MidJourney-generated neon-lit brain, CC 2023

Common sense is about understanding and reasoning about everyday life based on one’s experiences and oneself. A fundamental aspect of common sense is a deep understanding of how things and people are interconnected in time and space and how ideas and theories are related. What sets humans apart from AI models is the level of deep understanding, human instincts, and subjective understanding that comes from our sociocultural background, upbringing, friends, and family.

GPT-4 and similar language models (a complete list) will never have memories of a time and place other than those generated through conversations and prompts. For example, it will never know what it smelled like when the snow melted on Västerhöjden, Hofors, in the spring of 1981, or how it felt to ride a tricycle. It will also never know what it feels like to play guitar with one’s body and create a song out of nowhere. However, it might perform some of these tasks artificially without personal experience. The difference is that it is not personally made, empathetically, and without a deep understanding.

In this article, I will explain what is missing in GPT-4 and why it will not replace me in my design work. Yet.

Deep understanding

GPT-4 can only play roles and is not yet a conscious being with thoughts and feelings. It has the ability to simulate, but these simulations are not yet real. It cannot have personal memories or experiences to provide context. It is quite creative, but not on the same level as humans. Additionally, GPT-4 has no emotions. It can understand semantic analysis in text, but that is its limit. GPT-4 cannot see, smell, hear, touch, or taste anything. These limitations combined make it impossible for GPT-4 to provide the same level of understanding or insight as a human in certain situations.

In other words, GPT-4 lacks deep understanding; its intelligence is artificial and relies on shallow connections. To illustrate this point, I recall an experience during a prototyping course where a design professor commented on my code, saying

“You don’t know what you’re doing; you’re just taking a chance.”

He was correct — I lacked a deep understanding of the structure and context behind the code. However, the prototype still worked. Similarly, large language models like GPT-4 can create something without truly comprehending the meaning or context behind them, relying instead on superficial associations.

Deep understanding is a complex concept, but it can be understood through abilities such as connecting the dots, understanding context, transferring knowledge, and self-awareness.

Connecting the dots

“I don’t believe in conspiracies. I’m just connecting the dots.” Image credits: user joid75, https://imgur.com/t/knowledge/6K8GvyO

Connecting the dots is, in essence, the capacity to put together the pieces of an unknown puzzle and make it meaningful. It involves seeing how concepts, ideas, and theories connect and shape a unity. For example, in 1987, a group of Japanese scientists tried to understand the genetic code of E-coli bacteria in yogurt. However, it wasn't until 2012 when two biochemists named Jennifer Doudna and Emmanuelle Charpentier connected the dots and found that the genetic code in these bacteria could be reprogrammed, changing the genes in plants, animals, and humans - a discovery named CRISPR for which they won the Nobel Prize. Recently, Alphabet (Google) used CRISPR to wipe out an entire population of lab mosquitoes. Thus, connecting the dots is a mighty ability.

⊕ GPT-4canprocess and synthesize large chunks of information fast and precisely.

⊖ GPT-4can’tconnect the dots in a meaningful way informed by its personal experience, value, and goals.

💡 However, a designer can connect the dots using human-centered design methods, empathizing, gathering insights from context and user research, working in a diverse and inclusive team, and getting feedback from stakeholders.

Contextualization

This category refers to the ability to understand the context in which knowledge is situated. It involves seeing how social, cultural, historical, or other factors shape the meaning and significance of a particular piece of knowledge.

In Hofors, I grew up in an environment where everyone spoke their mind freely. When someone talked, it was common to talk over them, and sometimes my grandmother and her friends would speak at the same time in loud voices. Speaking simultaneously was the norm in other circles in my upbringing as well. Surprisingly, as I grew up, I realized that people in other places don’t appreciate that type of communication. However, I still feel comfortable with it, so whenever I return to Hofors, we still talk on top of each other.

Understanding social and cultural context leads to better communication between humans. For example, we can apply this learning in different contexts and be aware that the same knowledge and behavior may be perceived differently depending on the context. Humans develop a nuanced understanding of what influences knowledge in the real world.

⊕ GPT-4 can retrieve information about social, cultural, historical, or other factors that can shape the meaning and significance of knowledge.

⊖ GPT-4 can’t fully appreciate the nuances of context factors in how humans can.

💡 A designer can conduct extensive research and gain a deeper understanding of the context, ask experts, employ human-centered design methods, empathize with users, and receive stakeholder feedback.

Transferability

Testing the research of Melanie Mitchell, trying to train AI to think with analogies, asking, “If the string ‘abc’ changes to the string ‘abd,’ what does the string ‘ijk’ change to?” GPT-4, does not understand that the last letter shifted one letter ahead.

The correct answer is “ijl” if you can’t figure it out. However, since December until now, it seems like it has evolved a bit more in transferability. Asking a riddle: Adam’s mother has four sons: Beta, Cesar, David, and who is the fourth son?” will result in the answer Adam.

One clear example of the transferability of knowledge and skills is when you master the skill of playing guitar and understand music theory, you can transfer the knowledge and skills and play piano or even drums. Likewise, if you jam in a band, you can use that creativity and collaborative methods in the workplace. The same thing goes if you work in one place and know the culture you can quickly transfer that knowledge to another workplace. And the same goes for programmers, when they can learn to program in c+, they can use their coding skills in other languages.

https://en.wikipedia.org/wiki/File:Karate_kid.jpg

In the movie Karate Kid, Mr. Miyagi forces Daniel to wax many cars using the “Wax on, Wax off” technique. Initially, Daniel doesn’t understand why he’s being asked to do this, but later on, he realizes the method has helped him to win the Karate championship.

⊕ GPT-4can be trained to perform specific tasks or solve particular problems,

⊖ GPT-4lacksthe transferability of knowledge and skills that humans possess to apply knowledge and skills in new or unfamiliar situations from a subjective perspective

💡 A designer can encourage continuous learning, attend conferences and workshops, and discuss with peers. Collaborate between team members with different skills and expertise, and use a human-centered perspective.

Metacognition

René Descartes — “Cogito, ergo sum”

I think, therefore, I am (Descartes, 1637)

It is the subjective experience, and the real consciousness of self that ChatGPT is not capable of. Blake, an engineer at Google, was blowing the whistle at Google when he found a “ghost in the machine”. Google rebutted and claimed that LAMDA is only following user prompts, similar to what ChatGPT does.

👉Read about Blake’s encounter with a possibly sentient AI in his own words here.

So for GPT-4 and the likes to reach consciousness, they need to be able to stop, reflect and evaluate learnings, understand gaps, seek out new information and reflect on their learning process. Today, an LLM can not walk outdoors in a live, ongoing context and reflect on its own learning.

⊕ OpenAI mY develop algorithms that allow future GPT-5 to reflect on and adapt their learning strategies

⊖ GPT-4 is unaware of itself and cannot see itself from within and regulate its thinking processes.

💡 A designer can design AI services that reflect on and adapt their learning strategies to become aware, sentient, and a ghost in the shell. But is this something humans want? Recently, Microsoft released a paper saying that GPT-4 is almost AGI. So I asked GPT-4 to act as an LLM and asked if it considers itself “almost AGI”, and it did. Experts and leading voices within AI have written a list where they call on all AI labs to immediately pause for at least six months the training of AI systems more powerful than GPT-4.

Meaning-making: How we create meaning for ourselves and others

Creating meaning is a dynamic and ongoing process involving reflection, interpretation, and communication. It is a personal and subjective experience where we create our narrative and make sense of our place in the world to others. First, we connect with our experiences, values, and goals to create meaning. This personal understanding is then used to gain a meaningful understanding of a context, topic, or domain.

For me, my morning ritual gives me meaning. Making coffee, turning on the radio, making breakfast, checking the news, and taking a morning walk all hold significance because the rituals connect me to the world and society on a daily basis with purpose.

⊕ GPT-4 can generate text that is relevant and useful for specific purposes,

⊖ GPT-4 lacksthe personal experiences, values, and goals humans possess to create personal meaning and relevance from knowledge.

💡As a designer, I create meaning by matching the expectations and the needs of my design product or service. Meaning is made by doing user research and empathizing with the user’s context. Also, creating stories can captivate, resonate, and ultimately boost customer conversions.

Things ChatGPT can do related to a deep understanding

GPT-4 possesses functions capable of imitating and resembling deep understanding. Firstly, it can instantly process large amounts of information, allowing it to connect the dots artificially. Secondly, GPT-4 can artificially connect and contextualize social, cultural, and historical data, creating an artificial contextual description. However, an extensive prompt is required for this capability. Moreover, there is a possibility that future models can be trained to improve problem-solving and transferability. One potential approach is for OpenAI to enable GPT-4 to adapt its learning strategies to enhance its deep understanding capabilities.

GPT-4 lacks elements to be able to use deep understanding

Although GPT-4 has impressive features closer to deep understanding, it lacks the fundamental capabilities that make up for a deeper understanding.

1. GPT-4 cannot connect the dots aligned with real personal life.
2. GPT-4 has trouble understanding context and cannot fully sense or understand nuances and complexities in the same way as humans.
3. GPT-4 lacks the transferability of knowledge and skills that humans possess, making it difficult to apply knowledge and skills in new or unfamiliar situations from a subjective perspective.
4. GPT-4 is also incapable of metacognition, as it is unaware of itself and cannot see itself from within and regulate its thinking processes.

Lastly, GPT-4 does not have the personal experiences, values, and goals humans possess, making it challenging to create personal meaning and relevance from knowledge. While GPT-4 has impressive capabilities, it still has limitations that require consideration when utilizing its deep understanding abilities.

💡From the designer’s perspective —always get your hands dirty

https://pixabay.com/photos/people-woman-girl-hand-dirty-palm-2597941/

Again — get your hands dirty! A designer should work less on quantifications and more on empathizing. It involves working more with human-centered, more-than-human, or non-human design methods to connect the dots by empathizing with and gathering insights from the context and users with research, iterative experimentation, and having a stakeholder dialogue.

It is essential to get in context by practicing continuous learning, attending conferences and workshops, and discussing and collaborating with a diverse team. Additionally, start working on your AI-related initiatives and influence the development by sharing your work. For example, as a designer, you might become empowered by using AI as an assistant, but be aware that your prompts and actions could serve as training data for future models.

Understanding the purpose of AI development is critical since many projects are built on the idea of technological singularity (an AI to rule us all). However, that notion is ethically problematic because what could that lead to in the end?

Conclusion

In conclusion, while there has been progress, and potential with the development of GPT-4 and other models, deep understanding and empathy for the environment and other people are still missing where designers are crucial. GPT-4 and similar models cannot connect the dots based on personal experience, see small nuances in context changes, or even transfer knowledge like “wax on, wax off.” 😉 Additionally, these models cannot understand the importance of human-centered methods, continuous learning, collaboration across borders, or genuine interest and commitment to stand out among other designers and AI models. While AI can act as a powerful and agile assistant, it cannot yet replace a human designer’s critical and empathetic touch.

According to GPT-4: “human designers will remain a critical component of the design ecosystem for the foreseeable future.”
“The future of design will likely involve a collaborative approach where humans and machines work together, leveraging each other’s strengths to create innovative and meaningful designs.”

By the way, it is highly problematic that GPT-4 considers designers a critical component in a system. I don’t feel I am a component yet.

More on the topic

Moriarty, J. (2023). Design in the age of ChatGPT. UX Collective. Retrieved from https://uxdesign.cc/design-in-the-age-of-chatgpt-3c80e6fc8cf7

Wodecki, B. (2023, March 23). Goldman Sachs: Generative AI could replace 300 million jobs. AI Business. https://aibusiness.com/nlp/goldman-sachs-generative-ai-could-replace-300-million-jobs

Yuce Gun, O. (2023). When and how will your work be replaced by AI? UX Collective. Retrieved from https://medium.com/user-experience-design-1/precisely-when-and-how-you-will-be-replaced-by-ai-b4554da44391

References

Descartes, R. (1637). Discourse on the Method of Rightly Conducting the Reason and Seeking Truth in the Sciences. Retrieved from https://www.gutenberg.org/files/59/59-h/59-h.htm

Hannibal046. (2022). Awesome-LLM. GitHub. https://github.com/Hannibal046/Awesome-LLM

Future of Life Institute. (2023). A Letter Calling for the Pause of Giant AI Experiments. https://futureoflife.org/open-letter/pause-giant-ai-experiments/

Mitchell, M. (2021, July 14). Melanie Mitchell Trains AI to Think With Analogies. Quanta Magazine. https://www.quantamagazine.org/melanie-mitchell-trains-ai-to-think-with-analogies-20210714/

Mitchell, M. (n.d.). Postdoc project description. Retrieved from https://melaniemitchell.me/PostdocProjectDescription.pdf

OpenAI. (n.d.). GPT-4. Retrieved March 31, 2023, from https://openai.com/research/gpt-4/