How People Detect AI Risk Through Behaviour in Digital Life — 2025 Data

This page summarises findings from the Human Clarity Institute’s AI Safety, Risk Perception & Boundary Behaviour 2025 dataset, based on 301 valid responses across six English-speaking countries. The research focuses on how people detect and respond to AI-related risk through behavioural signals such as inconsistency, unpredictability, and changes in system behaviour.

View the AI Safety, Risk Perception & Boundary Behaviour 2025 Dataset

What the data shows

The clearest signals in this dataset show that people detect AI risk primarily through behaviour. When AI systems feel inconsistent, unpredictable, or difficult to interpret, unease rises quickly and people adjust how they use them.

Overall, the data suggest that AI risk is not judged abstractly. Instead, people rely on behavioural cues — such as inconsistency, instability, and unpredictability — to decide whether a system feels safe enough to use.

By the numbers (from HCI data)

Patterns observed in the data

Behaviour is the primary signal of risk

People do not rely on technical understanding to assess AI safety. Instead, they use behavioural cues such as inconsistency, unpredictability, and sudden changes to judge whether a system feels reliable.

Unpredictability drives disengagement

When AI systems feel unstable or confusing, many users reduce how much they rely on them. This suggests that perceived instability directly influences behaviour.

Predictability is interpreted as safety

Stable, consistent systems are preferred because they are easier to interpret and feel more controllable.

Risk boundaries are context-dependent

Perceived risk increases in specific use cases such as health, children, and opinion-shaping, indicating that people apply different thresholds depending on context.

IN-DEPTH REPORTS

Digital Trust

Digital Trust examines how people judge credibility, navigate uncertainty, and respond to authenticity concerns in AI-shaped digital environments.

Values vs Noise

Values vs Noise explores how digital distraction, noise, and misalignment can erode clarity — and why human values still matter when deciding what deserves trust.

Methodology

This dataset forms part of the Human Clarity Institute’s Human–AI Experience research programme, examining how people perceive AI safety, assess risk, and define acceptable boundaries for AI use in everyday digital life. The study uses a cross-sectional online survey design and focuses on descriptive patterns in safety concern, reliability judgement, verification behaviour, boundary-setting, and trust in AI systems.

The dataset was collected on 1 December 2025 via Prolific as part of the Human–AI Experience research programme. All participants provided explicit consent for anonymised open publication.

Sampling & participants

• Final n: 301
• Countries: UK, US, Canada, Australia, New Zealand, Ireland
• Eligibility: Fluent English
• Recruitment platform: Prolific

The resulting dataset should be interpreted as a non-probability convenience sample and is not intended to represent national populations.

The cleaned dataset, variable dictionary, and reuse terms are publicly available through the HCI dataset repository: AI Safety, Risk Perception & Boundary Behaviour 2025 Dataset →

Data integrity

All percentages reported on this page are calculated from valid responses in the cleaned dataset (n = 301). Percentages are rounded to the nearest whole number for readability. Unless otherwise stated, summary percentages combine respondents selecting 5–7 on the 7-point agreement scale (slightly agree, moderately agree, or strongly agree).

No approval-rate filter, attention checks, or AI-deception trap items were applied in this study. Prolific IDs were removed and timestamps were stripped before publication as part of the anonymisation process.

This dataset is exploratory and descriptive in nature. It does not support causal inference and results should be interpreted as observed patterns within the survey sample.

This dataset is released as open research to support transparent analysis of AI risk perception, behavioural trust boundaries, verification behaviour, and everyday safety judgement in the AI era.

Suggested citation:
Human Clarity Institute. (2025). AI Safety, Risk Perception & Boundary Behaviour 2025 (Dataset). https://doi.org/10.5281/zenodo.17782046

Data use and reuse terms are outlined in our Data Use & Disclaimer.

Explore further analysis on Human Clarity Insights, or browse the full collection of HCI research reports.