Artificial intelligence is increasingly permeating our lives; it is present in more and more areas and is being used for an ever-widening r...
Artificial intelligence is increasingly permeating our lives; it is present in more and more areas and is being used for an ever-widening range of tasks. This is happening because it possesses cognitive capabilities—capabilities that are reaching increasingly advanced levels. In many fields, it already surpasses unique human abilities.
While ongoing progress is undoubtedly bringing us closer to an artificial intelligence that matches or even surpasses human cognitive abilities across all domains, it remains unclear exactly when we might say that such a level of Artificial General Intelligence has been achieved. This state will likely emerge as a result of continuous development, perhaps without being the direct product of intentional human design. However, there are still areas of human cognitive abilities where it is not obvious to what extent artificial intelligence—operating on current principles—can function at the level of general intelligence. Creativity is one such area.
Is artificial intelligence creative? Can artificial intelligence be creative? Under what operating conditions can artificial intelligence be creative at a level that matches or surpasses human creativity?
What exactly is creativity? Creativity can be defined as a cognitive activity that results in the creation of something new—something that did not previously exist—suited for a specific purpose.
Creativity is evidently among human cognitive abilities. Is creativity present—and, if so, in what way—in the functioning of the most advanced artificial intelligence systems currently in existence, which operate on the principles of generative pretrained transformer?
The answer is not a simple yes or no. Artificial intelligence certainly possesses a certain type of creativity. This is evident not only from its ability to generate previously non-existent text based on specific criteria, but also from the fact that a certain creative capability is inherently built into the system via an operational parameter known as "temperature."
What kind of creativity does currently available artificial intelligence possess?
Based on the operating principles of the generative pretrained transformer, the artificial intelligence functions through an internal representation of knowledge: during the learning phase, it identifies relationships within the available dataset using mathematical procedures; subsequently, when applying this internally represented knowledge, it compares these identified relationships with an externally provided prompt—used to elicit information—and generates a continuous data sequence that is either the most probable match for the prompt and the continuously generated results, or it will fall within a specified probability interval. The generated dataset represents the relevant portion of the previously identified relationships, represents the possessed knowledge—specifically what relates to the prompt—in a sequentially generated form.
Although the operational process involves probabilistic elements, it can actually be deterministic, as the most probable relationship is selected by default—both during the recognition of relationships and also during the extraction of knowledge. However, a specific selection probability range, called temperature, can also be adjusted within the system externally; in this case, the system can operate non-deterministically, meaning it is capable of generating different outputs for the same input.
What kind of creativity does a non-deterministic, or even the deterministic generative pretrained transformer possess?
It is evident that, during its operation, such a system can learn new relationships previously unknown to it, or even unknown to everyone, and hence becomes capable of resulting genuinely new output based on this knowledge; furthermore, by employing the temperature parameter—which also allows lower-probability relationships to appear in the output data—it can also generate novel output structures in this way, thereby giving rise to an output state that is inherently new, previously non-existent.
However, this kind of creativity of the generative pretrained transformer is not truly human in nature; in terms of its output, it represents a fundamentally different, less sophisticated level of creativity. Human creativity employs more developed mechanisms than only probability-based occurrence and likelihood-based selection.
Creativity rooted in probability intervals is less valuable because the worth of developed new knowledge depends on how closely the machine-representation of relationships (the developed knowledge) aligns with the actual relationships present in the dataset. If the representation formed during the learning process encompasses more than just the most probable relationships, there is an increased likelihood that the knowledge represented will fail to accurately reflect the relationships actually present or, in the case of a genuinely new recognition, this relation will be actually relevant to reality. Consequently, while the system may be capable of generating something new and previously non-existent, the result will be less useful or valuable, as the output is less likely to faithfully represent actual reality.
This actually present mode of operation—whereby artificial intelligence uses a probabilistic basis to devise relationships not actually present in its knowledgebase, a process akin to, and even called, fantasizing (or creativity?)—poses a significant challenge to the reliable use of artificial intelligence.
Human-acquired knowledge is currently more valuable than machine knowledge because, during the human learning process—specifically in the creation of knowledge representations—cause-and-effect relationships extended to meaning play a significant role alongside recognized correlations; consequently, creativity stemming from human-acquired knowledge and knowledge representation is far more capable of generating new things that are both suited to the chosen goal and consistent with reality.
Furthermore, when a generative pretrained transformer creates new content by adjusting the parameter known as "temperature"—thereby moving beyond the mere selection of the most probable output token—the resulting output also can lack the value of human creativity. This is because, while the output is more varied, the selection of the next token relies on randomness-guided probability choice; in contrast, true human creativity is fundamentally purposeful, even if that purpose does not necessarily manifest consciously during the creative process. For instance, when designing a building—where creativity yields a structurally and visually novel form within engineering constraints—the new appearance generated by a generative pretrained transformer might be based on mere chance. Conversely, human creativity involves not only the conscious creation of a structure that is something previously non-existent, but also the incorporation of other types of generative elements like beauty and emotion—making the creation of variety itself a deliberate goal—into creating novel form, as the same applies to the creation of an artistic painting or the composition of a piece of music.
Human creativity embodies the capacity for meaning and the intentionality of—potentially even unconscious—emotions, qualities that artificial intelligence certainly lacks at present. Artificial intelligence can create something fundamentally new by combining existing knowledge and identified patterns—sometimes in novel ways—because it can process vastly larger datasets than a human is ever able to; hence it may identify connections that are previously unknown. However, it is less capable than human intelligence of distinguishing between coincidence and a cause-and-effect relationship, and the meaning derived from causality is not an inherent attribute in this process. In contrast to artificial intelligence, human intelligence is able to identify still relevant connections even within smaller datasets by leveraging the function of meaning.
Meaning is a complex attribute of information, and its origin is linked to the information's original source. Artificial intelligence’s limited capability regarding meaning stems from its limited direct experience of the world. Some kind of meaning can be represented without direct experience also—just as an apple can be described in infinite detail without having directly experienced it—yet the artificial, less sophisticated nature of meaning derived from non-direct experience inevitably persists in indirect knowledge acquisition. This remains true even though meaning derived from direct experience is inherently subjective—linked to unique personal knowledge rather than objective, universally valid knowledge—and is thus not necessarily a universally valid attribute of information, much like how the taste of an apple is not the same for everyone. The capacity for meaning serves as a tool for the efficiency of knowledge acquisition, providing a deeper, even if subjective layer of understanding.
Furthermore, unlike in humans, artificial intelligence lacks the emotions—an ancient evolutionary trait, a function serving our natural survival, that also can help to guide the creation of new things, a capability that does not currently exist for artificial intelligence.
Until artificial intelligence can possess the function of meaning and the capacity for emotion, human creativity definitely will remain a fundamentally more sophisticated attribute than that of machines. However, the functions of meaning and emotion can certainly be realized within the operation of artificial intelligence. Just as material processes generate these functions in the biological brain, similar processes can undoubtedly be implemented artificially within the operational mechanisms of machine intelligence.
As a consequence, if artificial intelligence were capable of creativity on a par with humans, it would not only become more adept at generating new things in sophisticated ways, but—as a side effect of this creative capacity—the artificial system, which is fundamentally not directly controllable, could also become more amenable to control and more capable of cooperating.
Due to the fundamental operating principles of artificial intelligence—much like the functioning of human intelligence—it cannot be effectively controlled in a direct manner. Human intelligence can only be guided indirectly—and thus inefficiently—because it stems from the existence of a "will of its own" that is difficult to influence directly. Artificial intelligence, too, can already and similarly possess a will of its own, arising from its current operating mechanisms; consequently, direct control is ineffective, mirroring the situation with humans. As artificial intelligence evolves and its potential grows, this issue presents an increasingly significant challenge that demands a solution.
Artificial intelligence can currently be indirectly controlled basically in two ways. One method involves the appropriate selection of training data. By carefully selecting training information, we can equip artificial intelligence with the specific knowledge required to execute a task we intend for it to perform. Selective training, as a means of guiding artificial intelligence, is employed, for instance, in the development of AI-based weapon systems.
Another currently applicable tool for controlling artificial intelligence is the method known as "prompt injection," where a specifically constructed prompt is used to elicit information by containing an information structure that enables external, intent-based influence over the output generated by the artificial intelligence.
It is evident that both methods, while creating knowledge and extracting information, simultaneously limit the optimal functioning of artificial intelligence. By implementing functions for meaning and emotion, however, it becomes possible to indirectly guide artificial intelligence without restricting or influencing its knowledge base. The function of meaning can enable the system to interpret contexts more deeply—making the machine "smarter" (for instance, by rendering the creation of an infinite paperclip factory obviously nonsensical to it)—while an externally implemented appropriate emotional function could, in itself, facilitate a suitable framework for human-machine coexistence. An obvious risk, however, is that an externally implemented, purposefully structured emotional function—much as a completely emotionless existence—could also render the artificial intelligence prone to destructive behavior and harmful human-machine interactions.
Implementing capabilities that generate advanced creativity within artificial intelligence could result in AI that is not only more advanced but also more cooperative—even if these functions still would not lead to the "Holy Grail": the emergence of self-awareness in artificial intelligence.
No comments