If 2023 was the year of AI fever, 2024 was characterized by a steady drumbeat of advancements as systems became more intelligent, faster, and less expensive to operate after ChatGPT was released in late 2022. In addition, AI started to communicate through voice and video and reason more deeply—trends that leaders and experts in the field predict will pick up speed. What to anticipate from AI in 2025 is as follows.
Better and more AI agents
Ray Kurzweil, an AI futurist, predicts that in 2025, we’ll start to witness a shift from chatbots and image generators and toward “agentic” systems that can operate independently to do tasks without just responding to queries. Anthropic gave its AI model Claude computer-based capabilities in October, including the capacity to click, scroll, and type, but this may be the beginning. Experts say agents will be able to do complicated jobs like writing software and making appointments.
According to Ahmad Al-Dahle, vice president of generative AI at Meta, these systems will continue to advance in sophistication. Although he sees a future in which AI agents serve as virtual coworkers, Jaime Sevilla, director of the nonprofit organization Epoch AI that forecasts AI, predicts that by 2025, AI agents will primarily be focused on their novelty. Professor Melanie Mitchell of the Santa Fe Institute cautions that, especially if agents have access to financial or personal data, their errors could have “big consequences.”
A national-security priority
Dan Hendrycks, head of the Center for AI Safety, predicts that governments will increasingly consider AI from a national security perspective: It’s how many of the big decisions about AI will be made. While Meta and Anthropic have strengthened their relationships with U.S. intelligence agencies by granting them access to their AI models, the U.S. has restricted China’s access to vital chips. According to Amandeep Singh Gill, the U.N. Secretary-General’s envoy on technology, political developments worldwide are directing us in the direction of ongoing competition, highlighting the necessity of maintaining “pockets of collaboration” between the United States and China.
Governance is rushing to catch up
As developers strive to create ever-more-intelligent systems, governments worldwide are vying to control them. The EU takes the lead with its AI Act. According to Markus Anderljung of the Centre for the Governance of AI, the E.U.’s Code of Practice, which is expected to be finalized by April and go into effect in August, is one of the first laws aimed at frontier AI developers. Many of the requirements will probably affect how businesses operate globally unless they choose to adopt different strategies in different markets. In the United States, where over 100 legislation have been introduced in Congress, Anderljung says that “very little will happen” at the federal level this year, though states might take separate action.
Facing the investment test
Humane Intelligence CEO Rumman Chowdhury states that the upcoming year “will be a year of reckoning.” Now that billions have been invested, businesses must demonstrate value to customers. That usefulness appears to be evident in the healthcare industry; for instance, the FDA is anticipated to approve more AI diagnostic tools, and AI may also be helpful in identifying and tracking the long-term effects of various drugs. However, in other places, the need to prove returns could lead to issues. The Global South may be subjected to the imposition of incorrect models due to the pressure to recoup all of these investments, according to AI policy expert Jai Vipra, who notes that these markets are not as closely monitored as those in the West. She cites the automation of already exploitative jobs, such as call center jobs, as a cause for concern in India.
AI-powered video becomes widely used
Both OpenAI and Google published outstanding video models in December. Access delays marred OpenAI’s Sora launch, while Google’s Veo 2 was made available to a limited number of customers. As developers figure out how to make video-generation tools more affordable to use, Sevilla anticipates that they will become more widely available. Al-Dahle of Meta envisions a not-too-distant future in which systems analyze video from smart glasses to provide real-time assistance across many tasks, including mending a bike. He expects that video will also become a crucial input for AI.
