Generative artificial intelligence, often referred to as generative AI or GenAI, involves artificial intelligence that has the capability to produce various forms of content such as text, images, or other media, utilizing generative models. These models learn the intricate patterns and structures present in their input training data and subsequently generate novel data with akin characteristics.
During the early 2020s, significant advancements in transformer-based deep neural networks paved the way for a multitude of generative AI systems. Notably, these systems could comprehend and respond to natural language prompts. Examples of such systems encompass expansive language model chatbots like ChatGPT, Bing Chat, Bard, and LLaMA, as well as text-to-image artificial intelligence art systems such as Stable Diffusion, Midjourney, and DALL-E.
Generative AI finds application across diverse industries including art, writing, software development, product design, healthcare, finance, gaming, marketing, and fashion. The early 2020s witnessed a surge in investment directed towards generative AI, with major corporations like Microsoft, Google, and Baidu, alongside numerous smaller enterprises, actively involved in the development of generative AI models. However, apprehensions persist regarding the potential misuse of generative AI, encompassing concerns about cybercrime, fabrication of fake news, and the creation of deepfakes, all of which have the potential to deceive or manipulate individuals.
This blog aims to delve into the history, modalities, and various applications of Gen AI, discussing its impact on society, its role in job markets, and the ethical considerations surrounding it.
History of Generative Artificial Intelligence
The academic field of artificial intelligence was established during a research workshop at Dartmouth College in 1956 and has seen multiple waves of progress and optimism in the years that followed. From its inception, researchers in this domain have grappled with philosophical and ethical questions concerning the essence of the human mind and the implications of crafting artificial entities with intelligence akin to humans. These contemplations have been explored in myths, fiction, and philosophy since ancient times.
The notion of automated artistry can be traced back at least to the automatons of ancient Greek civilization, where inventors like Daedalus and Hero of Alexandria were credited with creating machines capable of writing text, producing sounds, and playing music. This tradition of creative automatons persisted through history, exemplified by Maillardet’s automaton crafted in the early 1800s.
Since the inception of AI in the 1950s, artists and researchers have leveraged artificial intelligence to generate artistic creations. As early as the early 1970s, Harold Cohen was generating and showcasing generative AI art through AARON, the computer program he devised for painting generation.
Within the realm of machine learning, statistical models, including generative models, are frequently utilized for data modelling and prediction. The late 2000s witnessed the advent of deep learning, propelling advancements and research in image classification, speech recognition, natural language processing, and other domains. During this period, neural networks were primarily trained as discriminative models due to the inherent challenges of generative modelling.
In 2014, significant breakthroughs like the variational autoencoder and generative adversarial network led to the development of practical deep neural networks capable of learning generative models for complex data, such as images. These pioneering deep generative models were the first to not only assign class labels to images but also generate complete images.
By 2017, the Transformer network facilitated progress in generative models, culminating in the inception of the first generative pre-trained transformer (GPT) in 2018. Subsequently, in 2019, GPT-2 showcased the ability to unsupervisedly generalize across various tasks as a Foundation model.
In 2021, the unveiling of DALL-E, a transformer-based pixel generative model, alongside Midjourney and Stable Diffusion, marked a turning point, ushering in practical, high-quality artificial intelligence art driven by natural language prompts.
In March 2023, GPT-4 was introduced, with a team from Microsoft Research suggesting that it could be perceived as an early, albeit incomplete, version of an artificial general intelligence (AGI) system.
Modalities of Generative AI
Generative AI systems are built by applying unsupervised or self-supervised machine-learning techniques to a given dataset. The functionalities of a generative AI system are contingent upon the nature and composition of the dataset being utilized.
Generative AI systems can be categorized into two main types: unimodal and multimodal. Unimodal systems exclusively process a single type of input, while multimodal systems possess the capacity to handle multiple types of input concurrently. For instance, OpenAI’s GPT-4 is an example of a multimodal generative AI system, as it is capable of accepting both text and image inputs. This multimodal capability broadens the scope of the AI’s understanding and creative output, allowing for a more comprehensive and diverse generation of responses and creations.
- Text Generation
Generative AI systems, honed through training on words or word tokens, constitute a significant facet of artificial intelligence advancements. Noteworthy examples within this domain encompass GPT-3, LaMDA, LLaMA, BLOOM, and GPT-4, among others (refer to the List of large language models for an extensive overview). These AI systems showcase prowess in natural language processing, machine translation, and natural language generation. Moreover, they serve as foundational models that underpin a spectrum of other AI-driven tasks, showcasing their versatility and applicative potential.
The training of these generative AI systems heavily relies on diverse datasets, including but not limited to BookCorpus and Wikipedia. These datasets furnish the models with the linguistic richness and contextual understanding essential for proficiently generating meaningful and contextually appropriate language. Through exposure to such varied textual data, these AI systems learn to mimic and innovate human-like expressions, making them invaluable tools for various applications in communication, creative content generation, and language-centric endeavours.
- Code Generation
Generative AI can also create code, ranging from simple scripts to complex algorithms. This can streamline software development, automate repetitive tasks, and enhance productivity in the tech industry. Furthermore, large language models possess the capability to be trained not only on natural language text but also on programming language text. This expanded training enables them to generate source code for the development of novel computer programs. Prominent instances of such models include OpenAI Codex.
- Image Generation
The creation of exceptional visual art stands out as a significant application of generative AI. Numerous artistic creations generated by these AI systems have garnered public accolades and acknowledgement. Models like DALL-E and BigGAN have pushed the boundaries of image generation. They can create high-quality, realistic images from textual descriptions, opening up avenues in art, design, and visual content creation.
Generative AI systems that have been trained on datasets containing images and accompanying text captions, such as Imagen, DALL-E, Midjourney, Adobe Firefly, and Stable Diffusion, among others (refer to Artificial intelligence art, Generative art, and Synthetic media for more), play a pivotal role in this domain. These systems are widely utilized for tasks like text-to-image generation and neural style transfer. Noteworthy datasets employed in this context include LAION-5B and others (see Datasets in computer vision).
- Music Generation
Gen AI can compose original music, mimicking various styles and genres. It’s a boon for musicians, producers, and the entertainment industry in general, automating the composition process and inspiring creativity. Generative AI systems like MusicLM and MusicGen are designed to learn from audio waveforms of existing recorded music and accompanying text annotations. This training enables them to produce novel musical samples based on textual descriptions. For instance, the AI can generate music samples that align with specific text descriptions, such as a soothing violin melody complemented by a distorted guitar riff.
- Video Generation
Advancements in video generation have led to the creation of deepfake technology, enabling the synthesis of realistic videos featuring individuals who never participated in the original footage. While this technology has raised ethical concerns, it has potential applications in filmmaking, special effects, and entertainment. Generative AI, when trained on annotated video data, has the ability to create video clips that maintain temporal coherence. Notable examples of this capability include Gen1 and Gen2 by RunwayML as well as Make-A-Video by Meta Platforms. These AI systems utilize annotated video data to generate video sequences that align with the provided annotations, resulting in a cohesive and synchronized visual output.
- Molecule Generation
In the field of drug discovery and chemistry, generative AI can design new molecules with desired properties. This can significantly accelerate the drug development process and potentially lead to the discovery of novel treatments. Generative AI systems can undergo training using sequences of amino acids or molecular representations, such as SMILES notation representing DNA or proteins. Noteworthy among these systems is AlphaFold, extensively utilized for protein structure prediction and drug discovery purposes. Datasets employed in training these AI systems encompass a range of biological datasets, contributing to their efficacy in predicting protein structures and advancing drug discovery efforts.
- Planning and Robotics
As of the early 1990s, generative AI planning systems had already achieved a significant level of maturity as a technology. These systems were notably proficient in generating crisis action plans tailored for military applications. Generative AI plays a vital role in planning and optimizing processes, helping robots make intelligent decisions and improving automation in various industries. This includes applications in logistics, manufacturing, and even healthcare.
Generative AI systems have played a pivotal role in robotic decision-making, notably demonstrated in the context of autonomous spacecraft prototypes during the 1990s. These systems have been effectively trained to understand and interpret the motions of robotic systems, enabling the generation of new trajectories for motion planning and navigation. For instance, UniPi, a creation of Google Research, utilizes specific prompts like “pick up blue bowl” or “wipe plate with yellow sponge” to control the movements of a robotic arm. Additionally, multimodal “vision-language-action” models like Google’s RT-2 showcase the capability to engage in rudimentary reasoning in response to user prompts and visual input. An illustrative example is the model’s ability to pick up a toy dinosaur upon receiving the prompt “pick up the extinct animal” in a setting filled with toy animals and other objects.
- Software and Hardware
Generative AI can assist in software and hardware design, optimizing performance and improving user experiences. This ranges from automated UI design to generating hardware layouts that maximise efficiency.
Generative AI models are at the core of various applications and products, showcasing their versatility and impact across different domains. These models power chatbot products like ChatGPT, programming tools such as GitHub Copilot, text-to-image products like Midjourney, and text-to-video products such as Runway Gen-2. Integration of generative AI features into established commercial products is also widespread, as seen in Microsoft Office, Google Photos, and Adobe Photoshop. Many generative AI models are available as open-source software, including Stable Diffusion and the LLaMA language model.
The scalability of generative AI is evident, with smaller models, ranging up to a few billion parameters, capable of running on smartphones, embedded devices, and personal computers. For instance, LLaMA-7B, a version with 7 billion parameters, can run on a Raspberry Pi 4, and a version of Stable Diffusion can operate on an iPhone 11. Larger models, in the tens of billions of parameters, can effectively run on laptop or desktop computers but may require accelerators like GPUs from Nvidia and AMD, or Neural Engine in Apple silicon products, for optimal speed. For example, the 65 billion parameter version of LLaMA can be configured to run on a desktop PC.
On a grander scale, language models boasting hundreds of billions of parameters, such as GPT-4 or PaLM, typically run on data centre computers equipped with arrays of GPUs (e.g., Nvidia’s H100) or specialized AI accelerator chips (e.g., Google’s TPU). These immense models are typically accessed as cloud services over the Internet.
In 2022, the United States imposed new export controls on advanced computing and semiconductors to China, impacting the export of GPU and AI accelerator chips utilized in generative AI. Consequently, specialized chips like the Nvidia A800 and the Biren Technology BR104 were developed to comply with the sanctions.
Impact on Society
The rise of Gen AI has brought about significant societal changes, both positive and negative. The advancement of generative AI has sparked significant concerns from various stakeholders, including governments, businesses, and individuals. These concerns have led to protests, legal actions, appeals for pausing AI experiments, and interventions by multiple governments. In a July 2023 briefing to the United Nations Security Council, Secretary-General António Guterres underscored the dual nature of generative AI, stating, “Generative AI has enormous potential for good and evil at scale.” He highlighted that AI has the potential to “turbocharge global development” and could contribute between $10 and $15 trillion to the global economy by 2030. However, he also cautioned about its malicious use, warning that it “could cause horrific levels of death and destruction, widespread trauma, and deep psychological damage on an unimaginable scale.” The Secretary-General’s remarks emphasize the critical importance of managing and regulating generative AI to maximize its positive impact while mitigating potential harm.
The automation potential of Gen AI raises concerns about job displacement. Routine tasks in various industries could be automated, potentially leading to job losses. However, it’s crucial to remember that new jobs often emerge as technology evolves, and retraining the workforce for more complex roles is essential.
Since the early stages of AI development, ethical debates have been ongoing, spearheaded by individuals like Joseph Weizenbaum, the creator of ELIZA. These discussions revolve around the ethical consideration of whether tasks that can be performed by computers should indeed be delegated to them. This deliberation stems from the fundamental disparities between computers and humans, as well as between quantitative calculations and qualitative, value-based judgments.
By April 2023, the implications of AI, particularly in image generation, became tangible. Reports indicated that AI-driven image generation had led to a substantial loss of jobs, with 70% of video game illustrator positions in China being affected. Furthermore, the advancements in generative AI played a role in the 2023 Hollywood labour disputes, an event that underscored the growing concern regarding AI’s impact on creative professions. Fran Drescher, the president of the Screen Actors Guild, expressed apprehension during the 2023 SAG-AFTRA strike, stating that “artificial intelligence poses an existential threat to creative professions”. These instances emphasize the pressing need for thoughtful consideration of the societal and economic implications of generative AI, particularly in the creative and labour sectors.
Deepfakes, a dark side of generative technology, involve creating highly convincing fake videos or audio recordings. Misuse of this technology for misinformation, propaganda, or damaging someone’s reputation is a significant concern. Deepfakes, a term derived from “deep learning” and “fake,” represent a form of AI-generated media. These creations involve the substitution of an individual’s appearance in an existing image or video with the likeness of someone else, achieved through artificial neural networks and deep learning techniques. The implications of deepfakes have gained significant attention and raised serious concerns due to their potential misuse.
One of the most alarming applications of deepfakes involves the creation of non-consensual explicit content, such as deepfake celebrity pornographic videos and revenge porn. Additionally, deepfakes are used to manipulate and spread misleading information, contributing to the dissemination of fake news, hoaxes, and financial fraud. These malicious uses have prompted substantial worry and prompted responses from both industry and government entities to address and mitigate the adverse impacts of deepfakes.
Efforts have been made to develop detection technologies and limit the malicious utilization of deepfakes. These include advancements in deepfake detection algorithms and the implementation of regulations and policies to curb the creation and distribution of harmful deepfake content. The proactive response from various stakeholders underlines the necessity of taking decisive action to combat the potential harm caused by deepfakes in the digital landscape.
Generative AI could be used for malicious purposes, such as creating convincing phishing emails or generating realistic disguises for identity theft. Addressing these risks and implementing robust security measures is imperative.
The ability of generative AI to produce highly convincing fake content has unfortunately been leveraged for malicious purposes in various forms of cybercrime, with phishing scams being a notable example. Deepfake technology, encompassing both video and audio manipulation, has been utilized to generate disinformation and perpetrate fraud. Shuman Ghosemajumder, a former Google fraud czar, has forewarned that although deepfake videos initially caused a sensation in the media, they would progressively become more prevalent and consequently more perilous. This progression underscores the potential for an escalation in the misuse of deepfake technology.
Cybercriminals have actively exploited generative AI’s capabilities by developing large language models tailored specifically for fraudulent activities. Examples of these models include WormGPT and FraudGPT. These AI models are engineered to assist in the planning and execution of fraudulent activities, highlighting a concerning dimension of the evolving landscape of cybercrime. The proliferation of such models amplifies the necessity for intensified efforts in cybersecurity, detection, and prevention to mitigate the impact of fraudulent applications of generative AI.
Regulation and Ethical Considerations
Given the potential misuse and ethical concerns, regulation is essential. Striking a balance between innovation and ethical use of Gen AI is a challenge that policymakers, technology developers, and society need to address collectively. The regulatory landscape for generative AI is evolving globally, reflecting the need to address the potential risks and ethical implications associated with its applications. Here’s a summary of recent developments in key regions:
European Union (EU):
The proposed Artificial Intelligence Act in the EU contains provisions focused on enhancing transparency and accountability in generative AI usage. This includes requirements to disclose copyrighted material used for training generative AI systems and to clearly label any AI-generated output as such. These measures aim to ensure transparency and protect intellectual property rights.
In the United States, a voluntary agreement was reached in July 2023 between major tech companies like OpenAI, Alphabet, and Meta, in collaboration with the White House. This agreement involves watermarking AI-generated content, aiming to enhance traceability and accountability for AI-generated outputs.
In China, the Cyberspace Administration introduced the Interim Measures for the Management of Generative AI Services, a regulatory framework to govern public-facing generative AI. This framework mandates the watermarking of generated images or videos and emphasizes regulations on training data quality, label quality, and restrictions on personal data collection. Additionally, it outlines that generative AI should align with “socialist core values,” highlighting a focus on cultural and ethical considerations.
These regulatory efforts demonstrate a shared recognition of the importance of regulating generative AI to ensure responsible usage, protect intellectual property, and uphold societal values. It’s essential to strike a balance between promoting innovation and mitigating potential misuse or harm associated with this rapidly advancing technology.
Indeed, Generative Artificial Intelligence (Gen AI) presents a complex and multifaceted landscape. As with any transformative technology, it carries both promising prospects and potential challenges. The key lies in recognizing the dual nature of Gen AI and adopting a strategic approach to maximize its positive impact while mitigating its negative consequences.On the one hand, Gen AI has the potential to disrupt job markets, requiring individuals to adapt and reskill. However, it also paves the way for new opportunities, particularly in fields like AI specialization, ethics, policy-making, and AI technology management. Collaboration between humans and machines will be essential, amplifying creativity and problem-solving capabilities.On the other hand, Gen AI’s potential for creativity, innovation, and problem-solving is vast and could revolutionize various sectors positively. Yet, misuse and societal disruptions are real risks, necessitating a responsible and ethical approach to its development and deployment. Proactive regulation and thoughtful consideration of the societal implications of Gen AI will be vital in steering it towards beneficial outcomes for humanity.
Ultimately, the future with Gen AI hinges on our ability to harness its capabilities judiciously, balancing innovation with responsible practices to usher in a future where AI serves as a powerful tool for human advancement and societal progress.