Research News

In the race to make AI models not just reason better but respond faster, latency — the delay before an answer appears — is often treated as a purely technical constraint, something to minimize and move past. But how is this relentless push for speed actually impacting the people using these systems every day?

There is a rich body of work in human-computer interaction linking faster response times to better usability. But AI models are fundamentally different from the deterministic systems that previous research was built on. When you wait for a file to download or a page to load, the outcome is fixed and predictable. AI models are probabilistic — you cannot anticipate the precise response. Their conversational interface means users naturally read human social cues into the interaction. A pause might be read as the AI "thinking," for instance. Users are increasingly asked to choose between faster models and slower, deeper-reasoning ones, without guidance on what that choice actually means for their experience.

A recent study presented at CHI’26 explored how response timing shapes the way people use and evaluate AI systems. Felicia Fang-Yi Tan and Technology Management and Innovation Professor Oded Nov recruited 240 participants and asked them to complete common knowledge work tasks using a chatbot. Some tasks focused on creation, such as brainstorming ideas or drafting text. Others centered on advice, like evaluating decisions or offering recommendations. Crucially, the system was engineered to respond at different speeds. Some participants received answers after just two seconds, while others waited nine or even twenty seconds.

The results challenge a long standing assumption in human-computer interaction that faster is always better.

“People assume faster AI is better, but our findings show that timing actually shapes how intelligence is perceived,” says Tan. “A short pause can signal care and deliberation, making the same response feel more thoughtful and useful, even when nothing about the underlying AI model has changed.”

Surprisingly, how quickly the AI responded did not significantly change how people behaved (e.g., frequency of prompting, copy-pasting). Participants prompted just as much and interacted with the system in broadly similar ways regardless of whether they waited two seconds or twenty. Instead, behavior depended more on the type of task. Participants attempting creation tasks (which involve producing new content such as writing) prompted more back and forth, with users refining and iterating on ideas. Advice tasks (which involve providing guidance, critique, or evaluation) led to fewer, more focused exchanges.

Where timing did matter was in perception. Participants who received two-second responses consistently rated the AI’s answers as less thoughtful and less useful. In contrast, those who experienced longer delays tended to view the same kinds of responses more favorably. Many interpreted the pause as a sign that the system was “thinking,” attributing greater care and deliberation to its output.

This effect highlights a subtle but powerful feature of human psychology. In everyday conversation, pauses carry meaning. A quick reply can feel impulsive, while a measured delay suggests reflection. People appear to apply these same social expectations to machines, even when they know they are interacting with software.

The implications extend beyond user experience. Given that latency is an inherent feature of today's AI models, perhaps the more productive question is not how to eliminate it, but what it can be designed to do. Positive friction refers to intentional slowdowns designed to promote cognitive benefits such as reflection. Rather than treating every millisecond of waiting as waste, designers might ask: what can this pause do?

The study also surfaces important ethical considerations. If people equate longer response times with higher quality, they may place undue trust in slower systems, regardless of whether the output is actually better. This raises ethical questions about whether AI systems should be designed to manage timing in ways that shape user perception. And if so, whether users should be informed when they are.

Hydrogen is often described as the fuel of the future — a clean, energy-dense way to store renewable power and decarbonize industries from steelmaking to shipping. But inside the devices that produce it, a surprisingly small and familiar phenomenon is getting in the way: bubbles.

In water electrolysis, electricity splits water into hydrogen and oxygen gases. Those gases naturally form bubbles on the surfaces of electrodes. For decades, researchers have focused on improving catalysts and materials to make this process more efficient. Yet a new paper by postdoctoral researcher Darjan Podbevšek and Sustainable Engineering Initiative Director and Associate Professor of Chemical and Biomolecular Engineering Miguel A. Modestino published in the journal Joule argues that the real bottleneck may be far more mundane.

“Bubble dynamics represent a largely overlooked bottleneck that can account for significant efficiency losses.” says Podbevšek.

At first glance, bubbles might seem harmless, even expected. But their presence sets off a cascade of problems. As bubbles stick to electrode surfaces, they block the very sites where reactions are supposed to occur. They also disrupt the flow of charged particles in the liquid, increasing electrical resistance. As more bubbles accumulate, they can create uneven conditions across the electrode, further degrading performance.

In some cases, these effects are not trivial. The authors note that bubble-related losses can range from about 5 percent to as much as 25 percent of the total energy input, depending on operating conditions. In a technology where efficiency directly affects cost and scalability, that’s a major obstacle.

What makes the problem especially challenging is that bubbles behave in complex ways across multiple scales. At the smallest level, they begin as tiny nuclei, often forming at microscopic imperfections on the electrode surface. Their growth depends on subtle forces, including surface tension and gradients in temperature or concentration.

As bubbles grow and detach, they interact with one another, merging into larger bubbles or forming tiny bubble layers that blanket the electrode. These microscale “bubble carpets” can fundamentally alter electrode/electrolyte interactions and how reactants and products are transported in the system.

“It’s how and where they evolve, grow, detach, and interact that determines their impact,” the authors emphasize.

This multiscale complexity helps explain why the problem has been so difficult to tackle. Many experiments focus on single bubbles under highly controlled conditions, but real electrolyzers operate in messy, turbulent environments filled with countless interacting bubbles. Bridging that gap, the authors argue, is essential for making meaningful progress.

A typical electrolyzer is a tightly sealed, windowless box, operating at high pressure and temperatures (>30bar, >80°C), with highly alkaline or acidic conditions, making it difficult to observe bubbles directly. The authors suggest that it is at least part of the reason there are comparatively few bubble-related studies in the field.

The emerging view is that bubbles are not just a nuisance but an engineering challenge — one that can be managed, and perhaps even exploited. Researchers are exploring ways to design electrode surfaces that encourage bubbles to detach more quickly, preventing them from blocking reactions. Others are experimenting with flowing the liquid electrolyte more aggressively, using motion to sweep bubbles away.

Some of the most intriguing approaches involve changing how the electricity itself is applied. In “pulsed electrolysis,” the current is switched on and off rapidly. During the brief pauses, bubbles have time to dissipate, reducing their buildup and the associated energy losses. “Dynamic operation introduces additional control parameters,” the authors note, opening new possibilities for optimization.

Artificial intelligence is also beginning to play a role, helping scientists analyze complex bubble patterns and optimize operating conditions in ways that would be difficult by hand.

The implications extend far beyond the lab. Global demand for hydrogen is expected to grow dramatically in the coming decades, driven by efforts to cut carbon emissions. Making electrolysis more efficient — even by a small margin — could have an outsized impact on cost and energy use at scale.

“Connecting microscale bubble phenomena to macroscale electrolyzer performance” is therefore critical, the authors argue. In other words, understanding the physics of tiny bubbles could help unlock the full potential of hydrogen as a clean fuel and help improve other gas-evolving electrochemical reactions (or reactors).

Seen this way, the future of green hydrogen may hinge not just on breakthroughs in chemistry or materials science, but on something more fluid and elusive: the behavior of countless small bubbles rising, colliding, and disappearing inside a reactor. Managing them effectively could be the key to turning a promising technology into a practical one.

When a screenwriter told New York University researchers last year that letting AI do her work would make her "miserable inside," she was onto something.

A follow-up study from NYU’s Tandon School of Engineering and Stern School of Business finds that the instinct to compete with generative AI, rather than simply embrace it, is associated with meaningful long-term benefits for writing professionals.

The catch: rivalry alone isn't enough either.

The 2026 study, led by Rama Adithya Varanasi, a postdoctoral researcher in Tandon's Technology, Management and Innovation Department, alongside Tandon Professor Oded Nov, and Batia Mishan Wiesenfeld, a professor of management at Stern, surveyed 403 professional writers across marketing, publishing, education, and the arts. Findings will be presented at the CHI Conference on Human Factors in Computing Systems this month.

The work extends a 2025 qualitative study by the same team, which interviewed 25 experienced writers and introduced the concept of "AI rivalry" — the idea that some writers proactively compete against AI rather than simply avoid it, targeting what they see as its weaknesses, such as its difficulty producing content rooted in specific communities or geographies.

The new research asked a larger question: what actually happens to writers' careers, skills, and satisfaction depending on how they orient themselves toward AI?

The study finds risks at both extremes. Writers who reported strong collaborative attitudes toward AI also reported higher short-term productivity and job satisfaction, but invested less in maintaining their own skills — the risk of over-reliance.

Writers who perceived AI as a rival reported stronger skill maintenance and greater investment in peer relationships, but that perception showed no significant association with productivity or satisfaction — the risk of under-reliance.

"The concern isn't that workers use AI," said Varanasi. "It's that they stop developing the capabilities that make humans irreplaceable. What this study tells managers is that they can't measure success purely by output. If the workflow removes the need for human judgment, the skill atrophies and that cost doesn't show up until it's too late."

Notably, rivalry attitudes didn't reflect a rejection of the technology. The data showed these writers reported more experience with generative AI than those who held neither orientation strongly. They studied the AI competition, rather than ignoring it.

The most striking result came from writers who scored high on both orientations simultaneously. This group showed the strongest associations with job crafting and skill maintenance across nearly every dimension measured, and posted productivity levels closer to the pure collaboration group — though satisfaction remained higher among pure collaborators — without sacrificing the long-term skill maintenance that pure collaborators showed less of.

"What surprised us is that rivalry and collaboration don't cancel each other out," said Wiesenfeld. "Writers who hold both orientations seem to use AI more deliberately. They get the productivity benefits without outsourcing the judgment."

The study is among the first to measure this tradeoff across a broad set of outcomes — relationships, tasks, cognition, skills, satisfaction, and productivity — drawing on expertise in both human-computer interaction and organizational behavior.

The implications for employers are direct. Organizations that push widespread AI adoption to boost efficiency may be optimizing for the wrong thing, particularly if those workflows come at the cost of workers practicing core human skills.

"Most organizations right now are still developing policies on how employees should relate to AI. " said Nov. "Our findings suggest that the relationship workers have with AI matters as much as whether they use it."

The researchers call for a new design structure that builds productive "friction" into AI tools, calibrating how much assistance is offered based on a user's reliance attitudes rather than defaulting to maximum engagement.

The team's next phase will test that concept directly. They are building prototypes of AI tools designed to promote appropriate reliance, and plan to expand the research beyond writing to other creative professions including game developers, graphic designers, and visual artists.

Funding for this research was provided by the National Science Foundation.

When a customer calls to complain about a billing error or a delayed package, the person on the other end of the line is doing more than answering questions.

They are managing their emotions, suppressing frustration, projecting warmth, absorbing anger, often dozens of times a day.

Researchers at KAIST in South Korea and NYU Tandon School of Engineering say the routine logs generated by call center software may be the most powerful tool yet for detecting when that work is taking a serious toll.

Their study, being presented this month at the CHI Conference on Human Factors in Computing Systems, found that records of call duration, inquiry type, and the notes agents type during conversations ranked among the strongest predictors of post-call stress.

A Cornell University study revealed that 87 percent of call agents experience high levels of stress, contributing to depression, burnout, and turnover rates that plague an industry valued at nearly $30 billion globally. Most workplace stress research focuses on knowledge workers, leaving call agents largely unexamined.

To examine the problem, the researchers spent a month inside a South Korean city government call center. Eighteen agents wore Fitbit trackers, sat beside sensors monitoring CO2 and temperature, and used tablets that recorded the force and rhythm of their typing. After every call, agents rated their stress on a five-point scale, generating more than 7,400 records.

The study was designed around the job's natural rhythm, alternating between customer calls and brief intervals for notes and preparation.

The team fed data from the call center's own server into their models. Every modern call center logs details about each interaction: when a call started, how long it lasted, what the customer's issue was, whether it ended in a complaint. That information outperformed heart rate and movement data.

Long calls with unresolved problems were especially predictive of stress, as were calls requiring repeated explanations. "Even though it's stressful when the customer is unpleasant," one agent said, "it's even more challenging when they inquire without knowing their issue."

Vedant Das Swain, an assistant professor in NYU Tandon's Technology, Management and Innovation Department and a co-author of the study, said the findings reveal a blind spot in how workplaces think about stress. "Most stress research focuses on the worker's body, tracking heart rate, sleep, movement," he said. "But we found that the most revealing signal was the work itself. The call log tells you more about how someone is feeling than a wearable ever could."

Stress also looks different from person to person. Some agents went quiet and still after a hard call. Others typed harder, exhaled loudly, or left their desks. Models calibrated to individual workers outperformed general ones, though only after accumulating around 300 calls of personal history.

The researchers argue that any technology built from these findings should help workers, not surveil them. Stress predictions should go to the agent, not to managers. Aggregated data could help organizations identify systemic problems, but should never be used to evaluate individuals.

"Call centers already know a great deal about their workers," said Uichin Lee, a professor at KAIST's School of Computing and the study's corresponding author. "Our goal was to show that this data could be turned toward the worker's benefit, not used against them."

"In my own earlier research, I looked at designing tools to help call agents regulate their emotions in the moment," Das Swain added. "This paper answers the harder questions first: when to intervene, who to intervene on, and why."

Some agents said that tracking their own stress after each call made them more aware of their emotional state going into the next one. For a job defined by the feelings of others, that turned out to mean something.

The paper's first author is Duri Lee, a researcher in the School of Computing at KAIST. Co-authors are Heejeong Lim of KAIST's Graduate School of Data Science and Das Swain, who shares co-last authorship with Uichin Lee.

NYU and KAIST have built an expanding relationship in recent years, formalizing a partnership in 2022 and introducing a dual degree master's program with NYU Tandon in technology management in 2024.  The current study did not originate as part of the NYU-KAIST Partnership, but Das Swain and Uichin Lee are now continuing to work together through this institution-to-institution collaboration.  

Funding for the study was provided by the Institute of Information and Communications Technology Planning and Evaluation and the National Research Foundation of Korea, both funded by the Korean government, as well as Microsoft's Accelerating Foundation Models Research program and the National Institute on Drug Abuse, part of the National Institutes of Health.

Rivers are rarely the calm, orderly streams we imagine on maps. Over time, their winding paths — called meanders — shift, bend, and occasionally snap off in sudden “cutoff” events that shorten loops and reshape the landscape. While scientists have long suspected that such cutoffs inject a dose of unpredictability into river evolution, a new study published in Communications Earth & Environment demonstrates that these abrupt events are, by themselves, enough to produce chaos in river channels.

Harvard Ph.D. candidate Brayden Noh and NYU Tandon Assistant Professor Omar Wani used a widely-used computational model to explore how meandering rivers evolve over time. This model isolates the essential dynamics: bends migrate laterally in proportion to curvature, and loops are occasionally severed through cutoffs. Other real-world complexities — like sediment transport, bank composition, and vegetation — are treated as secondary, allowing the researchers to focus squarely on the geometry-driven behavior of rivers.

To test the role of cutoffs, the team simulated rivers starting from nearly identical initial shapes, then introduced infinitesimally small perturbations to each of the multiple copies. They tracked how the channels diverged over time by mapping their evolving shapes onto a fixed grid and measuring differences cell by cell. In a striking counterfactual experiment, when cutoffs were disabled, the two channels stayed nearly identical over large time horizons. When cutoffs were allowed, even tiny initial differences grew exponentially, a hallmark of deterministic chaos.

The researchers quantified this sensitivity using the finite-time Lyapunov exponent, a metric from dynamical systems theory that measures how fast nearby trajectories diverge. They found that the rate of divergence depended primarily on the speed at which bends migrated, not on the specific cutoff threshold. In other words, faster meander migration amplifies chaos, while the geometric criteria for triggering a cutoff mostly determine how frequently the river “resets” its local shape.

Importantly, this chaotic behavior was robust across a wide range of initial river geometries. Whether the model started with gentle or pronounced bends, the presence of cutoffs consistently created sensitive dependence on initial conditions. The team also showed that the predictability of a river’s course is bounded: beyond a certain horizon, roughly the number of cutoffs expected in one Lyapunov time, deterministic forecasts of channel position become unreliable.

The study highlights a subtle but powerful insight: continuous meander migration creates gradual stretching of the river planform, while cutoffs act as abrupt topological resets. Together, these processes produce a hybrid system that is both structured and inherently unpredictable. The finding resonates with broader observations of natural rivers, where cutoffs cluster or cascade, triggering sequences of rearrangements along the channel.

While the model is simplified — it does not include full fluid dynamics, sediment heterogeneity, or flood variability — it provides a clear counterfactual experiment: no real river can evolve without cutoffs, but simulations can, revealing the mechanism behind chaotic divergence. This approach connects geomorphology with fundamental concepts from chaos theory, offering a concrete way to quantify a river’s predictability horizon.

Ultimately, the research suggests that some limits to forecasting river evolution are intrinsic. Even in the absence of storms, landslides, or human intervention, the combination of smooth bend migration and occasional cutoffs ensures that lowland rivers retain a degree of inherent unpredictability. For engineers, ecologists, and planners, this work underscores the importance of probabilistic frameworks over deterministic predictions when assessing river migration and floodplain evolution.

In physics, some waves behave in a surprising way: instead of spreading out and fading, they hold their shape as they travel at constant speeds. These unusual waves, called solitons, have interested scientists since they were first observed in canals in the 19th century. Today, researchers study solitons in everything from optical fibers to biological systems.

A new study published in Proceedings of the National Academy of Sciences, shows that these stubborn waves can be guided and steered through materials by carefully designing internal strain, offering new ways to move energy or information at microscopic scales.

The research focuses on liquid crystals, the same class of materials used in LCD screens. But beyond displays, liquid crystals are prized by physicists because their internal structure can be manipulated with remarkable precision. Molecules inside them tend to align in a common direction, but that alignment can be twisted, bent, or reoriented with electric fields or surface chemistry.

In the new work, a research team from NYU and Cornell created special liquid-crystal cells where the molecules were forced to align differently at two opposite surfaces. One surface caused molecules to lie flat, while the other made them stand upright. The result was a continuously bent molecular orientation across the film — a built-in strain field inside the material.

When the researchers applied a high-frequency alternating electric field, something interesting happened. Tiny, localized pulses called “soliton bullets” began shooting through the liquid crystal. These bullets are not physical particles. Instead, they are traveling distortions of molecular alignment, moving through the material while maintaining a stable shape.

Earlier experiments showed that in uniformly aligned liquid crystals, these bullets typically move in just one direction. Under the new conditions, instead of following a single straight path, the soliton bullets traveled along two slanted trajectories, forming diagonal routes through the material. Even more intriguing, the direction of these paths could be tuned simply by adjusting the frequency of the electric field.

To understand why, the team combined experiments with theoretical models and computer simulations. The key turned out to be a phenomenon called flexoelectricity, a coupling between electric fields and mechanical distortions in liquid crystals.

Because the background molecular alignment in the strained cells was already bent, the electric field produced uneven torques on different parts of the soliton structure. Each soliton has two “wings,” regions where the molecular orientation tilts in opposite ways. In the strained environment, one wing becomes stronger than the other, generating a sideways push that sends the soliton along an angled path.

“In these systems, the material itself becomes a way to steer nonlinear signals,” said Juan de Pablo, Executive Dean at NYU Tandon and a coauthor of the study. “By engineering strain into the liquid crystal, we can control how these localized waves move.”

The finding illustrates a broader principle in materials science: the geometry and internal stresses of a material can shape how energy moves through it. In this case, carefully designed strain fields turn a simple liquid-crystal film into a kind of microscopic racetrack for solitons.

Such control could eventually help researchers design active or autonomous materials: systems that move energy, particles, or signals without mechanical components. Previous work has already shown that soliton waves in liquid crystals can transport tiny particles or even trigger droplet formation at fluid interfaces.

While practical devices may still be years away, the study highlights how liquid crystals serve as powerful model systems for exploring nonlinear physics.

“Controlled propagation of soliton bullets in an engineered strain field,” Alexis de la Cotte, Xingzhou Tang, Chuqiao Chen, S. J. Kole, Noe Atzin, Juan J. de Pablo, and Nicholas L. Abbott PNAS, #2025-18064R.