DEBRA MATHEWS
Associate director for research and programs

AI for speech and language

AI-driven speech and language technologies are transforming fields ranging from medicine to cybersecurity. Berrak Sisman, who leads the Speech & Machine Learning (SMILE) Lab at the Johns Hopkins Center for Language and Speech Processing, highlighted three key trends in how AI is reshaping speech and language processing. Her team at Hopkins develops state-of-the-art AI models to improve how machines understand and express themselves through human language.

Expressive speech generation speech generation and deepfakes STEM from the same powerful technologies

The tools that bring voices to life in healthcare, education, or accessibility can also be exploited to create disinformation. Studying these technologies together, both their capabilities and vulnerabilities, is key to developing effective deepfake detection and building more secure, trustworthy systems.

Emotional intelligence is the missing layer in today’s AI

As AI systems get better at generating speech that sounds human, the next frontier is how they say things: tone, empathy, and emotion. In an AI-driven world, emotional intelligence is essential for trust, safety, and effective communication.

Large language models are changing what machines can understand and say

LLMs are already transforming education, healthcare, and national security. They’re not just chatbots; they’re becoming the backbone of how machines reason, generate, and speak.

BERRAK SISMAN
Assistant professor

AI for computer vision

Computer vision focuses on enabling machines to interpret and understand visual data, such as images and videos. It has broad applications across sectors, including autonomous driving, assistive technologies, homeland security, medicine, and public health. The overarching goal is to automate visual tasks that the human visual system performs naturally.

There are a variety of impactful real-world applications, from commerce to health care, including:

• Identifying missing children
• Aiding people with low vision
• Enabling drones and robots to “see”
• Facial phenotyping for stroke monitoring and evaluation
• Mosquito identification to slow the spread of mosquito-borne disease
• Traffic flow estimations
• Self-driving cars, which embody the integration of vision with robotics and autonomous systems

AI for computer vision raises critical challenges, too.

• Data hunger. Heavy reliance on large datasets
• Black-box nature. The lack of transparency in model decision-making
• Bias. Sensitivity to domain shifts, concerns about privacy, and fairness. A common question Rama Chellappa, chief scientist at the Johns Hopkins Institute for Assured Autonomy, said he asks is, “Will it work everywhere, and will it work for everyone?”
• Security. Vulnerability to adversarial attacks

These limitations underscore the importance of addressing technical and generalization issues for real-world deployment, Chellappa said.

RAMA CHELLAPPA
Bloomberg Distinguished Professor

AI for national security

AI is fundamentally changing war both on and off the battlefield. For instance, it’s enhancing drones, making them increasingly autonomous. It’s also impacting the processes, such as the discovery of new materials and code generation, that help develop new military tech.

The broad accessibility of AI produces asymmetric threats. Entities no longer need major physical infrastructure to pose a threat anymore. Now, threadbare militaries can overwhelm stronger forces by leveraging AI capabilities.

“Ukraine, in particular, is giving us a glimpse into the vulnerability of much of the heavy, large equipment that we have, and also an indication of how we might deploy some of our capabilities to be more effective in the future,” said James Bellingham, Bloomberg Distinguished Professor of exploration robotics and executive director of the Johns Hopkins Institute for Assured Autonomy.

Other ways AI can impact national security include:

• The development of and protection against cybersecurity threats
• Improving the work of intelligence agencies by providing more situational awareness to manage conflicts
• The ability to anticipate the consequences of attacks
• Better precision and control of attacks

Dimensions of AI safety and assurance for national security

Many experts note that the same elements that enable AI to support national security can also generate many risks for it, including unclear attribution, better spoofing and deception, and outpacing the human control loop.

One of the top risks is the uneven nature of AI performance, which Bellingham said prompts a need for research on the science of AI and data with a particular focus on providing performance guarantees.

“We need systems-engineering processes with good guidelines for how you construct your data sets, how you validate your data sets, and then when you construct your AI, how you validate your AI,” he said, adding that experts should consider the application domain—land, sea, air, or space, for example—when designing this process.