Artificial Intelligence as Designers: Assistants or Substitutes?
610 Purdue Mall
West Lafayette, IN, USA, 47906
Though Artificial Intelligence technology is widely used in daily life, the long-standing concept is it is impossible for AI to be creative. However, on the contrary, some AI programs are already developed and used in some creativity required areas, like graphic design, composing and cyber gaming areas. Based on the analyzation of the machining learning style and theoretical basis behind those programs, a discussion is put forward about how Artificial Intelligence technology could contribute to the design field, what change would happen in the work content and job responsibility of designers, and what role designers would play in the predictable future. In short, Artificial Intelligence technology has the possibility of replacing designers’ position on cumbersome low-end tasks and enabling designers to focus more on evaluating and decision-making
Keywords- Artificial Intelligence, Design Process, Creativity
1.1. Current application of Artificial Intelligence technology
Artificial Intelligence technology is not an unfamiliar term to people nowadays, since it is broadly used in our daily lives. For instance, laptop battery optimization, customized advertisements on Facebook and Google, daily recommendation on Spotify—those are all achieved based on Artificial Intelligence technology, and from this aspect, Artificial Intelligence technology is providing great help on increasing human’s life quality. However, the development of Artificial Intelligence technology could also threat human in some ways. Rather than what Hollywood movies depict that Artificial Intelligence would form an iron army, defeat human and conquer the world, the real threat lies in the possibility of Artificial Intelligence technology could replace workers due to its high efficiency, productivity, learning ability, and “unlimited energy”. Automation technology has already been widely used in manufacturing industries, causing the unemployment of assembly line workers. Self-driving cars have gained much popularity and are becoming a hot future trend, led by Tesla Inc. What’s more eye-catching, the computer Go program AlphaGo developed by Google Deepmind beat Jie Ke, the world No.1 ranked player at the time at the 2017 Future Go Summit, which triggers the argument of if Artificial Intelligence could surpass human’s ability in some certain areas and replace the human working positions.
1.2. Artificial Intelligence technology in the creative area
The same kind of threat has seldom been mentioned in design or other creative areas since there exists a long-standing concept that AI could only learn from existing human works and could never have the chance to be creative. However, the appearances of several AI “artists” shake this solid idea. Grid or Wix are using simplified AI in the field of web design to help users come up with their website design conveniently and efficiently. The AI graphic designer “Luban” customized and generated 170 million banners on one night in 2016 according to the preference of each user. In the interview, the developer said it is still far away from its capability limit—if there exists one—and has the potential to become a “real designer” in the future. What’s more, in the songwriting area, Sony Computer Science Laboratories developed Flow Machine, the AI music maker, who has already composed its first single “Daddy’s Car” in the style of the Beatles, and their goal is to research and develop Artificial Intelligence systems able to generate music autonomously or in collaboration with human artists. All these examples prove that AI programs could gain the ability to generate artworks, and it is more thought-provoking that whether this ability equals the creativity of AI programs.
Generally speaking, machine learning is the basis of Artificial Intelligence development, and the evolution of AI programs’ creativity is largely dominated by the progress of machine learning. Therefore, studying the theoretical basis of those “creative” AI programs would help indicate the limitation and bottleneck of AI programs’ creativity developing. In this article, two AI programs in the design areas are analyzed and generalize to show how AI programs generate artworks, how Artificial Intelligence technology would affect the design area and how the working focus of designers would change in the predictable future.
First, the research about the basic concept of machine learning reveals how the machine gains the ability to “learn” from the external input and to what extent the machine relies on human’s feedback. Then the theoretical basis and design generating process of two specific Artificial Intelligence programs in creative fields—Luban and Flow Machine—are evaluated to show the specific steps for programs to generate artworks and the role human plays in the whole process. Based on that, the limitation of AI programs’ artistic creation ability is generalized.
After that, the AI design model and human design process are compared from multiple aspects to show the similarities and differences, and the capacity limit of AI programs is taken into consideration of which part of human design process could be assisted by AI programs and which steps could not be substituted. Also, the subject matters of researching papers and books talking about the Artificial Intelligence’s development on creativity are studied and inducted to see the potential design ability that machine could achieve in the predictable future. At last, a whole new design process of human designers collaborating with AI programs would be modeled, and a conclusion of what would be the future role of designers with the development of Artificial Intelligence technology would be put forward.
3. Main Body
3.1. Theoretical Basis of Artificial Intelligence Technology
To have a better understanding of how Artificial Intelligence could affect the current design process, one important point is that if it is possible for Artificial Intelligence programs to generate its creativity. To answer this question, it is necessary to understand how those Artificial Intelligence programs “learn” the knowledge and become independent in working. Generally speaking, the theoretical basis of the Artificial Intelligence technology is Machine Learning. It is one certain field of computer science that studies the practical and efficient way to give the programs ability to “learn” without specially programmed. Based on the availability of training labels in the training process, machine learning methods could be divided into two categories: supervised learning and unsupervised learning. For supervised learning, the program is provided with examples or templates and the desired goal. It is like the program is given a “teacher” to help it learn the process of how to arrange the given “materials” to achieve expected results. This kind of machine learning relies largely on the external input and human’s feedback to improve and refine the process. On the contrary, for unsupervised learning, no examples or desired results are provided. The computer program needs to find the relationship and structure of external input on itself, and there is no desired or “planned” result given.
The success of AlphaGo, which might be the most well-known Artificial Intelligence program in the past few years, shows the theory is applied well to practice. According to its introduction on Google Deepmind and the researching paper Deepmind has published, Alpha Go is presented with 30 million games played by human Go players from the database at the beginning4. AlphaGo would study and analyze how human experts would play under certain circumstances and try to imitate human experts’ movement in the real game. What’s more, it would also play with itself and calculate the winning rate to evaluate what move would most likely to lead to win, like establishing the “policy” of the movement in the Go game. Meanwhile, the value network of AlphaGo is trained to predict the outcome from the certain position by applying the “policy” to both players, and output a prediction based on the value function to guide the game.
It is clear that the research team is using supervised learning principle to train AlphaGo in the process of development, and researchers’ task is providing a large number of examples for AlphaGo to study and integrate, and provide the feedback on AlphaGo’s Go skill (compete with AlphaGo in the real game). For AlphaGo, it learns and imitates human experts’ movement in the real games, and also gain the ability to predict the future movement by practicing with itself. More simply, AlphaGo generates the work while human input data, set the desired result, and evaluate the result.
Fig. 1 Monte Carlo tree search in AlphaGo.
Similarly to example, the type of machine learning methods could also imply the possibility of creative ability of the Artificial Intelligence design programs indirectly. One significant feature is that for programs using supervised learning, they are actually “learning” from human’s work and learn the method of how human create artworks. Therefore, those AI programs may become a designer with super high efficiency but could never create works that surpass human’s art pieces (according to human’s definition of “good” art). In other words, the machine learning type of the AI programs in the creative field would reveal if the program is only mimicking human’s work or generating its unique artworks under no set criteria.
3.2. AI programs in Creative Field
There are already some well-developed Artificial Intelligence programs in the design field, or more broadly, creative fields, and two of them, Luban and Flow Machine, are selected for analyzation and evaluation to show the design algorithm and process of programs.
To begin with, the Artificial Intelligence graphics program, Luban, is very eye-catching in the creative field. It is an AI graphic design program developed by Alibaba AI Design Lab to support Alibaba’s online marketing (Alibaba is one of the biggest e-commercial companies in China). After Luban showed its excellent efficiency by creating 170 million advertising banners in one night on the “Promotion Day” in 2016, there are even some frightening hearsays discussing if this program owns the ability to replace graphic designers and influence the overall design market. According to an interview with one of the Luban project researchers, the inspiration of Luban is their marketing strategy about providing customized products recommendation for every user with the help of the algorithm and big data analyzation in 2015. After that, the team was considering if it would be possible to customize stronger marketing oriented advertising resources, like posters and banners design, for every user based on their preference. To accomplish the enormous design task following this plan, they decided to develop the Artificial Intelligence program Luban to help.
Generally speaking, the development of Luban consists of 4 main steps.6
Step 1. To let machine understand the elements of design. The research team divide the original design into layers to assign labels to different elements and summarize the design styles for the program to learn how elements could be arranged.
Step 2. Establishing the design database. After the machine understands the basic design framework, a design database is established for the machine to extract features from design elements and classify them.
Step 3. Generating designs. The machine would try to put design elements into the templates that the team imports. In this step, the team adopts the reinforcement learning method, while the machine serves as its own “teacher” and improve through continual trying.
Step 4. Evaluating design works. The machine-generated design works are evaluated by experts from its aesthetics and commercial values.
Fig. 2 Human’s and AI’s role in the design process of Luban. The top half contains the activities of design team and the lower half contains the activities of Luban. Created by Xiao Ma on Sep 15, 2018.
The design process model of Luban in figure 2 is created to classify the activities of Luban and design team. In the whole design process, the main work of Luban is trying to arrange design elements with templates imported, while design team is in charge of providing samples, evaluating the design works, and making the final decision.
Another example selected is the Artificial Intelligence “Composer” Flow Machine. According to its introduction on the official website, the key idea of the Flow Machine is to relate the notion of creativity to the notion of “style”. If the machine could understand the style of a musician or a certain kind of music, then it may be able to understand the creation behind that style and gain the ability to create something based on that. So far, Flow Machine could not compose as an individual composer since it still needs musician’s help on polishing and improving the demo. Instead, it fits better with the position of the musician’s assistant. For instance, if a musician wants to compose a new song, he could select the style and let Flow Machine compose a rough demo first, then improve on the demo or get inspirations from it.
Technically speaking, the songwriting process of Flow Machine is similar to the creating process of Luban. It is given more than 15000 songs to learn from and analyze. What’s more, since the rhythms and arranging for each song have its special characteristics, songs are presented in the category of styles to Flow Machine, like Jazz database, Rap database, Pop database. Then Flow Machine would analyze the songs in the same style set, figure out what is recurrent, and establish it composing logic, like what kind of chord progressions and melodic sequences should be arranged with what kind of chord. For short, it is like Flow Machine generalize and collect all the composing probabilities for the specific style.
Fig 3. Human’s and AI’s role in the creativity process of Flow Machine. AI’s activities are listed on the left and the research team’s activities are listed on the right. Created by Xiao Ma on Sep 15, 2018.
The next step in the songwriting process requires human’s intervention. Though the flow machine could work autonomously on its own, to write a song of high quality it still needs the feedback from human experts. Since a song consists of various parts—verse1, pre-chorus, chorus, verse 2, bridge, etc—the researcher would work with musicians and let Flow Machine know which part is qualified or which part still needs to be improved. In this process, the role of composer changes from writing a whole new song from scratch to picking up great parts of the song. Benoit Carre, the composer who worked with the Flow Machine research team on its first piece, has stated that “The machine pushes you to your limits. The melodies generated confront you to a choice. The machine helped me to create a song for it. There are no limits to you ‘creativity’, it’s up to you to give it your choices, to establish benchmarks. Once the framework is defined, if it is sufficiently coherent, the melodies generated can be very inspiring.”
In conclusion, the flow machine could learn how to compose based on human musical works and generate composing works, while human provide good examples to flow machine and make the final decision on the quality of the song. It is similar to human’s role in the designing process of Luban, and the commonality of these two Artificial Intelligence programs is they are mimicking human’s work and generating artworks with extremely high efficiency, while human makes the final decision of if the artwork is qualified and could be adopted. To have a clearer view of what's human’s role and what’s AI program’s role and how they affect and assist each other in the overall process, human’s and AI program’s behaviors are put in categories in the following figure. From the chart it is clear that human’s role is more about defining the problem, providing instruction to programs, evaluating and finalizing works, while AI program’s role is more about practically learning samples, generating and improving works. After comparing this chart with human designers’ design process model nowadays, it would be easier to see how would the role of designers be affected in the future.
Fig 4. Generalization of human’s and AI’s role in the creativity process of AI programs.
Created by Xiao Ma on Sep 15, 2018.
3.3. Predictable future of the change in the design process
To understand what effect Artificial Intelligence Programs would have on human designers, it is necessary to study which part of the human design process could be replaced by Artificial Intelligence programs. In addition, the process of how human designers generate “creativity” should also be taken into consideration.
Since there is already some design process in different fields, several design models are combined and integrated into the universal design process model. This model is more focused on the general procedures to make it suitable for the overall design area.
Requirement & Empathize: When designers receive a specific requirement from clients, sometimes the client does not know what the core problem is or in which direction should they try to find the solution. The design team needs to analyze the case and figure out what is the critical problem they need to solve.
Define: After the critical problem is narrowed down, the design team would set up the plan and research on related areas to clear the goal for the final solution or result. Multiple or parallel goals could be made at this moment.
Design & Prototype: Based on the research data the design team would start the creating of new concepts or blueprints. Some of the potential concepts will be made into prototypes or frameworks and tested.
Evaluation & Production: The design team would evaluate each concept based on their performance in the testing and select the concept that is great enough for production or adopt the qualified features for further design.
When the universal model is compared with the human’s role in the creative process of AI programs, it is clear that it is still necessary for human designers to get the requirements from the client to discover the critical problem. The difference happens after defining the problem. Usually speaking, designers nowadays would come up with several ideas and proposals for senior managers or clients to evaluate and select, and this process is very time consuming since it is usually not that easy for designers and clients to reach an agreement on all the details at the beginning and designers would have to make changes several times— sometimes useless works. However, with the intervention of AI technology, designers could rely on it to generate different concepts and proposals after it is specially trained. This process covers brainstorming, sketching, prototype making, rendering and so on—which would take so much time for human designers but not for AI program due to its ultra-high efficiency—and designers would save much time and focus more on the evaluating and improving the ideas put forward by AI programs. What’s more, designers can let the AI program try some non-mainstream or unsure ideas, and sometimes great works are inspired by those ‘wired’ or ‘deviant’ ideas.
Based on the comparison above, the new design process model is established based on the current technology and the expected development of Artificial Intelligence. Some steps are designed especially for AI programs or designers while the other steps require the collaboration of both sides.
To start with, designers could work with AI to improve the researching and analyzing efficiency. Actually, the search engine nowadays is one kind of AI technology, and with the development of AI big data analysis—that is how Google Suggest works—it is possible for AI to recommend designers the materials or information related to the project. What’s more, AI could assist designers in collecting and generalizing the materials and improve the efficiency of case studies and information analyzing.
After the critical problem is defined by designers, they would set the goal for the desired result and establish the database of related design works for AI to learn the certain design styles and establish its design logic. Since the more AI program has learned before, the more design skills and styles it has mastered, so with more experience the AI program has gained it may not be necessary for it to learn new styles for every project, but just call the proper database.
The next step would be letting the AI program do all the ‘repeating’ design works after designers preset some additional limits on the project. At this step, AI would try to arrange the design materials based on presets and its design logic in all possible ways. All the ideas would run through a self-assessment where the criteria are set by the AI program based on its generalizing of design work samples input. Ideas survive the assessment would be presented to designers, and designers’ task would be evaluating AI program’s works and selecting the qualified ones or providing feedbacks for the AI program to improve the idea. This process would help the AI program to optimize its design algorithm and become more ‘qualified’ or ‘skilled’ in design. Meanwhile, designers would pick the qualified ideas and move into the finalizing process.
The evaluation would be the most critical step in the overall design process model for several reasons. Firstly as I mentioned before, the AI program is not truly ‘creative’ and it is just mimic the design style which it thinks best matches the requirements input. Since it is impossible for AI program to generate the innovated idea, designers need to make the judgment on if the design style AI program uses is the best one for the project and how could it be improved to match clients’ criteria. Also, the design aesthetics and trend are always changing according to designers and consumers. Occasionally designers would come up with unprecedented design styles, and the ones that gain popularity among consumers would become the new trend. AI program could learn the trend but cannot lead the trend, and it would still be designers’ duty to develop the design styles.
Fig 5. New design process with the collaboration of designers and AI programs. The AI program’s roles are listed on the left and designer’s roles are listed on the right. Created by Xiao Ma on Sep 16, 2018.
4. Conclusion and Discussion
Based on the current technology and predictable future of Artificial Intelligence, the appearance of AI programs should not be considered as the threatening of AI could replace designers’ position. On the contrary, AI programs could serve as an effective assistant in the design process. The success of Luban and Flow Machine have proved that AI programs nowadays are not truly creative, and what they do is learning design samples, forming its design logic and mimicking human’s artworks, while in the final stage designers are still needed to evaluate design works and specify improving directions. Therefore, designers could apply AI’s ultra-high efficiency, ‘unlimited energy’, and superior learning ability to their design process to improve their design efficiency to a higher level.
If applied properly, AI programs could help with data collecting and generalizing to guarantee a more thorough researching process; their high efficiency in generating design concepts would enable designers to try all the possible design direction and have a more diverse brainstorm; and designers could co-create with AI to produce multiple design solutions by simply putting in the desired goal. All those would lead to a more efficient and elaborate design process, and those are just what Artificial Intelligence technology is capable of nowadays. With the development in the future, it could be expected that the application of some AI knowledge related to the creative fields, like Generative Adversarial Networks, would bring more positive effects to design area. In short, it would be better for designers to embrace the new technology and make good use of it rather than getting scared of it. The appearance of AI programs in creative fields would not change designers’ position, but their roles in the overall design process. With the intervention of AI programs, designers could concentrate more on the evaluation and innovation stage and put more energy into leading the new trend and elevating the overall design aesthetics.
 Avinash Royyuru, “AI, humanity and the future of product design”, https://hackernoon. com/ai-humanity-and-the-future-of-product-design-4dc566be372e, Sep. 14, 2017.
 David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Yimothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel & Demis Hassabis, “Mastering the game of Go with deep neural networks and tree search”, Nature, Vol. 529, pp. 484-489, Jan 28, 2016.
 Kees Dorst, Nigel Cross, “Creativity in the design process: co-evolution of problem–solution”, Design Studies, Vol. 22(5), pp. 425-437, Sep 2001
 Kyle Jennings, “Developing creativity: Artificial Barriers in Artificial Intelligence”, Minds and Machine, Vol. 20(4), pp. 489-501, 2010
 Y. Erden, “Could a created being ever be creative? Some philosophical remarks on creativity and AI development,” Minds and Machine, Vol. 20(3), pp. 349-362, 2010
 Ben Dickson, “Can Artificial Intelligence be truly creative?” PC Magazine, pp. 95-104, 2018
 Hung-Cheng Tsai, “Computer Aided Product Color Design with Artificial Intelligence.” Computer-Aided Design & Applications, 4(1-4), pp. 557-565, 2007
 Hui Ma, “Application of Artificial Intelligence in Industrial Design.” Revista De La Facultad De Ingenieria, 32(14), pp. 309-315, 2017.
 Alec Radford, Luke Metz, Soumith Chintala, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks”, under review as a conference paper at ICLR 2016.
 Dartnall, T. (1994). Artificial Intelligence and Creativity An Interdisciplinary Approach (Studies in Cognitive Systems, 17).
 Pereira, F. (2007). Creativity and artificial intelligence a conceptual blending approach (Applications of Cognitive Linguistics [ACL]). Berlin; New York: Mouton de Gruyter.
 Leo Hickman, “How Google can guess what you are thinking”, https://www.theguardian.com/ technology/2009/nov/19/google-suggest-guessing-thoughts, Nov 18, 2009
 Britton Hennessy, “AI can do a lot, but it still can’t be creative”, https://drakecooper.com/ai-sill- cant-be-creative/#, June 22, 2017.
 Mindy Weisberger, “Can Machine be creative? Meet 9 AI ‘artists’”, https://www.livescienc e.com/62713-can-machines-be-creative-ai-humans.html, June 1, 2018.
 Fabricio Teixeira, “How AI has started to impact our work as designers”, https://uxdesign. cc/how-ai-will-impact-your-routine-as-a-designer-2773a4b1728c, Oct 31, 2017.
 Lucy Jordan, “Inside the Lab That's Producing the First AI-Generated Pop Album”, https://www. seeker.com/tech/artificial-intelligence/inside-flow-machines-the-lab-thats- composing-the-first-ai-generated-pop-album, April 13, 2017.
 Elsa Ferreira, “How Sony Flow Machines’ AI takes itself for the Beatles”, http://www.makery.info /en/2016/10/14/comment-lia-de-sony-flow-machines-se-prend-pour-les-beatles/, Oct 14, 2016.
 Miklos Philips, “The Present and Future of AI in Design”, Toptal, https://www.toptal.com/designers /product-design/infographic-ai-in-design, 2018
 Avinash Royyuru, “AI, humanity and the future of product design”, https://hackernoon.com/ai-humanity-and-the-future-of-product-design-4dc566be372e, Sep 14, 2017.
 Wikipedia contributors, "AlphaGo”, Wikipedia, The Free
 Wikipedia contributors, "Machine learning," Wikipedia, The Free Encyclopedia, https://en.wikipedia.org/w/index.p hp?title=Machine_learning&oldid=860632691 (accessed September 25, 2018)
 Deepmind, “Exploring the mysteries of Go with AlphGo and China’s top players,” https://deepmind.com/blog/expl oring-mysteries-alphago/, April 10, 2017.
 David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Yimothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel & Demis Hassabis, “Mastering the game of Go with deep neural networks and tree search”, Nature, Vol. 529, pp. 484- 489, Jan 28, 2016.
 Yuan Cheng, “1.7 billion banners are created by ‘Luban’ AI design system”, UISDC, https://www.uisdc.com/alibab a-luban-ai-banner, April 20, 2017.
 Sony CSL, Flow Machine, http://www.flow-machines.com/#
 Lucy Jordan, “Inside the lad that’s producing the first AI-generated pop album”, Seeker, https://www.seeker.com/tech/artificial-intelligence/inside-flow-machines-the-lab-thats-composing-the-first-ai- generated-pop-album, April 13, 2017
 Lucy Jordan, “Inside the lad that’s producing the first AI-generated pop album”, Seeker, https://www.seeker.com/tech/artificial-intelligence/inside-flow-machines-the-lab- thats-composing-the-first-ai-generated-pop-album, April 13, 2017
 Miklos Philips, “The Present and Future of AI in Design”, Toptal, https://www.toptal.com/designers/product- design/infographic-ai-in-design, 2018