g., pepper spray & tactical baton) as part of their duties and therefore must be well-versed in case law regarding use of force as well as proper deployment in response to aggression. After previous successful partnerships with the eLearning team (i.e., Consular Notification and Access), CID saw an opportunity to create a self-paced module to supplement in-service training on Use of Force. The visual storytelling in the project, particularly around the two court cases that set Use of Force precedent, quickly lent itself as an opportunity to leverage Generative AI to develop a series of images in a cohesive and compelling style. I experimented with ChatGPT, Adobe Firefly, Adobe Express, and Adobe Photoshop's generative capabilities as well as Articulate Storyline's AI tools for narration and quiz question writing to create the final product. I discuss the challenges and wins of working with the various technologies below.

Project Highlights

Look and Feel

For the look and feel of the course, I was inspired by wikiHow. In my previous life as an English as a Second Language (ESL) instructor, I'd relied on wikiHow to visualize a process or procedure. I liked their signature illustration style - simple cartoons that showcase the same characters completing an action panel by panel. Despite being a cartoon, the style worked well for adult learners. I was able to use this point of reference to create similar panel-by-panel images to show the context behind two Use of Force court cases. I used ChatGPT to create the initial images, prompting it to create all images in "a watercolor style in sepia tones."

I relied on details about the case to recreate settings like the convenience store shown here. I touched up the scene using Photoshop's generative fill to create the word "Mart" over the store and to change the look of the police officer in the squad car.

Then, I used Photoshop, Adobe Express, or Window's photo app remove background tools and isolate pieces of the scene (i.e., the squad car, main characters, etc.) so that I could use lighting and subtle movement to tell the story and add visual interest during narration of the case facts.

Character Consistency for Stories and Procedures

The second court case illustrates how well ChatGPT was able to maintain characters and scenery over multiple prompts in order to tell a story scene by scene. While I often used Adobe Firefly to create the initial scene, it was not capable of maintaining characters in the same way as ChatGPT.

This scene was challenging to create because it depicts violence. Both ChatGPT and Adobe Firefly would tell me when my prompts violated their policy, especially around the depiction of persons of specified races conducting violent actions. ChatGPT, however, helpfully told me what it could do instead and guided me toward editing my prompt to stay within guidelines.

This video clip showcases how well ChatGPT was able to create multiple scenes with the same two characters, which I was then able to animate within Storyline. I prompted each scene with how the characters should be interacting, and it maintained the background and character features fairly well. It's not perfect (e.g., I don't love how the suspect's hands look white when he's facing the officer), but it's still incredibly effective. This clip also showcases one of Storyline's AI voices (so good!).

Procedures

ChatGPT's ability to maintain characters was also handy for demonstrating the steps of a procedure. I needed to show the five steps for deploying pepper spray. I used five different prompts to show the various actions of the officer and suspect through the procedure. The characters do change slightly throughout the poses but not enough to disrupt communication of the message.

Perspective Change

This project was challenging because it involved a high volume of explanatory narration. I was able to manage some of the volume with the addition of markers to spread out the text and on-screen text to highlight key information. As I became more proficient with prompting, however, I found I was able to use different perspectives on the same scene to add visual interest as well. The two images below show the same scene from two different perspectives. I used Photoshop to cut out the characters, leaving them in sepia tones while I changed the background the black and white, drawing the eye to the action taking place.

This was the first scene I prompted, depicting the use of pepper spray on a suspect. On an interesting note, up to this point, I'd been prompting "police officer" and allowing ChatGPT to create its vision of a police officer. I finally realized that I could be providing more specific details on how I wanted the officers to be dressed so that I could more closely match them to the actual uniforms of our CID officers (i.e., black polo and khaki pants).

ChatGPT did a good job of shifting the perspective so that we can now see the two officers as well as some 'innocent bystanders' behind them. I will note that it was a bit lazy in creating the female officer. These twins could be the start of their own brother-sister cop show!

Combating AI Bias

Another challenge I encountered with AI image generation was a bias toward sexy pictures, particularly sexy ladies. ChatGPT was better at creating 'normal' people, but I often had to rely on Adobe Firefly because I was using the free version of ChatGPT that limited me to five images a day. This bias is illustrated below when I tried to prompt flushing someone's eyes with water after they were sprayed with pepper spray.

You can see above the progression of my prompts as I tried to strongarm Firefly into creating a non-sexy image. More often than not, the water was pouring into a pair of voluptuous lips - despite my best prompting to show "eyes clinched shut" and "mouth in a grimace".

I eventually got to the image above after much tweaking and convincing Firefly that the person in question was in pain, not posing for the latest swimsuit edition of Sports Illustrated.

I really liked this more realistic depiction of a woman from ChatGPT. However, it took some finessing to prompt an image of someone who had been sprayed with pepper spray. ChatGPT said my prompts of someone being sprayed violated their policies, so I had to prompt someone in pain and coughing with water droplets on their face.

The woman on the right above is another good example of ChatGPT's ability to create less biased images of women. I did have a hard time getting the AI to create her facing an aggressive person. I eventually used Photoshop's generative fill to create the person on the left.

Quizzes

Lastly, I was very impressed with Articulate Storyline's AI Quiz generation tool. It allowed me to select slides from which I wanted it to draw question content (e.g., all the slides regarding pepper spray). The questions it wrote were equivalent to something I would write myself (though it did it much faster), so nothing special there. The answers it wrote, however, were probably better than I could write. The distractors, in particular, were so good. It's clear that Articulate trained its AI with expert educators. I've captured four AI-generated quiz questions and their answers below.

Page updated

Google Sites

Report abuse