On the veranda of her family\u2019s home, with her laptop balanced on a mud slab built into the wall, Monsumi Murmu works from one of the few places where the mobile signal holds. The familiar sounds of domestic life come from inside the house: clinking utensils, footsteps, voices.<\/p>\n
On her screen a very different scene plays: a woman is pinned down by a group of men, the camera shakes, there is shouting and the sound of breathing. The video is so disturbing Murmu speeds it up, but her job requires her to watch to the end.<\/p>\n
Murmu, 26, is a content moderator for a global technology company, logging on from her village in India\u2019s Jharkhand state. Her job is to classify images, videos and text that have been flagged by automated systems as possible violations of the platform\u2019s rules.<\/p>\n
On an average day, she views up to 800 videos and images, making judgments that train algorithms to recognise violence, abuse and harm.<\/p>\n
Monsumi Murmu in forest near her home. Photograph: Anuj Behal<\/p>\n
This work sits at the core of machine learning\u2019s recent breakthroughs, which rest on the fact that AI is only as good as the data it is trained on. In India, this labour is increasingly performed by women, who are part of an workforce often described as \u201cghost workers\u201d<\/a>.<\/p>\n \u201cThe first few months, I couldn\u2019t sleep,\u201d she says. \u201cI would close my eyes and still see the screen loading.\u201d Images followed her into her dreams: of fatal accidents, of losing family members, of sexual violence she could not stop or escape. On those nights, she says, her mother would wake and sit with her.<\/p>\n In terms of risk, content moderation belongs in the category of dangerous work, comparable to any lethal industryMilagros Miceli, sociologist<\/p>\n Now, she says, the images no longer shock her the way they once did. \u201cBy the end, you don\u2019t feel disturbed \u2013 you feel blank.\u201d There are still some nights, she says, when the dreams return. \u201cThat\u2019s when you know the job has done something to you.\u201d<\/p>\n Researchers say this emotional numbing \u2013 followed by delayed psychological fallout \u2013 is a defining feature of content moderation work. \u201cThere may be moderators who escape psychological harm, but I\u2019ve yet to see evidence of that,\u201d says Milagros Miceli, a sociologist leading the Data Workers\u2019 Inquiry<\/a>, a project investigating the roles of workers in AI.<\/p>\n \u201cIn terms of risk,\u201d she says, \u201ccontent moderation belongs in the category of dangerous work, comparable to any lethal industry.\u201d<\/p>\n Studies indicate content moderation<\/a> triggers lasting cognitive and emotional strain, often resulting in behavioural changes such as heightened vigilance. Workers report intrusive thoughts, anxiety and sleep disturbances.<\/p>\n A study of content moderators<\/a> published last December, which included workers in India, identified traumatic stress as the most pronounced psychological risk. The study found that even where workplace interventions and support mechanisms existed, significant levels of secondary trauma persisted.<\/p>\n A slab extending from the mud wall of her house serves as Murmu\u2019s desk. She uses a secondhand laptop to do content moderation work. Photograph: Anuj Behal<\/p>\n As early as 2021, an estimated 70,000 people in India were working in data annotation, which had a market value of about $250m (\u00a3180m) in 2021<\/a>, according to the country\u2019s IT industry body Nasscom. About 60% of revenues came from the US, while only 10% came from India.<\/p>\n About 80% of data-annotation<\/a> and content moderation workers are drawn from rural, semi-rural ormarginalised backgrounds. Firms deliberately operate from smaller cities and towns, where rents and labour costs are lower, and a growing pool of first-generation graduates are seeking jobs.<\/p>\n Improvements in internet <\/a>connectivity<\/a> have made it possible to plug these locations directly into global AI supply chains, without relocating workers to cities.<\/p>\n Women form half or more of this workforce. For companies, women are seen as reliable, detail-oriented and more likely to accept home-based or contract work that could be seen as \u201csafe\u201d or \u201crespectable\u201d. These jobs offer rare access to income without migration.<\/p>\n A sizeable number of workers in these hubs come from Dalit and Adivasi (tribal) communities. For many of them, digital work of any kind represents an upward shift<\/a>; cleaner, more regular and better-paid jobs than agricultural labour or mining.<\/p>\n A data annotation office in Ranchi, Jharkhand. Tech firms often set up offices in smaller cities. Photograph: Anuj Behal<\/p>\n But working from or close to home, can also reinforce women\u2019s marginal position, according to Priyam Vadaliya, a researcher working on AI and data labour, formerly with the Bengaluru-based Aapti Institute.apti Institute in Bengaluru.<\/p>\n \u201cThe work\u2019s respectability, and the fact that it arrives at the doorstep as a rare source of paid employment, often creates an expectation of gratitude,\u201d she says. \u201cThat expectation can discourage workers from questioning the psychological harm it causes.\u201d<\/p>\n Raina Singh was 24 when she took up data-annotation work. A recent graduate, teaching had been her plan, but the certainty of a monthly income felt necessary before she could afford to pursue it.<\/p>\n She returned to her home town of Bareilly in Uttar Pradesh and each morning logged on from her bedroom, working through a third-party firm contracted for global technology platforms. The pay \u2013 about \u00a3330 a month \u2013 seemed reasonable. The job description was vague, but the work felt manageable.<\/p>\n I can\u2019t even count how much porn I was exposed to. It was constant \u2026 the idea of sex started to disgust meRaina Singh, data worker<\/p>\n Her initial assignments involved text-based tasks: screening short messages, flagging spam, identifying scam-like language. \u201cIt didn\u2019t feel alarming,\u201d she says. \u201cJust dull. But there was something exciting too. I felt like I was working behind the AI. For my friends, AI was just ChatGPT. I was seeing what makes it work.\u201d<\/p>\n But about six months in, the assignments changed. Without notice, Singh was moved to a new project tied to an adult entertainment platform. Her task was to flag and remove content involving child sexual abuse.<\/p>\n \u201cI had never imagined this would be part of the job,\u201d she says. The material was graphic and relentless. When she raised concerns with her manager, she recalls being told: \u201cThis is God\u2019s work \u2013 you\u2019re keeping children safe.\u201d<\/p>\n Raina working on her laptop: \u2018It didn\u2019t feel alarming, just dull. But there was something exciting too.\u2019 Photograph: Anuj Behal<\/p>\n Soon after, the task shifted again. Raina and six others on her team were instructed to categorise pornographic content. \u201cI can\u2019t even count how much porn I was exposed to,\u201d she says. \u201cIt was constant, hour after hour.\u201d<\/p>\n The work affected her personal life. \u201cThe idea of sex started to disgust me,\u201d she says. She withdrew from intimacy and felt increasingly disconnected from her partner.<\/p>\n When Singh complained, the response was blunt: \u2018your contract says data annotation \u2013 this is data annotation.\u2019 She left the job, but a year on, she says the thought of sex can trigger a sense of nausea or dissociation. \u201cSometimes, when I\u2019m with my partner, I feel like a stranger in my own body. I want closeness, but my mind keeps pulling away.\u201d<\/p>\n Vadaliya says job listings rarely explain what the work actually involves. \u201cPeople are hired under ambiguous labels, but only after contracts are signed and training begins do they realise what the actual work is.\u201d<\/p>\n Remote and part-time roles are promoted aggressively online as \u201ceasy money\u201d or \u201czero-investment\u201d opportunities, and circulated through YouTube videos,<\/a> LinkedIn posts, Telegram channels and influencer-led tutorials that frame the work as flexible, low-skilled and safe.<\/p>\n Hyderabad is home to India\u2019s AI industry \u2013 far removed from the scattered rural locations where data is actually labelled. Photograph: Anuj Behal<\/p>\n The Guardian spoke to eight data-annotation and content-moderation companies in India. Only two said they provided psychological support to workers; the rest argued that the work was not demanding enough to require mental healthcare.<\/p>\n Vadaliya says that where there is support, the individual has to seek it out, shifting the burden of care on to workers. \u201cIt ignores the reality that many data workers, especially those coming from remote or marginalised backgrounds, may not even have the language to articulate what they are experiencing,\u201d she says.<\/p>\n