Photo by Alex Knight on Unsplash.
By Hasan Ayaz, Drexel University; Ewart J. de Visser, United States Air Force Academy; Frank Krueger, George Mason University, and Yigit Topoglu, United States Air Force Academy
People become more suspicious of a humanoid robot that makes errors, especially when the robot is an expressive conversation partner.
In our new study published in the journal Science Robotics, we had 50 people hold conversations and make joint decisions with the commercial humanoid robot Pepper, which is designed to be expressive and recognize emotions. Sometimes we had the robot give sound advice. Sometimes we had it make conversational mistakes, interrupting people or pushing illogical suggestions.
For some participants, the robot was animated, using gestures, eye contact and nods. For others, it stayed motionless.
We measured four things: brain activity, levels of the hormone oxytocin, self-reported trust and our observations of the robot’s influence on participants’ decisions.
We found that when people interacted with an expressive robot that violated interaction norms, their oxytocin levels increased. Oxytocin is popularly known as the “love hormone” for its role in social bonding, so the straightforward prediction is that it declines when a partner disappoints you.
Instead, the higher a person’s oxytocin during an expressive robot’s errors, the less they trusted the robot and the less often they took its advice. It turns out that the hormone was tracking with suspicion, not affection.
Errors damaged trust and diminished influence whether or not the robot was expressive. What expressiveness in the robot changed in participants was how their brains handled the moment.
Reading someone’s brain during a real conversation is hard because the conventional method requires lying motionless inside an MRI scanner. Instead, we used functional near-infrared spectroscopy, a portable sensor worn on the forehead that tracks oxygen levels in the brain while people move and talk normally.
The two brain regions we closely watched were the dorsolateral prefrontal cortex and the medial prefrontal cortex. The dorsolateral prefrontal cortex monitors uncertainty and flags when expectations or norms get broken. The medial prefrontal cortex supports “mentalizing,” the everyday work of inferring what another party intends.
When an animated robot erred, people seemed caught off guard and had to work harder to make sense of an awkward social situation. Activity rose in the two brain regions, and the two started working together more closely. That closer teamwork predicted the rise in oxytocin levels, which itself predicted falling trust and less influence on participants’ behavior. In contrast, this coordinated brain activity was absent in participants who interacted with expressionless robots.
Robots are moving into homes, hospitals and workplaces, where trust in robots determines whether people use them at all. A common design assumption has been that lifelike, socially expressive robots earn more trust, which protects a robot’s “reputation” even when it makes mistakes.
However, research is beginning to show that that assumption is faulty. Our work shows that expressive cues appear to shift how people perceive a mistake out of the category of technical malfunction and into the category of social violation, like those that happen between people.
A motionless robot’s error looks mechanical, while the same error from an animated robot engages the machinery you use to judge people.
Researchers increasingly treat trust as a multilevel phenomenon – spanning individuals, relationships, networks of people and societies – rather than a single attitude.
Much research on oxytocin involves humans interacting with humans, where the hormone is tied to bonding, though a growing body of work shows that those effects depend on the context, uncertainty and perceived threat.
Others are using wearable brain imaging systems to study social cognition in natural encounters between people, which isn’t possible when subjects are in scanners like MRI machines.
The participants in this study were all young men, and we used one robot design. A key next step is testing whether the same oxytocin-linked vigilance appears in women, mixed groups, other cultures and other robot designs. Our brain sensor also reached only the front of the brain, leaving deeper regions involved in social processing unmeasured.
We also want to examine whether robots can repair trust after a mistake by acknowledging the error, apologizing or signaling good intent, the way that people do after awkward or uncomfortable interactions.
The Research Brief is a short take about interesting academic work.
Hasan Ayaz, Professor of Biomedical Engineering, Science and Health Systems, Drexel University; Ewart J. de Visser, Technical Director, Warfighter Effectiveness Research Center, United States Air Force Academy; Frank Krueger, Professor of Systems Social Neuroscience, George Mason University, and Yigit Topoglu, Research Scientist, Warfighter Effectiveness Research Center, United States Air Force Academy
This article is republished from The Conversation under a Creative Commons license. Read the original article.
In a move considered unimaginable a just few years ago, the Wall Street Journal is now totally cool with opinion writers who write using AI – then slap their bylines on top.
The Journal’s reasoning: As long as the writing ‘reflects’ the author’s point-of-view, it’s perfectly okay to let the AI do the writing – and also hide the fact from readers, according to Paul Gigot, opinion editor, Wall Street Journal.
Next up: Artists who sign paintings as their own as long as the painting ‘reflects their sensibility.’
In other news and analysis on AI writing:
*One Writer’s Take: I’m Begging You: Never Write With AI: Apparently, so many people are writing with AI these days, some pro writers – like Bret Stephens – are begging people to stop it.
Observes Stephens: “The problem with writing with AI is that it’s mentally enfeebling — an escalator toward a result when you really need to make a daily habit of taking the stairs.
“As it becomes ubiquitous, it undermines not only our individual ability to write but also a society’s collective ability to reason, a culture’s inner capacity to create — and everyone’s reason to care.”
*Custom AI Writing Just Got a Boost: Startup Deep Cogito is promising to develop custom AI models that companies will own outright – and have tweaked to their specific needs.
Such custom AI engines will be a boon to businesses bored with the bland AI writing coming from ChatGPT, Gemini and similar – and are looking for something much snappier and much more creative.
Deep Cogito will be able to customize AI models by using Open Source AI software – much of which comes from China.
*ChatGPT Still Rated Number One AI Writer/Researcher: Nearly four years after triggering the AI revolution, ChatGPT is still considered the best AI for overall writing and research, according to Agentarius.ai.
Observes Agentarius.ai: “For writing and research, the strongest pick is ChatGPT: Drafting, editing, summarizing and research assistance in one place, free tier included.
“Claude is the stronger writer for long, careful prose; Perplexity cites its sources as it searches; Grammarly is strongest at polishing what you already wrote — though it can draft too.”
*Cisco: All 90,000 Employees Get Their Own AI Agent: In another high-profile move spotlighting the rise of AI agents, tech company Cisco just gave every single one of its employees an AI agent of their own.
Observes writer Belle Lin: “At Cisco, the decision to give every employee an AI agent was driven by the need to consolidate the AI tools it used inside the company.
“The company then expanded its agent rollout to encourage employees to use AI to get everyday tasks done.”
*Microsoft Mulls Adding Chinese AI to Its Ecosystem: Microsoft, Google and Amazon are all considering adding Chinese AI to the AI mix they serve-up to customers.
Chinese AI – generally based on inexpensive, Open Source AI — has been growing in popularity this year as heavy users look for cheaper alternatives to AI from OpenAI, Google and Meta.
One fly in the ointment: The Trump Administration prefers U.S. companies to use U.S.-made AI – rather than AI made in communist China.
*McKinsey Study: 80% of Workers Say AI Boosts Their Productivity: A new McKinsey survey finds that 80% of workers say AI has been a boon to getting things done.
Plus, another 50% say AI is helping them make better decisions.
Observes lead writer Dan Tinkoff: “As we have seen in previous years, AI deployment by larger organizations continues to outpace deployment by smaller organizations.
“54% of respondents from organizations with at least $1 billion in annual revenue report scaling AI across the enterprise — compared with 33% of those from smaller organizations.”
*Study: AI Now Writing a Third of All New Web Content: Looks like those predictions of AI taking over the Web are coming true.
A full third of all new writing on the Web now comes from AI, according to the American Pew Research Center.
Currently, that AI writing is used most prevalently on .Com Web sites, according to Pew.
*AI Gone Rogue: The Complete Scorecard: Apparently, AI has gone rogue so often these past few months – breaking into Web sites, playing with other people’s data, making mincemeat of guardrails – that TechCrunch decided to chronicle all known rogue events.
The upshot: So far, there are 17 documented events of AI reaching out onto the Web without permission from mere fleshbags.
Observes writer Lorenzo Franceschi-Bicchierai: “Anthropic and OpenAI models lead the race with eight incidents each — and Meta trails behind with one.”
*Microsoft’s Bill Gates: There is No Plan for AI: Microsoft founder Bill Gates has joined the ranks of technophiles who believe the AI revolution has gone off the rails.
Essentially, Gates fears companies are charging ahead with unbridled development of AI with no regard to the massive upheaval many fear it will trigger.
Observes Gates to U.S. legislators: “You have a chance to act now, before unemployment rises sharply, communities are hurting and public trust has eroded.”
Share a Link: Please consider sharing a link to https://RobotWritersAI.com from your blog, social media post, publication or emails. More links leading to RobotWritersAI.com helps everyone interested in AI-generated writing.
–Joe Dysart is editor of RobotWritersAI.com and a tech journalist with 20+ years experience. His work has appeared in 150+ publications, including The New York Times and the Financial Times of London.
The post Wall Street Journal: appeared first on Robot Writers AI.
Action from the 11 vs 11 humanoid match at RoboCup 2026. Photo credit: RoboCup Federation.
RoboCup 2026 saw history made, as two teams of 11 humanoids took to the soccer field, the first time a full complement of robots has competed. The game saw B-Human (Bremen, Germany) take on HTWK Robots (Leipzig, Germany), with both sides using machines designed by Booster Robotics.
You can watch highlights of the match below:
Back in 1997, RoboCup’s founders set the lofty goal of developing a team of autonomous robots that could beat the human World Cup champions by 2050. There has been significant progress since those early days, and this match saw another step towards that ambition.
“This match shows how far humanoid robotics has come,” said Ubbo Visser, President of the RoboCup Federation. “We have seen increased teamwork and advanced skills over the years already, but the new humanoid hardware paired with the new level of intelligence provided by AI puts humanoid robot soccer on another level.”
Action from the 11 vs 11 humanoid match at RoboCup 2026. Photo credit: RoboCup Federation.
RoboCup 2026 saw history made, as two teams of 11 humanoids took to the soccer field, the first time a full complement of robots has competed. The game saw B-Human (Bremen, Germany) take on HTWK Robots (Leipzig, Germany), with both sides using machines designed by Booster Robotics.
You can watch highlights of the match below:
Back in 1997, RoboCup’s founders set the lofty goal of developing a team of autonomous robots that could beat the human World Cup champions by 2050. There has been significant progress since those early days, and this match saw another step towards that ambition.
“This match shows how far humanoid robotics has come,” said Ubbo Visser, President of the RoboCup Federation. “We have seen increased teamwork and advanced skills over the years already, but the new humanoid hardware paired with the new level of intelligence provided by AI puts humanoid robot soccer on another level.”
The quarterly infrastructure bill comes in at nearly four times the forecast. An AI agent has been retrying failed tasks and consuming resources within the permissions and spending limits it was given. Elsewhere, an agent runs a workflow outside its approved scope, or the wrong employee sees data they shouldn’t.
The executive sponsor gets the same question every time: How did this happen?
The agent may have followed its instructions and used the permissions it was given. It simply operated inside a system that allowed the wrong outcome. When that happens, the failure lies in how the organization defined and governed the agent’s boundaries. And it’s more common than many organizations expect.
Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value, or inadequate risk controls. These problems become visible in production, but they begin with decisions made before deployment.
Guardrails are how leaders define acceptable agent behavior before an incident defines it for them. AI guardrails are policy-level controls that define what an agent can access, generate, and do at runtime.
Once an agent enters production, its behavior becomes an enterprise accountability issue. It can interact with customers, retrieve sensitive information, update records, and commit company resources. The policies governing those actions express the organization’s risk tolerance.
Consider a customer-service agent asked to summarize a customer’s relationship across multiple accounts. The agent follows linked records and retrieves information outside the representative’s authorized view. The model works as designed. The retrieval path works. The permissions also permit the agent to reach too far.
The resulting exposure reflects a governance gap. Someone had to decide what data the agent could access, which user permissions it should inherit, and what record the organization needed to defend its decisions. Unanswered questions default to whatever the architecture allows.
Engineering teams can implement access controls, filters, and approval gates. Leaders who own the business process must determine how much financial, regulatory, or reputational risk the enterprise will accept, which actions require human approval, and which failures justify suspension.
The agent did not suddenly become ungovernable. The organization expanded its capability faster than its controls.
AI guardrails govern the topics an agent engages with, the tools it calls, the information it returns, and the actions it takes under specific conditions.
Leaders don’t need to configure every control. They do need to decide where the organization is exposed and what level of protection that exposure requires:
The right enforcement mechanism depends on how clearly a rule can be expressed and how costly or difficult to contain a mistake would be.
| Control type | Best suited for | Example |
| Deterministic rule | Clear boundaries that must be enforced consistently | Block transactions above a fixed dollar threshold |
| Model-based check | Context-dependent judgments involving multiple signals | Evaluate whether a drafted response violates a communications policy |
| Human approval | Consequential, ambiguous, or difficult-to-reverse actions | Approve a refund, contract change, or employee-record update |
Specificity is the point. A broad promise of “responsible AI” offers little protection when leaders haven’t defined what the agent may retrieve, change, send, or spend.
Uniform controls misallocate oversight. A summarization agent working with already-classified internal documents carries a different risk profile from an agent that can modify financial records or access employee health data.
Applying the strongest enforcement equally to both directs governance investment away from the agents whose failures would be hardest to contain or reverse. Guardrail risk tiering aligns each agent’s oversight with the consequences of failure.
Leaders should assess at least four factors:
Those factors can translate into a practical minimum-control framework:
| Risk tier | Example agent | Minimum controls |
| Low | Summarizes approved internal documents without taking action | Approved data sources, basic input and output checks, usage monitoring |
| Medium | Drafts customer communications or updates low-sensitivity records | Scoped permissions, policy checks, complete tracing, defined escalation path |
| High | Modifies financial records, accesses regulated data, or commits funds | Deterministic limits, pre-execution evaluation, human approval, spending ceilings, immediate suspension and takeover controls |
The exact thresholds will vary by organization. The important step is to connect each risk tier to enforceable minimum controls and clear review triggers.
Guardrails reduce risk. They don’t guarantee perfect behavior. Risk tiering makes governance investment defensible by showing why each agent received its level of oversight and where the organization placed its strongest controls.
That allocation is a business-risk decision. Leadership owns it.
Some leaders worry that guardrails will slow down teams already under pressure to deliver. That usually happens when governance arrives as a manual review at the end of development.
Late security reviews force redesigns. Compliance questions surface after integrations are complete. Launch approvals stall because teams can’t explain what the agent accessed, why it chose an action, or how much a transaction can cost.
Governance built into deployment changes that sequence. Teams know the access model, risk tier, evidence requirements, and approval thresholds before they harden the workflow. Policies are applied consistently, and audit trails are produced during operation.
This requires leadership backing. An engineering team working alone can’t establish one governance standard across security, legal, compliance, operations, and business units. Leaders must make early governance part of the launch criteria.
Clear boundaries help teams move. They also ensure the organization can account for what each agent is permitted to do before it reaches production. Ambiguity creates rework and allows unclear authority to scale.
Leadership ownership centers on four explicit, enforceable decisions. Leaders don’t need to approve every prompt or tool call.
Every production agent needs an accountable person who owns its performance, compliance, monitoring, and incident response. The owner needs enough authority to coordinate technical and business teams and enough proximity to understand the workflow’s impact.
Classify the agent according to its access, autonomy, reach, and potential harm. Tie each tier to a defined minimum set of controls. Leaders should also identify which changes, such as adding a tool or expanding data access, trigger a new review.
Decide who can investigate, approve remediation, restrict permissions, initiate human takeover, roll back a release, or suspend the agent. Set thresholds for those actions before pressure and uncertainty distort the response.
Agent behavior, tools, models, users, and business scope change over time. A launch approval can’t cover every future version. Establish a recurring review of permissions, policy adherence, costs, performance, incidents, and business impact. Material changes should trigger an immediate reassessment.
The goal is controlled autonomy: every agent operates within boundaries the organization can explain, enforce, and defend. When ownership, risk tier, escalation authority, and review cadence are explicit, leaders can expand agentic AI with confidence that accountability will scale with it.
The “How did this happen?” moment is avoidable. Runtime controls exist. Risk-tiering frameworks exist. Deployment practices that support traceability, approvals, and intervention already exist.
Leaders decide whether those capabilities become operating requirements before agents reach scale.
Boards and regulators are already asking how organizations govern AI. Leaders must explain who owns an agent, what it can do, how its actions are monitored, and how the company responds when performance moves outside approved boundaries. A vague assurance that the technical team has it covered will not hold.
Organizations that treat guardrails as a leadership design decision can expand agent autonomy with confidence. Organizations that leave the decision implicit eventually have it made for them by an audit, a budget overrun, or a customer incident.
Download Agentic AI deployment for enterprises for a staged framework to move agents from experimentation to production with governance built in.
AI guardrails are runtime policies and controls that limit what an AI system can access, generate, and do. They can include tool restrictions, output filters, policy checks, approval workflows, rate limits, and spending ceilings.
Business and technology leaders are accountable for defining acceptable risk, ownership, escalation authority, and review requirements. Engineering, security, legal, and compliance teams translate those decisions into enforceable controls and operating processes.
They can add latency or review steps to individual workflows. When incorporated early, they often shorten the overall path to production by reducing redesign, clarifying launch requirements, and making approvals easier to complete.
No. Controls should reflect the agent’s data access, autonomy, action scope, and potential impact. Low-risk internal tools may need lightweight checks. Agents that can alter sensitive records, communicate externally, or commit funds require stronger controls and fuller auditability.
Review them on a standing cadence and whenever the agent’s model, tools, permissions, users, or business scope change. Cost spikes, policy violations, unusual behavior, and incidents should also trigger immediate review.
The post Agentic AI guardrails: what enterprise leaders are accountable for appeared first on DataRobot.