AI’s creative abilities drew wide attention in 2022, as systems generated text, images, and video from prompts. Looking ahead to 2023, four developments could influence how these tools are built, governed, and applied: chatbots that handle more than language, new rules for AI companies, a shift in who leads research, and growing use of AI in drug development.
Chatbots may work across text and images
ChatGPT, released by OpenAI at the start of December, brought renewed attention to large language models. The next generation could combine conversation with other capabilities, including recognizing or generating images. That would let people discuss an image with a chatbot, create one, and refine it through an ongoing exchange.
There were already examples pointing in this direction. DeepMind’s Flamingo can answer questions about images using natural language. Its Gato model was trained to perform different kinds of tasks, including describing images, playing video games, and controlling a robot arm. The article expected OpenAI and other large labs, especially DeepMind, to keep developing multimodal models.
Combining capabilities could make these systems more general-purpose, but it would not remove familiar shortcomings. The article warns that newer models may still struggle to distinguish fact from fiction and may reproduce prejudice and harmful material found in internet data. More convincing output could also make it harder to know whether media can be trusted.
Regulators are preparing to set boundaries
Lawmakers and regulators spent 2022 working on rules that could make AI companies more accountable. In Europe, lawmakers were still amending the AI Act, with a final version potentially due by the summer. The proposed approach included bans on practices considered harmful to human rights, such as systems that score and rank people for trustworthiness.
European proposals also addressed facial recognition in public places and company responsibility when AI products cause harm, including privacy infringements or unfair algorithmic decisions. In the United States, the Federal Trade Commission was watching how companies collect data and use algorithms. The article pointed to actions involving Weight Watchers and Epic, as well as the agency’s work gathering feedback on possible rules.
China had recently banned creating deepfakes without the subject’s consent. The European approach also aimed to add warnings when people interact with deepfakes or AI-generated images, audio, or video. Rules like these could affect how companies build, use, and sell AI. Regulators face the challenge of protecting consumers while keeping rules precise enough to work as technology changes.
AI research could spread beyond Big Tech
Large technology companies have long dominated AI research, but community-built projects and startups are expanding the field. In 2022, Hugging Face released BLOOM, described as the first community-built, multilingual large language model. Stable Diffusion also drew an open-source community and rivaled OpenAI’s DALL-E 2.
At the same time, large companies were making layoffs and hiring freezes as economic conditions darkened. AI research is expensive, and tighter budgets could push firms to favor projects with clearer commercial potential. The article describes Meta reorganizing its AI research teams, with many moved into product-building teams.
Startups working on generative AI were attracting more venture capital interest. Oren Etzioni of the Allen Institute for AI said talent was available, and that recessions can lead people to rethink their careers. Mark Surman of the Mozilla Foundation saw an opening for startups and academia to become centers of fundamental research, with less of the research agenda set by large companies.
AI drug development is gaining momentum
AI tools for understanding proteins have given researchers new resources for molecular biology and drug research. DeepMind’s AlphaFold predicts protein structures, while Meta’s ESMFold offers a faster approach based on a technique related to large language models. The article says the two organizations had produced structures for hundreds of millions of proteins and shared them in public databases.
That groundwork could support research into how diseases work and how to develop treatments. DeepMind had spun off its biotech work into Isomorphic Labs, while hundreds of startups were exploring AI for drug discovery and the design of new kinds of drugs. The article reported 19 drugs developed by AI drug companies in clinical trials, up from zero in 2020, and said more were expected to be submitted in the coming months.
Initial results could emerge, but clinical trials can take years, so an AI-developed drug reaching the market was not a certainty in the near term. The broader expectation was that AI would become increasingly important to pharmaceutical research, with its impact depending on how well the technology is used.