💻 computer science

"These cameras are just like the Eye of Sauron": A Sociotechnical Threat Model for AI-Driven Smart Home Devices as Perceived by UK-Based Domestic Workers

Through semi-structured interviews with UK-based domestic workers, this paper develops a sociotechnical threat model that illustrates how AI-driven smart home devices create unique privacy risks across both employer-controlled and personal households, highlighting the role of institutional power and data flows in shaping surveillance experiences.

Original authors: Shijing He, Yaxiong Lei, Xiao Zhan, Ruba Abu-Salma, Jose Such

Published 2026-02-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shijing He, Yaxiong Lei, Xiao Zhan, Ruba Abu-Salma, Jose Such

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a professional chef working in a high-end restaurant. You do your job well, but you realize the owner hasn't just installed cameras to prevent theft—they’ve installed "smart" cameras that can tell if you’re chopping onions too slowly, if you look tired, or if you’re talking too much to a coworker. Even worse, the restaurant keeps a digital "scorecard" of your every movement, and even after you quit, that data stays in their system forever.

That is the core of this research paper. It explores the lives of Domestic Workers (DWs)—like nannies, cleaners, and caregivers—who live and work in "smart homes" filled with AI-driven cameras and voice assistants.

Here is the breakdown of the study in plain English:

1. The "Eye of Sauron" Effect (Workplace Surveillance)

The researchers found that for many workers, smart cameras aren't just security tools; they feel like the "Eye of Sauron" from Lord of the Rings—an all-seeing, judging eye that never blinks.

Because these cameras use AI, they don't just "see" a person; they interpret them. A camera might detect a baby crying and automatically swivel toward the nanny. To the worker, this feels like being constantly graded on a performance review you never signed up for. It creates a "panoptic" feeling: even when no one is watching the live feed, the worker feels they must "act" for the camera, losing their ability to just be themselves.

2. The "Ghost in the Machine" (Data That Won't Leave)

One of the scariest parts for these workers is the "digital footprint" they leave behind. They discovered that smart speakers often keep voice logs or profiles of previous workers.

Think of it like a "digital ghost." A worker might leave a job, but their voice, their habits, and their routines remain stored in the house like a haunting memory. They worry that their "data ghost" will stay in the employer's house long after they have moved on, potentially being used to judge them or being accessed by strangers.

3. The "Middleman" Problem (Agencies as Gatekeepers)

The study points out that the agencies that help people find jobs often act like "unwitting accomplices" to this surveillance.

Instead of protecting the worker, these agencies often act like a bridge that only goes one way. They might sign contracts that use vague language—calling a high-tech AI surveillance system just "security"—without telling the worker exactly how much the AI is actually "thinking" about them. They prioritize the employer's peace of mind over the worker's right to privacy.

4. The "Two-Sided Mirror" (Home vs. Work)

The most interesting finding is how these two worlds collide. The researchers looked at how these workers behave in their own homes.

From Work to Home: Because they’ve been "burned" by surveillance at work, these workers become "Privacy Ninjas" at home. They are much more likely to disable facial recognition, turn off microphones, and hide their devices because they know exactly how much power that data holds.
From Home to Work: Conversely, because they know how to manage their own smart devices, they feel even more frustrated at work. They know that "privacy modes" exist, so when an employer refuses to give them any privacy, it feels like a deliberate choice to control them.

The Big Picture Summary

The paper argues that we shouldn't just think about "hacking" or "bad actors" when we talk about AI privacy. The real threat isn't a hooded hacker in a basement; it's the power imbalance in the room.

The researchers are calling for:

Clearer "Rules of the House": Contracts should explicitly say, "We use AI to detect motion, but we do not use it to track your bathroom breaks."
Better Device Design: Devices should have physical, un-hackable lights that show when they are "thinking" or "recording," so a worker isn't caught off guard.
The "Right to be Forgotten": When a worker leaves a job, their digital data should be wiped clean, ensuring their "digital ghost" doesn't haunt their reputation.

Technical Summary: “These cameras are just like the Eye of Sauron”

1. Problem Statement

The rapid integration of Artificial Intelligence (AI) into smart home ecosystems (e.g., smart cameras with motion/crying detection and voice assistants with persistent memory) has created a new dimension of privacy risk for Domestic Workers (DWs). Unlike typical smart home "bystanders" (guests or family members), DWs occupy a unique socio-technical position: they are employees working in employer-controlled environments where they lack administrative control, yet they are also device owners/administrators in their own private households.

Existing threat models often treat privacy risks as abstract technical vulnerabilities or focus on single-household contexts. This paper addresses the gap in understanding how AI-driven inference capabilities (behavioral analytics, automated classification, and long-term data retention) and interconnected household dynamics (the flow of privacy norms and risks between work and home) impact the privacy and agency of this vulnerable population.

2. Methodology

The researchers employed a qualitative, human-centered approach:

Participants: 18 UK-based domestic workers (including nannies, caregivers, and cleaners) were recruited via social media and university networks. The sample included both live-in and live-out workers.
Data Collection: Semi-structured online interviews (averaging ~74 minutes) were conducted via Microsoft Teams. The interview guide focused on device usage, perceived AI capabilities, privacy boundary negotiations, and the role of employment agencies.
Theoretical Framework: The study utilized Communication Privacy Management (CPM) theory, which analyzes how individuals manage private information through three dimensions: privacy ownership, privacy control, and privacy turbulence (violations of boundaries).
Analysis: The researchers used inductive coding. Three authors independently coded transcripts to develop a unified codebook (stabilizing at 66 codes), achieving high inter-rater reliability ( $\kappa \approx 0.79–0.81$ ).

3. Key Contributions

A Novel Socio-Technical Threat Model: The paper proposes a threat model that moves beyond "malicious hackers" to include institutional adversaries (DW agencies) and insider threats (employers). It maps risks across two interconnected contexts: the workplace (employer home) and the personal home.
Identification of AI-Specific Threats: It distinguishes between traditional surveillance (passive recording) and AI-driven surveillance (algorithmic interpretation of behavior, such as "laziness" or "attentiveness" based on motion logs or crying detection).
Cross-Contextual Analysis: It provides a unique look at how "surveillance-induced vigilance" at work leads to stricter, more cautious device configurations in the workers' own homes.

4. Results

The findings are categorized into three primary domains:

In Employer-Controlled Homes:
- Panoptic Effect: Pervasive camera placement creates a "social spotlight," causing DWs to "act for the camera" rather than behave naturally.
- Algorithmic Judgment: AI features (e.g., motion detection, crying alerts) act as de facto performance metrics. Workers fear that automated logs will be used to justify disciplinary actions or disputes.
- Orphaned Data: Workers expressed significant anxiety regarding "residual data traces"—voice logs or behavioral profiles that persist in the smart home ecosystem long after their employment has ended.
- Blurred Roles: DWs are often pressured into "informal data stewardship," such as monitoring children’s interactions with smart speakers.
The Role of DW Agencies:
- Agencies were identified as institutional adversaries. They often prioritize employer interests, provide no privacy training, and use vague contractual language (e.g., "CCTV") to normalize intrusive AI monitoring.
In Personal Homes:
- Ownership vs. Gendered Control: While DWs have more agency at home, decision-making is often gendered, with male relatives frequently acting as the sole "administrators," leaving women as "passenger users."
- Mitigation via Experience: Workplace trauma leads to proactive privacy measures at home, such as disabling facial recognition or cloud backups.

5. Significance and Implications

The paper argues that privacy for DWs cannot be solved by individual settings alone; it requires systemic changes:

For Designers: Smart devices should include glanceable, non-app-based indicators (e.g., physical lights or audible cues) when AI analytics are active. They should also implement automatic data de-identification/deletion when a user profile (like a worker's) is removed.
For Policymakers/Agencies: There is a need for explicit privacy clauses in employment contracts and the integration of DW-specific protections into labor laws (e.g., prohibiting surveillance in intimate spaces like bedrooms).
For Theory: The study demonstrates that threat models must account for power asymmetries and the way digital safety risks "travel" between different social and physical environments.