5.4 Psychological Safety, Team Trust & Human Agency

5.4 Psychological Safety, Team Trust, and Human Agency

Core principle: Teams learn when people can question confident output, admit mistakes, and raise harm without retaliation. Psychological safety supports candor; it does not remove professional standards or accountability.

---

Psychological Safety

Psychological safety is commonly described as a shared belief that a team is safe for interpersonal risk. A person can say “I do not understand,” “the model may be wrong,” or “this use could harm someone” without being humiliated. It is not permission for careless work, forced positivity, or avoidance of conflict. Respectful challenge and accountability coexist.

Google’s Project Aristotle reported psychological safety as an important dynamic in its team research. That finding does not prove one variable is always the single greatest cause of performance in every organization. Team composition, purpose, clarity, resources, leadership, workload, and broader culture also matter.

AI Adoption Pressures

People may fear replacement, exposure of skill gaps, loss of craft, surveillance, or unequal access to tools and training. Those concerns should be investigated rather than dismissed with a promise that AI will “free everyone for higher-value work.” AI can reduce effort on some tasks while adding review, correction, coordination, or security work on others.

Run bounded experiments. Define the task and outcome, allow an appropriate comparison, count verification and rework, and ask participants about cognitive load and learning. Make non-use possible where policy and task allow it. Provide training and time rather than evaluating people on adoption volume.

Surveillance and Metric Gaming

Prompt counts, accepted completions, typing speed, lines of generated code, and individual velocity are activity proxies. Turning them into targets can invite gaming—a pattern often summarized through Goodhart’s law—and can suppress honest reporting. Scrum does not prescribe individual productivity metrics. Product value, team outcomes, quality, and sustainable learning are stronger inspection targets.

Transparency in Scrum does not mean universal personal monitoring. Employment monitoring may raise privacy, labor, bias, and legal issues. Use the minimum data for a legitimate purpose, involve appropriate specialists and worker representatives, make logic and consequences visible, provide correction, and avoid covert behavioral scoring.

Blameless Does Not Mean Consequence-Free

When AI-assisted code causes a defect, examine source quality, permissions, review, tests, workload, incentives, and the Definition of Done. The person who surfaces the problem should not be punished for candor. At the same time, the team corrects the product and maintains professional accountability. Repeated reckless or unauthorized behavior may require management action under clear policy; “blameless” is not a shield against all responsibility.

Team AI Working Agreement

A useful agreement may cover:

• approved tools, accounts, and data classes;
• when material AI assistance is disclosed;
• required evidence and review by risk;
• handling of prompts, outputs, logs, and retention;
• prohibited uses such as covert individual scoring;
• how to report a hallucination, leak, bias, or unsafe instruction;
• accessibility and equitable access to training;
• conditions for experiments, pausing, and incident response.

The agreement does not override organizational policy or law. A top-down restriction can be appropriate when a use is unsafe or unlawful. Blanket bans can sometimes drive use underground, but that is a risk to manage, not an inevitability. Clear reasons, usable approved alternatives, and responsive review reduce shadow use.

Human Craft and Learning

Use AI in ways that preserve understanding. Developers can predict an approach before seeing a completion, explain generated code, test edge cases, compare alternatives, pair with colleagues, and practise core skills without assistance. A team can rotate tool use and review so capability does not concentrate in one person.

An AI role prompt does not create a mentor. Training needs qualified people, practice, feedback, and time. Measure whether the team can operate, diagnose, and recover when the tool is unavailable or wrong.

Scrum Master Contribution

The Scrum Master is accountable for establishing Scrum and for the Scrum Team’s effectiveness. They coach self-management and cross-functionality, help remove impediments, facilitate stakeholder collaboration, and cause removal of barriers. They are not a sole “guardian” who owns everyone’s safety. The Scrum Team and organization share responsibility for the conditions people experience.

Healthy Signals

Look for questions, dissent, incident reporting, shared learning, source checking, realistic workload, and willingness to stop unsafe automation. A shared prompt library can help if it includes purpose, approved data, limitations, evidence, and owners—not merely clever phrasing.