The Rise of “Ghost Work”: Humans Quietly Fixing AI Mistakes Behind the Scenes

By Adrian Keller | Updated on May, 2026 | 🕓 9 minutes

Key Highlights

- What is ghost work, and how is it hidden behind AI systems?

- Which companies have relied on human labor to maintain “fully automated” experiences?

- What are the ethical, legal, and technical risks of ghost work for businesses?

- How can tech decision-makers, engineers, and consumers respond to ghost work responsibly?

- What practical steps can transform AI systems from “ghost-dependent” to human-machine collaborative?

On a Tuesday afternoon in early 2024, I attended an AI product launch in Berlin. The speaker showcased a "fully automated" customer service system, and the demonstration ran almost flawlessly. After the presentation, I noticed a subtle detail: during the Q&A session, the system took about 0.8 seconds to respond to several edge-case questions—a delay pattern remarkably similar to what I observed three years earlier at a remote customer support center in Nairobi, Kenya.

I am not implying that this system was "cheating." What I want to highlight is that we are living in a carefully orchestrated "automation theater," and the workers behind the scenes—referred to in academia as "ghost workers"—are the ones keeping this illusion alive.

I. The Hidden Humans: Four Real-World Cases

Scenario 1: Amazon Just Walk Out’s "Indian Team"

In April 2024, Amazon’s "Just Walk Out" technology sparked a public relations embarrassment. This technology was marketed as a "computer vision-driven, checkout-free" system—customers could take items and leave, while the system automatically charged them. However, an investigation by The Information revealed that about 70% of transactions in 2022 required manual review by a team in India. Amazon later admitted to employing human reviewers but denied claims that "1,000 Indian workers were watching shopping in real time," emphasizing that these workers were primarily involved in "data labeling and model training."

Regardless of the numbers, one fact is undeniable: the "seamlessness" of the consumer experience largely depends on the labor of remote workers. Amazon eventually withdrew the technology from most of its own stores and instead promoted Dash Cart—a more honest solution because it clearly informs users that a screen in the cart is calculating in real time.

Scenario 2: ChatGPT’s "Safety Filter"

In 2023, a TIME magazine investigation revealed OpenAI’s partnership with the outsourcing company Sama: workers in Kenya were paid $1.32–$2 per hour to filter toxic content for ChatGPT. Richard Mathenge, a data labeler with a university degree, processed 150–250 pieces of content per day, including detailed depictions of sexual abuse, torture, and suicide. He later shared in a podcast: "Our team started signaling—we didn’t want to continue seeing these texts."

Another team member, Mophat Okinyi, experienced severe consequences: prolonged exposure to extreme content led to insomnia, anxiety, depression, and panic attacks, ultimately causing his wife to leave him. "Making ChatGPT safe destroyed my family," he said.

Scenario 3: Content Moderators in a Barcelona Clinic

In 2022, a mental health clinic in Barcelona treated a 35-year-old female patient who had previously worked for five years in online content moderation, handling material such as child sexual abuse and sexual assault imagery. At the time of treatment, she suffered repeated daily panic attacks, intrusive images, insomnia, and nightmares. Doctors diagnosed her with post-traumatic stress disorder (PTSD), but medication provided only partial relief, and most symptoms persisted.

This case, published in an NIH academic journal, serves as a stark reminder: the psychological trauma of content moderation is not merely an "occupational risk" but a systematically neglected occupational disease.

Scenario 4: Uber’s "Ghost Drivers"

On an even more subtle level, some AI systems’ so-called "real-time intelligence" actually relies on humans intervening within millisecond windows. Uber initially relied on remote teams to handle approximately 1% of anomalous transactions; some AI customer service systems deliberately simulate "AI thinking" delays when humans take over, giving users the impression that the machine is responding, while in reality a human is typing.

II. Why Ghost Work Is Systematically Hidden

This is not merely a moral failing of individual companies—it is the result of structural mechanisms.

The narrative of capital. AI company valuations are built on the myth of "scalability." If investors understood that systems relied heavily on low-paid human labor, valuation models could collapse. Mary Gray and Siddharth Suri, in Ghost Work, point out that the tech industry has strong incentives to "dehumanize" human labor because "automation" is more attractive to capital than "human-machine collaboration."

Geographic arbitrage and labor reclassification. The Global North consumes AI, while the Global South provides the labor. Workers in Kenya, India, and the Philippines perform tasks for platforms such as Sama, Scale AI, and Appen. These platforms reclassify labor as "task-based" rather than "employment," circumventing minimum wage laws, mental health protections, and other labor rights.

Interface design for “dehumanization.” Chatbots with "typing..." indicators, voice assistants with instant responses, and customer service systems labeled "AI at your service" all deliberately erase traces of human involvement. A single API call may trigger the labor of hundreds of distributed workers, yet the end user remains completely unaware.

III. The Triple Backlash of Ghost Work for Tech Decision-Makers

If you are a tech decision-maker, product manager, or engineer, ghost work is not just an "ethical issue"; it can backfire on your business in three ways:

Compliance and legal risk. The EU AI Act requires high-risk AI systems to maintain transparency. The Sama/Meta case in Kenya and the collective lawsuit of 184 moderators signal a tightening global regulatory environment. Does your AI vendor disclose its data-labeling supply chain? This is becoming a required due diligence question.

Quality and brand risk. Amazon’s withdrawal of Just Walk Out serves as a warning: when the "automation theater" collapses, brand reputation suffers far more than technical adjustments cost. High turnover rates in content moderation—up to 50% annually—mean the quality of training data continually declines, creating a vicious cycle.

Hidden accumulation of technical debt. Over-reliance on human fallback mechanisms prevents AI models from genuinely learning edge cases. "Human-in-the-Loop" becomes "Human-as-a-Crutch," leaving system architecture fragile. Even more subtly, ghost workers hold the "tacit knowledge" of AI systems; if they leave, that knowledge disappears with them.

IV. From "Ghost" to "Symbiosis": Actionable Paths for Transformation

For tech decision-makers/product managers:

- Transparency strategy: Clearly disclose human involvement in AI systems. This is not a weakness, but a trust asset. Nike, for instance, built human supervision mechanisms and detailed audit logs into its AI shopping assistant from day one, allowing non-technical staff to monitor AI performance.

- Architecture design: Make Human-in-the-Loop an explicit module rather than a hidden patch. Unilever’s Beauty AI Studio is a good example: AI generates marketing content, but all final creative output must be reviewed by marketing personnel. The company invested in training 25,000 employees in AI collaboration skills.

- Vendor audits: When procuring AI services, ask at least five questions: Does the vendor disclose its data-labeling supply chain? Do workers receive mental health support? Are NDAs used to suppress worker organization? Is the pay at least local minimum wage? Are third-party labor audits available?

For developers/engineers:

- Ethical embedding at the code level: Retain auditable paths for human intervention. PepsiCo’s data validation pipeline uses Apache Airflow and Great Expectations to automatically route cases requiring judgment to business teams while recording the decision rationale.

- Design “worker feedback channels”: Allow labelers to report data quality issues directly rather than through multiple intermediaries. Airbnb’s Wall framework succeeds partly because it allows business roles to participate in quality control, breaking the information barriers between technical teams and labelers.

For ordinary users/consumers:

- Identify "pseudo-AI" using five signals: Overly perfect response times for complex issues (likely human-assisted); service pricing inconsistent with claimed automation; companies refusing to disclose AI tech stacks; edge-case handling is unusually smooth (human intervention traces); LinkedIn employee profiles show many "AI Trainer" roles while products are marketed as "fully automated."

V. An Unfinished Conversation

In 2024, when Amazon withdrew Just Walk Out from most of its stores, one executive remarked: "Customers in large supermarkets want a shopping assistant accompanying them." This is an interesting comment—it suggests a possibility: what if Indian workers had been designed as "collaborators" from the start rather than "hidden patches"? Would the outcome have been different?

The issue of ghost work is fundamentally not about "AI being insufficiently intelligent," but about "our lack of honesty." When the tech narrative shifts from "fully automated" to "human-machine symbiosis," and supply chain transparency becomes a competitive advantage rather than a burden, the workers behind the scenes may transform from "ghosts" into "visible contributors."

This will not happen automatically. It requires engineers to leave audit paths in code, product managers to honestly label human involvement in PR materials, and consumers to ask probing questions when issues arise. Most importantly, it requires acknowledging a simple fact: the smartest systems are often those that collaborate best with humans—not those that hide humans the best.

Here’s a fully detailed Frequently Asked Questions section with clear, concise answers based on your “Ghost Work” article:

FAQs

1. How common is ghost work in AI today?

Ghost work is widespread, especially in systems that are marketed as "fully automated." Many AI-powered services—such as customer support chatbots, content moderation tools, and “just walk out” retail systems—rely on human labor behind the scenes for tasks like verifying transactions, filtering toxic content, and correcting edge-case errors. Investigations show that even high-profile AI products often depend on hundreds to thousands of low-paid workers globally.

2. What psychological risks do ghost workers face?

Ghost workers, particularly in content moderation and AI training, face significant mental health risks. Exposure to violent, abusive, or disturbing material can lead to insomnia, anxiety, depression, panic attacks, and post-traumatic stress disorder (PTSD). Long-term exposure without adequate support has been shown to impact personal relationships and overall well-being.

3. Are there regulations governing transparency in AI systems?

Yes, regulatory efforts are increasing. For example, the European Union’s AI Act requires high-risk AI systems to maintain transparency about their human and algorithmic decision-making processes. Additionally, labor and human rights organizations advocate for ethical standards in outsourcing and AI data labeling. Companies are increasingly expected to disclose their AI supply chains, worker protections, and human-in-the-loop processes.

4. How can users identify when a system is partially human-operated?

Users can look for five common signals:

- Response times that are unusually perfect on complex queries, suggesting human assistance.

- Service pricing inconsistent with claimed automation levels.

- Companies refusing to provide technical details about AI implementation.

- Edge-case problem-solving that seems overly smooth or contextually aware.

- Public job listings for “AI Trainer” or similar positions while the product claims full automation.

5. Can companies make AI systems fully autonomous without ghost labor?

While some AI systems can function with minimal human intervention, most commercially deployed AI still requires human oversight to handle edge cases, ensure quality, and maintain safety. Fully autonomous AI is challenging due to limitations in current machine learning models, the complexity of real-world scenarios, and ethical or legal requirements. Human involvement remains a critical safety and quality mechanism in the foreseeable future.

References

1. Gray, M., & Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt.

2. Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.

3. The Information. (2024, April). Investigation into Amazon Just Walk Out and human review involvement.

4. TIME Magazine. (2023). Exposé on OpenAI’s content moderation outsourcing in Kenya.

5. NIH Case Report. (2022). Psychological impact of content moderation: PTSD in professional reviewers.

About the Author

Adrian Keller, MSc is an emerging technology analyst specializing in macro innovation trends, biotechnology commercialization, and socio-technical timing dynamics. He holds a Master’s degree in Technology Policy from MIT and has worked with venture studios and research groups analyzing how early-stage technologies fail or succeed based on market readiness rather than technical capability. His writing connects long-term technological shifts with real-world adoption patterns and systemic constraints.

Editorial Transparency Statement

This article is based on publicly available reports, journal articles, and verified interviews with workers and industry insiders. Sources have been cross-checked for accuracy, and all claims are supported by references. No sponsored content influenced the editorial perspective.

Disclaimer

The content provided is for informational and educational purposes only and does not constitute legal, financial, or professional advice. The author and website are not responsible for any decisions made based on this article.