Top Predictions and Trends for Generative AI in 2023

Matt White
11 min readJan 7, 2023

--

Stable Diffusion generated image. (prompt: cyberpunk style image of the future with robots and flying cars)

2023 will be a banner year for the field of generative AI. In the latter half of 2022 the world took notice of the power and potential as well as the pitfalls of generative AI.

Here are a list of my predictions for 2023 which include improvements, innovations, trends, new paradigms, issues, adoption and applications of generative AI.

Image Generation

  • Significant improvements will be made in AI image creation. Currently models struggle with generating text within images and handling multiple concepts (ie. multiple central objects like people, animals, etc.) Newer modeling techniques, including improvements in text-to-image modeling will yield results where faces are fully composed and photorealistic, body compositions are accurate and viable, and text-to-image will produce coherent text outputs within an image that yield accurate results.
  • New versions of Stable Diffusion, Imagen and Dall-E will be released that will outperform existing models and we are likely to see OpenAI or Google release a model with 15–20 billion parameters that can produce photorealistic images that are highly plausible.

Video Generation

  • We will see longer video segments produced from text prompts in 2023 that are more photo-realistic and temporarily consistent. I do not foresee an entire movie or television program being created with generative AI, but significant advances will move the dial forward and resolve the number of issues that are currently suffered by the state-of-the-art and generative AI tools will be used to create various aspects of movie production including script-writing and dialogs, special effects, voice synthesis, music and sound effects.
  • Text-to-video generation will also see idea-to-script begin to take off, which will allow a user to describe the type of video they want to produce with time-stepped prompts which will extend generative videos beyond just a single prompt that generates a short 10 second video of the same object and scene to creating multiple objects, scenes, and actions in a single synthetic video.

Audio Generation

  • Synthetic audio generation will allow artists to produce songs and albums more rapidly with less overhead and people involved in the production of music. We may hear the first synthetically generated song that will make Billboard’s top 100 hits for 2023, however we may not even know it. I do think that the music industry is in for a big disruption in 2023.
  • Those who don’t possess the talent and skills required to produce music with digital audio workstations, have the ability to play instruments or are capable of singing will be able to produce studio quality songs quickly and with very low effort.
  • Voice synthesis will improve to the point where it will be indistinguishable to the human ear what is real and what is synthetic. Work will be done to enhance IVR systems with large language models, natural language processing and voice synthesis to provide a better customer experience.

3D Generation

  • 3D asset generation is still in its infancy, this is an active area of research for Berkeley Synthetic. However, 2023 will bring innovations that will see photorealistic and physically accurate 3D models generated from text and voice prompts. These models will be high fidelity and usable within games and simulations without the need for manual post processing to clean up meshes and textures.
  • Instead of applications creating final results based on text prompts, iterative and editable solutions will come to market to allow for real-time modifications of the properties of 3D assets including textures, materials, rigging and other attributes.
  • Military simulations will depend more on more on generative AI to create scenarios and produce assets that are physically accurate and photorealistic so that less time is spent developing environments and more time used executing simulations for training, testing and war games.

Hyper-personalization

  • A new area will begin to emerge within the different segments of media which is hyper-personalization. Hyper-personalization is the ability to have content generated to cater to your specific preferences and in a unique way. Unlike personalization which curates previously generated content to a predefined set of classes, hyper-personalized content is auto generated by AI to suit your specific individual profile. An example of this is the shift from search engines like google, which curate results to large language models like ChatGPT which produce results in the form of very specific answers.
  • The area of hyper-personalization will be embraced by marketers who seek to attract and engage with consumers in ways that were not previously available with mass appeal and targeted marketing campaigns using static content. Although there will be no “Minority Report” moment in 2023, there will be a movement towards this degree of personalization in marketing that will be facilitated by generative AI.
  • Media in its many forms will also move towards hyper-personalization, from auto-generated images and headlines in online news stories that are more attractive and valuable to individual users to music and movies that will one day be partially or fully dynamically generated to be the most aesthetically pleasing and engaging to individuals. This will be an area of major development in the coming decade for which numerous startups will crop up.

From Search to Chat

  • ChatGPT has taken off and although the platform is still in public Beta, it has already captured the attention of the public, surpassing 1 million users in less than a week. With its dialog-based interaction it is like a chatbot on steroids, and consumers are coming to appreciate the quality and format of the results although the platform cannot yet be trusted to produce accurate results. ChatGPT represents a formidable challenger to Google’s search, in an industry that has seen little innovation over the past two decades. I expect that consumers will prefer this type of interaction and we will see dialog-based large language models that are active learners become more pervasive. Although OpenAI plans to charge for use of ChatGPT we will see other tech companies release free versions and these large language models will back existing services like the more primitive customer service chatbots used today and personal assistants like Siri and Alexa.

Multi-Modality

  • As AI researchers and companies like OpenAI continue to push towards creating Artificial General Intelligence (AGI), we will see more and more multi-modal models developed. Multi-modal models can effectively operate across different applications, for instance being able to generate both text and images. I expect this will be a strong area of research focus and investment in 2023.

Businesses and Startups

  • With VCs actively looking for the next unicorns in emerging tech, there will be significant investment in generative AI startups, including those that perhaps don’t have the strongest of business models. There is likely to be a lot of overoptimistic VCs willing to bet the farm on generative AI startups, and many of them will fail to live up to the hype causing investors to reel in in the latter part of 2023.

Big Tech and Responsible AI

  • In the world of AI research, big tech dominates. The research labs of Google Brain, DeepMind, Nvidia, Adobe, Meta and OpenAI (backed by Microsoft) produce the bulk of the innovations in generative AI working with PhD researchers from leading universities like Stanford, UC Berkeley, and MIT. This relationship gives big tech direct access to the best talent in the industry and has created an AI monopoly, where big tech with 100 million+ budgets, infrastructure and talent can produce innovations that are not attainable for startups and small companies. Generative AI startups are being forced to pivot or shut down before they can get products and services to market because big tech can move faster. I see this becoming a greater area of concern for fairness and competition in the generative AI space.
  • I foresee the issue of “accessible AI” becoming a more substantive topic of discussion as the AI research community adopts principles of “Responsible AI” which are built on the pillars of transparency & explainability, fairness & human-centeredness, security & privacy, and accountability. With wider adoption of generative AI and most applications and papers not providing datasets, models, weights and biases, consumers and advocates of privacy, safety, diversity and inclusion and other communities will raise the alarm regarding the lack of explainability and the bias underlying generative AI applications.

Social Media

  • 2023 will see the “rise of the synths”. Influencers will begin using generative AI tools to improve their appearance, to place them with people and locations they have never been and to engage with their followers using generated text content in their posts.
  • I predict that at least a few social media influencers will be fully synthetic. Their entire image and actions will be entirely generated using synthetic images, synthetic video, synthetic voices, and synthetic dialog. We have seen some very preliminary attempts at this thus far, but with new models yielding photorealistic and consistent results as well as the increased access to the technology and proliferation of platforms, we will see more and more synthetic images, videos, audio, and text in social media. I predict that by 2035 most online influencers will in fact be fully synthetic and our synthetic avatars will be extensions of ourselves as we move into the Metaverse.

Synthetic Data Generation

  • A Gartner study suggests that 60% of all data used in the development of AI will be synthetic rather than real by 2024. I believe that this is an overestimate. Synthetic data will no doubt become more pervasive in 2023 but if we have learned anything about the current rate of AI adoption with 23% of companies having successfully integrated AI into their processes, products, and services, it is unlikely we will see such a high adoption of synthetic data. I believe that number will be more in the area of 40% by the end of 2023 and that is at the high end of my estimates. Synthetic data generation requires sufficient real data to train models on, and that continues to be a challenge for many organizations as they try to wrangle data and privacy advocates and consumers push back on the collection and use of their data. Several startups have popped up recently to handle the generation of synthetic data, and I expect that trend to continue.
  • With many organizations treating synthetic data as a derivative of real data and not owned by the original sources of the true source data, we will see synthetic data exchanges established to help provide organizations with pre-generated datasets for their deep learning applications. However, it is possible data privacy advocates will object to the marketing of personally derived synthetic data and expect that is should be regarded under the same privacy laws as personal data.

Software Development

  • Generative AI models trained on code will improve in 2023, Copilot and Codex will likely include a reinforcement learning to be able to be active learners capable of improving upon their accuracy in producing code by taking real-time feedback from users.
  • Generative AI models will be used for performing code security audits using reinforcement learning to test simulated use cases.
  • Testing tools will also use generative AI to simulate user interaction to improve performance of backend and frontend software services and applications.

Medical Research & Drug Discovery

  • Google Brain’s AlphaFold has shown that generative AI can make significant contributions in the field of medical research, in the case of AlphaFold this was achieved by predicting and creating new plausible protein structures.
  • We will see more adoption of generative AI in scientific and medical research, and those methods will ultimately reduce the need for full animal and human trials using simulations and synthetic process emulation backed by deep learning models.
  • In the field of drug discovery, the use of AI will extend beyond just predicting candidate drugs to generating potential solutions which can then be tested in computer simulations to test potential efficacy.
  • Synthetic data generation will facilitate improved classification model accuracy which will assist medical professionals in making more accurate diagnosis of patient conditions in areas like medical imaging.
  • In the not-too-distant future generative AI will be used to tailor treatments to individuals to ensure maximum efficacy of medications while minimizing potential side effects. However, this area of work will need to evolve substantially over the next decade to be viable.

Security & Privacy

  • Generative AI will be used more and more by adversaries to attempt new methods to penetrate networks, especially by state-sponsored actors. New startups and products will be developed to counter these methods using the same generative technology.
  • Deep Fakes will continue to be an issue, currently the same methods used to generate them are used to detect them. However, there will come a point, the “Matrix moment” for digital media where we cannot tell the difference between synthetically generated content and real content. I suspect that this could happen in 2023.
  • Scammers will begin using synthetic images and real-time deep fakes to catfish and con unsuspecting victims.
  • The proliferation of synthetic Illegal and unethical content, whether images, video, text or 3D will become a problem. Synthetic child pornography and hate media are two areas that will be of serious concern. AI bots (autonomous agents) could be let loose into the wild with the directive to use generative AI to create hateful posts and comments or to provoke political instability.
  • The porn industry will widely adopt generative AI to produce new content, personalized content (ie. users appearing in synthetic videos, not just transposed faces like today’s deepfakes) and to reduce the overhead of producing real pornographic content. Webcam models will become synthetic, and their interactions controlled by autonomous agents that can act and converse based on input from the user. Unfortunately, this technology will be mis-used to create synthetic revenge porn or to defame others.

Authentic Content & Synthetic Detection

  • Watermarking will become a very important factor for services that leverage generative AI, to validate whether content is AI generated or original. This is important for academics whose students can produce reports, images and software that is being passed off as their original work. However, watermarks will be easily defeated using generative AI itself. This will be a very difficult area to address with new methods being defeated by adversaries.
  • The question of attribution and citations is a big question mark when it comes to generative AI. Content producers will likely pass off materials as being their own original thoughts or creations, but they will in fact be produced by a generative AI model.
  • Detection of generative AI works will be an area of growth for startups, but detection rates will decrease as quality of content increases.
  • Distributed ledger technologies like blockchain will be looked at as a method of tracking authenticity of human-produced content and works. People will actively seek out verifiable authentic human created content and avoid the huge amount of AI generated content that will proliferate throughout the Internet. Human-only communities will develop to protect their works from becoming training data for use by generative AI models.
  • Companies that use generative AI models will be forced to create opt-out mechanisms to allow artists, authors, and creators from having their works used to train a company’s models and having derivative works appear in AI produced content.

Business Models and Integration

  • Generative AI will go from being a stand-alone novelty to being integrated into existing tools and processes.
  • Business leaders will take generative AI seriously as a way to reduce overhead, increase productivity and to differentiate their products and services, however they will struggle with adoption.

--

--

Matt White

AI Researcher | Educator | Strategist | Author | Consultant | Founder | Linux Foundation, PyTorch Foundation, Generative AI Commons, UC Berkeley