A review finds that safety, fairness and human control require careful design alongside advances in AI capabilities.
The review examines how people and large AI systems can work together in problem-solving and decision-making. It considers human-guided development, collaborative design, ethical and governance frameworks, and applications in high-stakes areas.
The authors conclude that successful human-AI systems depend on deliberate design rather than model strength alone. They also outline continuing challenges involving safety, fairness and human control, and identify areas for future research.
The survey finds that stronger large AI models are not by themselves enough to create reliable human-AI partnerships. The review highlights four areas as important: involving people in model development, designing systems around collaboration, establishing ethical and governance frameworks, and studying use in high-stakes domains. It presents safety, fairness and control as persistent challenges.
Large AI systems are increasingly used for tasks involving problem-solving and decisions. The review suggests that their social benefits depend not only on what the systems can do, but also on how people guide, use and govern them. This places human-centered design alongside technical capability as a central concern for building partnerships intended to be reliable and trustworthy.
This is a survey article that synthesizes existing research rather than reporting a new experiment or a direct test of one human-AI system. The abstract does not provide details about the review’s search methods, the number or selection of studies, or how the authors assessed the evidence. Its conclusions identify broad research priorities and challenges, but do not establish that any particular design or governance approach will work in every setting.
ACM Transactions on Intelligent Systems and Technology · 2026 · DOI: 10.1145/3841472
Authors: Vanshika Vats, Marzia Binta Nizam, Minghao Liu, Ziyuan Wang, Richard C. Ho, Mohnish Sai Prasad, Vincent Titterton, Sai Venkat Malreddy, Riya Aggarwal, Yanwen Xu, Lei Ding, Jay Mehta, Nathan Grinnell, Liu Li, Sijia Zhong, Devanathan Nallur Gandamani, Xinyi Tang, Rohan Ghosalkar, Celeste Shen, Rachel Shen, Nafisa Hussain, Kesav Ravichandran, James Davis
Institutions: University of California, Santa Cruz
Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.
AI vulnerability detectors often perform less well on software projects outside their training data. A study finds that cleaner, more varied training data and an encoder-based design improved detection, including a reported 6.8% gain in recall on the BigVul benchmark.
Researchers developed a quantum method for the maximum cut problem and proved a stronger worst-case guarantee than a three-round version of a widely used approach. On IBM quantum devices with 72 and 148 qubits, the method exceeded a known classical hardness threshold, while the comparison method did not do so in the 148-qubit test.
