Introduction
Ethical considerations in video surveillance
Video surveillance has unlocked a world of new possibilities, from safeguarding individuals and communities to improving work efficiency and understanding customer behavior. Yet, concerns about its ethical implications and impact on privacy are growing.
In public health emergencies, where security and compliance are vital, the delay between capturing an image and taking action could be crucial. While humans ultimately make the decisions, AI support can ethically process video footage more effectively.
Challenges
Adapting to pandemic protocols through technology
During the COVID-19 pandemic, when global shutdowns mandated a two-meter distance in public, we collaborated with a drone manufacturer to create a system tracking breaches of these rules. The project was overseen by a public authority to boost public health awareness and adherence using live video feeds.
Balancing surveillance with privacy
Surveillance data from public spaces contains sensitive personal information. Protecting this information while retaining the effectiveness of surveillance was critical.
Ensuring transparency and auditability
For an AI-enabled compliance system to gain public trust, it must be transparent and auditable. As it directly influences societal behavior, clear and accessible guidelines on its operation and decision-making processes are necessary.
Upholding accountability with technology
Bringing together technology, policy, and legal frameworks to safeguard privacy and ensure responsible system use is complicated. A thorough oversight mechanism was needed to monitor and ensure the alignment with legal and ethical standards.
The solution
Building a Responsible AI (RAI) framework
We aimed to create a smart system for fast decision-making with instant alerts for social distancing breaches. This system would minimize human effort in CCTV monitoring. Due to the ethical implications of mass surveillance, the solution was designed within a framework of responsible AI.
AI algorithms are built on human assumptions, and their data reflect the context of their creation, not absolute truths. Still, AI must embody human values and be crafted to minimize associated risks through key principles and behaviors.
Foundational Principles
Transparency: Everyone deserves clear, understandable decisions.
Accountability: Humans are responsible.
Privacy & safety: Ensure respect and prevent misuse.
Fairness: Commit to inclusion and equity.
Encouraging Behaviors
Human-centricity: Continuously check for biases.
Contestability: Apply the “newspaper test” (i.e., assess whether processes are robust enough to withstand public scrutiny).
Explainability: Make decision processes transparent.
Adaptability: Regularly update and refine based on feedback.
To follow the RAI framework, we assembled a diverse team, including legal, domain, product, AI, and engineering experts, to assess each stage of the AI project’s development. This emphasized the importance of three main principles: privacy, transparency, and accountability, which were incorporated into every step of development.
Problem definition: What is the AI’s purpose? Who are the stakeholders and beneficiaries? Could the AI cause unintentional harm?
Data input: Is the data collection process clear? Are there any alterations to the data? Does the data accurately represent the real world?
Model building: Are the models unbiased and thoroughly vetted? What are the limitations of these models?
Output evaluation: What outputs will the model generate? Is the model reliable? Can we interpret the AI’s decisions?
Implementation: What decisions does the AI influence, and who is affected? How will the outputs be utilized? What are the risks of AI modifications? Who has access to specific information?
What We Provided: AI for social distancing compliance
We developed a system that monitored and reported social distancing breaches in real-time. The dashboard displayed important metrics like violation numbers and severity, tailored to each location. We further enhanced the system by embedding three fundamental principles of responsible AI. By addressing challenges related to these principles, we ensured strict adherence to ethical AI standards.
| Privacy | Transparency | Accountability | |
| Challenge | How do we safeguard personal privacy while assisting law enforcement with sensitive data? Enable the categorization of objects while maintaining their privacy. |
How should such an AI-enabled behavioral compliance system be moderated and be made audit-ready? An AI surveillance system directly impacts society and its behavioral traits. The solution needed to operate with transparent guidelines and be auditable. |
How do we ensure accountability while combining technology, policy, and legal frameworks to protect privacy rights and ensure responsible and effective system usage? Unblurred images provided information on a need-to-know basis. As an individual was identified through unblurring, strict records needed to be kept for actualizing a chain of accountability. |
| Solution | Blurred detected objects on the video stream to deliver privacy. The object was masked and encrypted to protect its identity. |
Under public authority supervision, we limited the visibility and access control of the transparency view. To ensure full transparency, only authorized stakeholders with encryption keys could unblur images and access object identity. |
Accountability becomes paramount because access to sensitive information is restricted based on necessity. Each access was recorded and auditable based on parameters including, but not limited to: ● Accessor details ● Accessed object details ● Timestamp ● Reason for accessing ● Retention period record |