Particularly in the IT & tech industry, many players like to talk about “disruptive innovations” and how much these can result in a massive shift in economic weights. If you take a closer look, only a few such disruptions actually took place until recently. The Metaverse is not getting sufficient user numbers despite gigantic investments, e-cars are displacing established internal combustion vehicles far too slowly, and voice assistants like Alexa are being abandoned by Google as part of a wave of layoffs.
Real disruption
But then, in November 2022, ChatGPT came onto the scene and after a short time it was agreed that we were dealing with an innovative disruption – and this time for real.
The signs of this are clearer than in most other cases for many reasons.
- It is a technology that has been able to reach 100 million users the fastest of all previous technologies
- ChatGPT has no dependencies on other technologies or platforms – it works out-of-the-box
- Even though the topic of AI is always associated with ethics and is therefore classified as critical, ChatGPT shows a path that has surprisingly little “ethical ice” in store
- The increase in productivity that comes from using ChatGPT is transparent and measurable for everyone, rather than just verbalized in marketing channels
It is precisely the measurable increase in productivity that is a key factor that is difficult to resist, even as a “disruption doubter” – especially if there are also no profound ethical concerns to counter, as is easily the case with genetic engineering, for example.
But even if ChatGPT has gained popularity through press, media and personal experiences, this technology is only one representative of a whole technology genre, namely “generative artificial intelligence”. In the same breath, one must therefore also mention previously known tools such as GitHub Co-Pilot, DALL-E or Midjourney. Obviously, these tools were still a bit too “niche” for some users at that time and could only come out of their shadowy existence somewhat through the radiance of ChatGPT. In the meantime, however, even these niche tools are being looked at more closely.
Proven productivity increase
GitHub Copilot, for example, became the focus of a study that wanted to answer to what extent the underlying generative AI results in a measurable increase in the productivity of a software engineer. An AI, by the way, that is even used in parts in ChatGPT.
It had already been scientifically proven before that the productivity increase brought about by AI has a significant impact on the job market, skills, and organizational structures. The results of the GitHub Copilot study are all the more impressive because they put the productivity increase into numbers: the group of developers with GitHub Copilot as support was almost 56% faster than the group without GitHub Copilot.
Now, it can be argued that writing code, or as examined in the study, specifically programming a web server in Javascript is not a creative task either, but is ultimately an automation of known hand movements.
UX, UI, Digital Design and AI
Creative industries such as design may benefit less from AI in terms of productivity. In terms of the IT industry, we are therefore talking about the disciplines of digital design, UX design and UI design. But the implementation part of these disciplines has long since been conquered. Not only can you create an appealing user interface design via Midjourney in a few minutes, at least for inspiration and as a basis for discussion, but dedicated UX & UI tools are already entering the market that promise to be able to generate entire screens in a way that can be processed further. It is still too early to say whether the promises of these tools can be fulfilled in practice, but the profession of the classic UI designer in its current form is definitely at least “ticked off”.
Digital Design and AI
This leaves the profession of UX designer or digital designer: here the focus is more on conceptual idea generation, deep domain understanding and good communication and sharing of UX specifications with software engineers. Surprisingly, however, ChatGPT’s domain understanding is so broad and deep at the same time that it can outplay a budding UX designer with ease. The following sample questions and answers impressively demonstrate that ChatGPT brings a solid domain understanding to the table.
ChatGPT: “The manual tasks that an operator must perform with the samples in the meantime when performing a semi-automated SARS-CoV-2 test on a laboratory machine can vary depending on the type of test procedure used and the laboratory machine. However, some typical manual tasks may include:
Sample collection: The operator must collect samples from patients and prepare them according to the test manufacturer’s instructions before placing them into the laboratory machine.
Sample processing: In some cases, the test may require manual processing of the samples before they can be used in the laboratory machine. This may involve steps such as washing, diluting, or preparing the sample.
Sample loading: The operator must load the samples into the laboratory machine, either manually or using an automated loader.
Monitoring the test progress: The operator must monitor the progress of the test to ensure that it is being performed properly and that all required steps are being carried out.
Evaluating the results: After the test is completed, the operator may need to manually evaluate the results if they cannot be automatically interpreted by the laboratory machine.
Cleaning and disinfecting: The operator must clean and disinfect the laboratory machine according to the manufacturer’s instructions to avoid cross-contamination between the samples.
It is important that the operator works carefully throughout the entire testing process and follows the test manufacturer’s instructions precisely to ensure accurate and reliable results.”
Research 2.0
It is easy to imagine that designers can increase their productivity immensely, especially in research work, if they make use of such a powerful tool – however, the question remains whether domain understanding and bite-sized communication then continue to be a key task of the designer at all or whether Subject Matter Expert and Software Engineer do not now also come to the required insights in a direct way at this point and a designing mediator role is therefore no longer needed. A possible answer to this question will be given later for the benefit of the following consideration.
Idea generation
Indeed, if we shift our gaze to the topic of “generating conceptual ideas”, the profession of UX and digital designer seems to stand somewhat safer at first glance. We are entering the field of “creativity”.
Good design depends on good ideas and creativity is therefore a necessary ingredient. However, it is only worth turning into a constructive implementation process when precisely fitting ideas are developed for given requirements. The creativity of the designer can therefore only be exercised in limited corridors along certain requirements. This is a good thing, because it is precisely such constraints that ultimately inspire the idea generation process. At the latest with the advent of mainstream methods such as “brainstorming”, it has been proven that many ideas also produce more good ideas, i.e. there is a correlation between the quantity and quality of ideas. However, time pressure and limited budgets often cap this calculation in practice. Moreover, in a human-centered design process, the true quality of an idea is not yet known at the time of idea development – this assessment can often only be made after implementation and the collection of user feedback.
Human limitations
Before the potential of an idea-generating AI can be brought into play, it is worthwhile to first look at the human limitations in the idea process:
- Emotions: Idea generators are often proud of their own ideas and therefore close themselves off to the ideas of others or react negatively emotionally to criticism
- Misdirection: Ideas often miss the mark without this being noticed at the moment the idea is generated
- Unpredictability: The process of idea generation is unpredictable and the quality and quantity of ideas strongly depends on the current work environment, the facilitation skills as well as experience of the participant group
- Inaccessibility: The people who have relevant domain or requirements knowledge do not necessarily have access to creative design processes and therefore need “enablers” who can “elicit” this knowledge and translate it into ideas.
- Bias: Existing ideas, “group-thinking,” anticipatory obedience, or other biases limit the openness and imaginative capacity of idea generators, which ultimately also prevents new, divergent ideas from emerging
- Limitation: No matter how creative a person may be – no one can generate new ideas without further input. So there must always be external impulses or new experiences, for example through further UX research or through mutual exchange in the group.
Potentials and limitations of AI
Based on these human constraints, a more concrete assessment can be made of how generative AI stacks up in terms of more effective idea generation.
- Emotionality: an AI has no emotions per se, so the point goes quite clearly to the AI.
- Direction error: the AI is – similar to a UX designer – dependent on the information and contexts it is provided with for idea generation. Ideally, it is provided with rich contextual information on the basis of which more precise ideas can then be generated.
- Unpredictability: AI is not necessarily deterministic in its results, as a same prompt may sometimes lead to a different result the next time. Nevertheless, a prompt can be controlled more easily than work environments, facilitation skills, or participant motivations, especially in a heterogeneous idea workshop group.
- Inaccessibility: an AI combines solid domain knowledge with solid idea generation expertise. But an AI also needs an “enabler” to deliver useful results. For example, a domain or requirements expert must contribute his or her deep specific domain knowledge so that this can subsequently lead to accurate ideas.
- Bias: with the help of an AI, contexts or biases can be deliberately considered or avoided, depending on what is used as the “source of truth” for idea generation. Specifically, the more solution-agnostic requirements, domain knowledge, and constraints are formulated, the less likely the AI will incorporate bias into the ideas generated from them – and vice versa.
- Limitation: AI is basically limited only by the data with which it was originally trained. This “longer long-term memory” means that it will reach its limits much later in practice than a human. However, it is certainly also at a disadvantage when it comes to original or “genuinely new” ideas, since it is indirectly dependent on ideas that were once expressed by another human being.
Restriction in the generation of ideas | Human | AI |
---|---|---|
Emotions | Proud of own ideas. Negative reaction to criticism. | Basically unemotional. |
Misdirection | Superficial understanding of requirements. Incorrect understanding of requirements. Missing requirements. | Depending on the scope, detail, accuracy and freedom of solution of the requirements. |
Unpredictability | Depends on the working environment and the moderation performance. | Controllable environment through training data and prompts. |
Inaccessibility | Transferring domain knowledge into ideas may require "enablers". | Domain expert must himself act as an enabler. |
Biased | Cognitive distortions | Depending on prompt design and training data. |
Limitation | New ideas need new impulses and experiences. | Depending on the size of the training data, but in principle not limited. |
Limitations in idea generation as a comparison of human vs. AI
Everything is relative
On the human side, it can be assumed that, thanks to their profession, UX and digital designers are somewhat less affected by some of the limitations listed, since, for example, an “agile mindset” contributes to a strong reflectiveness and good critical faculties. The egocentric pride in dealing with one’s own ideas is “trained away” from the profession.
Conversely, on the AI side, it must always be assumed that the underlying model, the training data, as well as the exact nature of the prompt will have a decisive influence on whether or not the results can be superior to human ideas.
And of course, despite all the hype about ChatGPT and co., the limitations of AI must be clearly pointed out at this point. For one thing: the outputs of ChatGPT may sound extremely convincing and yet can be completely wrong when it comes to delivering real facts. The AI is then “confidently wrong”.
However, this does not apply to the generation of ideas, because here, on the one hand, it is less about “the truth” than about “sparks of inspiration” that move the process forward and still have to go through some validation steps in the course of the user-centered process anyway.
It is also doubtful that the AI comes up with ideas that have not already been written down in some form and thus had to be explicitly part of the training data. However, this pool is very large and expands automatically over time.
Despite the inclusion of this relativization, the comparison shows one thing above all: in generating more ideas in less time, i.e. in “idea productivity”, AI is superior to humans in general – and also to designers in particular – in many respects. The benefits of even a purely text-based AI in everyday design practice are therefore already beyond question.
Prompt Engineering for UX and Digital Design
In terms of gaining a productivity boost, UX and digital design is not yet fundamentally different from other disciplines, as even teachers and students are increasingly becoming prompt engineers if they don’t want to be at a competitive disadvantage to the others. So where does the core lie when it comes to prompt engineering specifically for idea generation in UX and Digital Design?
The fact is: stereotypical prompts are “rewarded” with stereotypical answers, as the following example shows in the case of a time tracking app.
ChatGPT: “Of course. Here are five ideas on how to make a time tracking app more motivating:
Gamification: use game-like elements like points, badges, or leaderboards to make time tracking more fun. You can also offer rewards such as coupons or discounts for those who regularly record their working hours.
Personalization: let users customize their app by allowing them to upload their favorite colors, fonts or images. This will help them identify with the app and feel more motivated to use it.
Ease of use: make sure the app is easy and intuitive to use. Straightforward tracking of work hours and quick navigation through the app can help users not feel overwhelmed.
Notifications: Send notifications or reminders to users to remind them to record their work time. Make sure notifications are friendly and not intrusive.
Collaboration: allow users to share their work hours with colleagues or supervisors. This can build team spirit and help users feel more responsible about recording their working hours.”
There is no doubt that there are some good ideas here, but they are hardly sufficient to derive a concrete app concept, let alone a gamification concept. Further context must be provided in order to arrive at answers with more substance. For example, interspersing domain knowledge is a good way to arrive at better ideas.
ChatGPT: “Working time limits: Show users the legal working time limits so they know how long they can work before they need to take a break. When a user reaches a certain limit, you can send them a notification to remind them to take a break.
Link to employment contract: Provide users with the ability to enter or import the terms of their employment contract into the app. This can help them better understand work hour laws and ensure they are in compliance.”
It is equally important to incorporate relevant insights from user research into the prompts.
Our user research has found that most users (nurses in hospitals) in our main target group are not very competitive, so the points and rankings would tend to scare them off. Moreover, they actually already want to record the times in their own interest, they are just often too stressed to think about it.
How could you specifically motivate this main target group to record their times regularly? Please give me only the top 3 ideas.
ChatGPT: “Based on the specific needs and insights of your target audience, here are three ideas to motivate hospital nurses to record their times regularly:
Make recording easier: Simplify the recording of working hours by providing an intuitive and user-friendly app. You can also offer features such as automation or advance scheduling to further facilitate the recording of working hours and reduce stress for nurses.
Emphasize the benefits: highlight the benefits that regularly recording work hours can have for caregivers. Show how tracking hours can lead to greater transparency, fairer working conditions, and possibly even higher wages. Also provide a way for caregivers to present their own data to show them how their work hours affect their lives.
Support collegiality: build features into the app that allow caregivers to share their experiences and challenges and support each other. By creating a community within the app, caregivers can encourage and support each other to track their work hours more regularly.”
The ChatGPT examples shown also demonstrate another interesting fact: the generation of a good idea is not a punctual or completed event, but an iterative dialog. One cannot expect to arrive at really good and less stereotyped ideas unless one approaches prompt design with a certain attention to detail. So it pays to get creative in designing prompts so that the AI generates more creative ideas, as the following example will illustrate.
Show a dialogue of a nurse talking to her colleague during a shift handover about her time tracking, describing ideas on how she already tracks time during individual activities.
ChatGPT: “Caregiver: “Hey, I’ve noticed that I have trouble keeping track of my work hours. I work so much and I often forget when I started or when I stopped.”
Colleague: “Yes, it can be difficult, especially during a stressful shift. But it’s important to keep accurate records of our times so we get paid fairly and equally.”
Caregiver: “Yes, I know. But I have some ideas about how to do it better. For example, I could record my time when I give patients medications. I could just note the time when I gave the medication and record it later.”
Colleague: “That’s a good idea. You could also record the time when you reoccupy a bed or fill out a document. Those are all activities where you need a time anyway.”
Nurse: “Right, that’s a good idea. I think I’ll try it out and see how well it works.”
Colleague: “Yes, and once you’ve tried it out, we can get together at the end of the shift and see how well it worked. Then we can support each other and make sure we’re recording our times accurately.”
Nurse: “Sounds good, thanks for your help. I think I’ll feel better knowing that I’ve recorded my hours accurately.”
Colleague: “Exactly! We need to support each other and make sure we are paid fairly and equally. It’s important for us and for our work.”
This prompt not only provides a creative new idea (“Record times later”), but also links it to the hospital domain (“Activities where the time is needed anyway.”) and furthermore delivers it in an exciting to read dialog format (“Storytelling”). Nevertheless, it remains unclear here whether time recording at the activity level is necessary at all in the hospital domain. Here, only domain experts and user researchers can provide the necessary validation by “doubting” parts of this dialog against the AI.
In everyday hospital life, however, not every single activity has to be recorded, but only the start and end times of the working day as well as breaks.
ChatGPT: “Nurse Lena: Right, you’re right, Tom. In our hospital we actually only have to record the start and end times of the working day as well as the break times. I got a bit lost in the details.
Nurse Tom: No problem, Lena. It’s still good that you thought about it. We should focus on recording our working hours and breaks correctly to make sure we comply with the law.”
UX and digital designers must therefore pay more attention to the question of how AI can deliver more accurate and valid ideas based on concrete domain facts and user requirements. The underlying model and the training data cannot really be influenced by a designer in his day-to-day work, but the nature of his prompts and the format of the output can. So designers are increasingly becoming “user-centric prompt engineers”: instead of focusing on generating greenfield ideas, they need to bring given contexts of use and user research insights to the AI. They need to describe domain requirements in such a way that the AI can process them as well as possible and define the output format in such a way that it can be consumed by a third party as well as possible.
Design and requirements engineering
The formal and, if possible, precise, solution-agnostic formulation and communication of requirements used to be the task of the requirements engineer. However, this role has become increasingly irrelevant in recent years due to agile principles and the requirement for “lesser documentation”, because documentation needs to be read and understood and the time required for this is no longer available for the really important things: implementing “running software”. The software development world has become more agile and classic requirements engineering is too time-consuming in terms of creating, reading and understanding requirements documents.
Documentation in AI times
This could now change again, because the AI “reads” and “writes” faster than all team members together. The role of the requirements engineer, i.e., the specialist who formulates requirements as precisely as possible and now finds an effective customer in the AI, could thus achieve a renaissance. User researchers will also become more relevant, since they can validate and refine requirements with real users.
In both cases, AI can also be used to summarize both requirements and generated ideas or other complex issues in a way that makes them easier to understand.
Can you summarize the ideas of all prompts of this chat as markmap code and please without duplicates and in short bullet points instead of long sentences?
ChatGPT: ”
– Ideas for improving time tracking:
– Offer rewards
– Include progress bar
– Allow individual goal setting
– Offer personalized feedback
– Make working time models more flexible
– Display working time limits
– Link work contracts
– Simplify recording
– Emphasize benefits
– Encourage collegial support
– Set up alarms
– Track time during activities
– Collaborative review and improvement
– Ideas for improving work-life balance:
– Set boundaries
– Take breaks
– Improve time management
– Set priorities
– Practice self-care
– Ideas for improving teamwork:
– Improve communication
– Distribute tasks clearly
– Build trust
– Provide feedback
– Set common goals
– Ideas to improve productivity:
– Set goals
– Set priorities
– Improve time management
– Avoid multitasking
– Optimize workspace”
With the increasing popularity of text-based visualization tools such as Markmap, it is even becoming possible to graphically prepare textual summaries in such a way that the development team can cognitively absorb this information in significantly less time. The following mindmap was generated based on the above textual Markdown output, only by “copy & paste”, i.e. without further manual adaptation.
For those who like it a little more flowery, the essence of an idea generation can also be handed over to the development team in the form of a short story, i.e. by means of “storytelling”:
Give me a story, some of it verbatim speech, of two characters in a hospital context talking about how it’s a good idea to record break times directly on an app.
ChatGPT: “On a warm spring morning, Anna, a young nurse, was working in the intensive care unit of a large hospital. She ran from one patient room to another, tending to the many needs of her patients. While she was pushing a medication cart, she noticed her colleague, Jonas, sitting at the nursing station counter staring at his smartphone.
“Hey Jonas, what’s going on? You look so engrossed,” Anna asked curiously.
Jonas looked up and smiled. “Oh, hi Anna. I’m just testing the new app the hospital developed for us. It allows us to record our break times right here. Pretty cool, right?”
“That sounds interesting. Let me see it!” asked Anna, stepping closer.
Jonas held up his smartphone and showed Anna the app. “See? You can just check in here when you start your break and check out when you come back. The app will then automatically calculate how long you’ve been on break.”
Anna nodded in agreement. “That sounds really convenient. But why is it so important to accurately track our break times?”
“Because the hospital wants to make sure we all keep our legally required break times,” Jonas explained. “Plus, it’s also important for us to get enough rest and not overwork. If we record break times directly in the app, we can better manage our work hours and see if we’ve taken enough breaks.”
The biggest argument against solid requirements engineering – namely dry, overflowing documentation that takes a lot of time to write and read – falls away in no time with the new AI-based capabilities.
However, what AI needs and cannot be supplied by anyone except “real humans” is deep user research information and best domain knowledge. And even in the transmission of this human knowledge, AI can assist: Interviews with users can first be transcribed by AI, then made available to AI in text form, then used to generate ideas, and finally fed back to stakeholders in a cleanly processed form.
Conclusion and outlook
There are few limits to the productivity of an AI in generating ideas. However, this is only true if the underlying prompts are not stereotypical, but are packed with contextual and requirements knowledge and the requesting human has the will to iteratively refine them.
So, in fact, UX and digital designers need to lean more towards user research and requirements engineering to become better designers. They must become active in projects sooner rather than later and – in the face of AI – document more (for AI) rather than less. Reducing information to the team can also be taken care of by the AI. Needless to say, those who have already set their focus on user research and requirements engineering will not have to worry about being replaced by the AI soon. UX and digital designers, on the other hand, who focus more on finding solutions (especially the rapid production of wireframes) and place less value on a deep understanding of problems, will.
At the end of the day, AI assistance will result in those designers who use it making those designers who don’t use it obsolete.
We have aroused your interest? Take a look at our services!