Workers in India treat AI as task tool, not workplace takeover

Artificial intelligence (AI) is entering everyday work in India less as a dramatic disruption and more as a quiet extension of familiar digital routines, according to a new study by Emilia Edwards and Dhiraj Murthy of the University of Texas at Austin.

The study, titled Beyond disruption and invisibility: Interactional continuity in everyday AI use in India, was published in new media & society. It examines how workers in Bangalore, India's tech hub, use embedded AI features and standalone generative AI tools, finding that AI becomes useful when it fits into specific task steps such as searching, drafting, revising, scrolling, planning and messaging.

The research challenges two dominant ways of describing AI. One frames generative AI as a disruptive technology that radically changes work, communication and media. The other treats embedded AI as invisible infrastructure that silently shapes what users see and do. The authors argue that both views miss how AI is actually absorbed into ordinary activity. They propose the concept of interactional continuity, a user-centered frame that explains how people incorporate AI across platforms when it helps them move from one step of a task to the next.

The findings are based on 28 semi-structured interviews conducted in July 2024 at an Indian company in Bangalore with employees working across sales, management, marketing, design, administration and internships. The study took place shortly after Meta AI began appearing inside popular apps in India, including WhatsApp and Instagram. That timing allowed the researchers to compare how workers used AI embedded in familiar platforms with standalone systems such as ChatGPT and Gemini.

Workers used AI when it moved a task forward

The study found that participants judged AI mainly by practical value, not by technical category. Users kept AI tools when they helped complete a task faster, more clearly or more conveniently. They dropped or ignored them when the output was irrelevant, required too much verification or created more friction than benefit.

Embedded AI was used most often through content recommendation and search assistance. Participants described Instagram feeds, search results and platform suggestions as part of everyday routines, even when they did not call those systems AI. Some workers used recommendations to find restaurants, weekend destinations or useful content. Others used search features inside platforms to locate specific information on topics such as zoology, automobiles, hotels or local places.

The study found three broad styles of engagement with embedded AI. Some users absorbed AI-driven recommendations into routine scrolling without much reflection. Others used them strategically to support planning or information discovery. A third group calibrated platform behavior by searching more deliberately or avoiding recommendations that felt too broad.

Embedded generative AI, especially Meta AI inside WhatsApp and Instagram, functioned differently. Because it appeared inside familiar communication apps, users encountered it as a low-friction extension of existing messaging habits. Participants used it for conversational information retrieval, content creation and text assistance.

Some asked Meta AI questions about health procedures, practical topics or everyday information. Others used it to generate promotional text, rewrite messages or produce short content for work. The tool's location inside WhatsApp made it easier for users to try it without leaving the app or opening a separate system. That platform placement mattered: the same kind of AI felt less like a separate innovation and more like another step in a familiar communication flow.

Standalone genAI tools were also used, but with different patterns. ChatGPT was the most common system mentioned, while Gemini appeared less often. Workers used standalone tools for grammar correction, email drafting, presentations, summaries, social media captions, trip planning, jokes, emotional support and content ideas. The uses often overlapped. A single ChatGPT session could involve search, drafting, rewriting and recommendation at once.

Participants did not treat standalone AI as a replacement for their work. Instead, they used it for first drafts, phrasing, structure or correction, then kept control through editing and selection. A marketing worker used ChatGPT regularly for emails, organizational writing and promotional lines. An architect used it to structure presentations, then modified the output. Others used it only once or occasionally when a specific need arose.

The study found that social influence shaped adoption. Workers often learned about ChatGPT or other tools through colleagues, friends, demonstrations or workplace cues. Peers helped reduce uncertainty by showing how to use the tools and what tasks they could support. In that sense, AI adoption was not only individual. It spread through practical examples and shared use scripts inside social and professional networks.

Recognition depended on interface, not frequency of use

The study found that people do not need to name something as AI in order to use it. Embedded AI was heavily used but rarely described as AI. On the other hand, chat-style assistants such as Meta AI, ChatGPT and Gemini were much more likely to be explicitly recognized and labeled as AI.

This recognition gap was not simply a matter of knowledge. It reflected interface design. When AI appeared as a separate chatbot or assistant, users were more likely to call it AI. When it appeared as a recommendation feed, search ranking or platform default, users often described it as the app, the algorithm or a tailored feed rather than as artificial intelligence.

The researchers identified a spectrum of AI recognition. Some participants explicitly named tools such as ChatGPT or Meta AI as AI. Others implicitly understood automated behavior without using the AI label. They described apps learning from them, showing tailored content or changing feeds based on activity. A third group misrecognized where AI was present, saying they did not use AI while still relying on algorithmically curated reels, feeds or search functions.

The study found that this naming pattern matters for AI measurement and governance. If researchers or policymakers ask only whether people use AI, they may miss a large share of embedded AI use that people experience as ordinary platform behavior. The paper argues that labels are weak evidence of engagement. Task placement and outcomes are stronger indicators of how AI actually enters everyday life.

This complicates both disruption and invisibility narratives. Standalone generative AI can appear disruptive because its assistant-style interface makes it visible and nameable. Embedded AI can appear invisible because it is packaged as a normal platform function. But in practice, both forms are often incorporated through the same logic: the tool remains in use when it fits the task and helps the user reach the next step.

That logic underpins the study's concept of interactional continuity. It has three main dimensions:

• task fit - refers to whether the tool helps complete the activity
• sequential placement - refers to where AI enters the task episode, such as before drafting, during rewriting, while searching or after receiving a suggestion
• interface packaging - refers to whether AI appears as a background default or a distinct assistant.

The researchers found that AI became routine at repeatable insertion points. Workers used it to draft, revise, search, select, summarize, plan or refine. It was most useful in tasks where the output could be quickly judged. If a rewritten message sounded better, it could be copied and sent. If a recommendation was useful, it could guide a decision.

If a presentation outline saved time, it could be adapted. But when tasks required stronger verification or carried higher consequences, users were more cautious. This helps explain why AI did not necessarily replace existing practices. It more often augmented them. Users still edited, selected, verified, ignored or abandoned AI outputs. They used systems as assistants, not as full substitutes for judgment.

AI adoption through routine, friction and trust

Participants were using AI in a platform-heavy environment where WhatsApp, Instagram and other digital tools already supported work coordination, social interaction and information seeking. In that setting, embedded AI and generative AI entered routines that were already shaped by mobile-first communication.

Meta AI's placement inside WhatsApp and Instagram was especially important. The study found that lower entry and exit costs made embedded assistants easier to try. Users did not have to move between apps, copy material into a separate browser or open a standalone platform. Convenience influenced adoption. One participant stopped using ChatGPT after WhatsApp AI became available because the embedded option reduced friction.

Standalone tools also retained value when they matched professional needs. ChatGPT became useful for workers who repeatedly needed polished writing, marketing copy, grammar support, summaries or structured content. The tool's usefulness depended less on novelty than on whether it could deliver a workable first pass.

The findings also show that AI use can extend beyond strictly work-related productivity. Some participants used generative AI for emotional reassurance, music recommendations or conversation when alone. Even these uses remained task-oriented in the study's analysis. The task was not always a formal workplace output; it could be stress relief, companionship or mood management. That broadens the understanding of AI use without treating it as a wholesale transformation of social life.

The study was based on a small, single-site urban office sample and relied on interview accounts rather than direct observation of live AI use. The findings are not presented as nationally representative. Instead, they offer a close look at how AI is narrated, recognized and inserted into everyday tasks in one Indian workplace.

The study calls for future research using observation, diaries, chat logs or screen recordings to examine how people sequence AI use in real time. Comparative research across regions, sectors and labor conditions could also test how far the concept of interactional continuity travels beyond this workplace setting.

Implications

AI governance should not focus only on highly visible tools such as ChatGPT. Embedded AI in feeds, search systems, messaging platforms and recommendation engines may shape behavior just as powerfully, even when users do not name it as AI. Regulation, transparency and accountability frameworks need to account for the continuum between background platform features and named assistants.

For media and technology researchers, the study offers a way to move beyond the debate over whether AI is disruptive or invisible. The authors argue that disruption and invisibility are not fixed properties of AI systems. They are outcomes of how tools are packaged, where they enter a task and whether users experience them as useful.

On the whole, the study insists that AI's social impact may be less dramatic at the point of arrival than in its repeated, practical incorporation into everyday decisions. In the Bangalore workplace, AI mattered when it helped workers draft a message, find information, polish a sentence, plan a trip, scan a feed or complete a routine more efficiently, showing that AI's influence often grows through continuity rather than rupture.