AI Could Make Extraction Cheaper Too

The most useful way to think about artificial intelligence is not as a friend or a threat, but as a cheapening machine. It makes some expensive things cheap: translation, coding help, medical information, lesson plans, image editing. But it also makes some harmful things cheap: fake video, personalized manipulation, synthetic celebrities, and infinitely replenished outrage feeds. The same general ability that can power a tutor can power a persuasion engine. The difference is not in the model. It is in what the surrounding system rewards.

Point C1 Generative AI dramatically lowers the cost of producing video, audio, text, and images at scale.

The New Economics of Content

Until recently, producing a convincing video, a polished article, or a realistic voice clone required time, money, and skill. Generative AI changes that equation. A single person with modest resources can now generate large volumes of text, images, audio, and video that look and sound professional. The Stanford HAI AI Index Report 2025 documents the rapid improvement and diffusion of these systems, while industry data shows training and inference costs falling sharply for many tasks.

This is not only an issue of volume. It is an issue of unit cost. When the cost of creating a persuasive video drops toward zero, the economics of attention change. Platforms that already optimize for engagement can now fill feeds with content that is cheaper to produce and more precisely tuned to keep users watching. The scarcity is no longer production capacity; it is the user’s attention itself.

The Indian context matters because the country combines cheap data, a large young population, and rapidly growing AI adoption. The same infrastructure that could deliver low-cost education and translation can also deliver low-cost distraction and deception. Which path dominates depends less on the technology than on the business models and design choices built around it.

Point C2 AI-powered recommendation can make feeds more personally addictive and harder to audit.

Personalization and the Black Box

Recommendation systems were already powerful. AI makes them more adaptive and harder to inspect. A modern feed can learn not just what categories a user likes, but the precise emotional triggers, pacing, and formats that extend session length. The result can feel less like a menu and more like a tailored environment designed to keep the user inside it.

This precision creates two problems. First, it can deepen the extraction of attention by delivering content that is harder to resist on a personal level. Second, it becomes harder for outsiders—including researchers, regulators, and even the user’s own future self—to understand why a particular piece of content appeared. The World Economic Forum’s Global Risks Report 2024 identifies AI-generated misinformation and the erosion of information integrity as major near-term risks, partly because the scale and personalization of synthetic content outpace verification capacity.

The opacity is not accidental. The most effective engagement systems are also the hardest to audit, because their logic depends on billions of user-specific signals that change by the hour.

flowchart LR
    A[User data] --> B[AI personalization]
    B --> C[Synthetic content]
    C --> D[Hyper-targeted feed]
    D --> E[Longer sessions]

Conceptual pipeline: when user data, generative synthesis, and personalized ranking combine, the cost of extending a session falls while the difficulty of auditing the system rises. Synthesized from Stanford HAI AI Index Report 2025, WEF Global Risks Report 2024, and Microsoft Digital Defense Report 2024.

Point C3 Deepfakes and synthetic influencers are already entering Indian political and consumer discourse.

Synthetic Media in the Wild

India is not waiting for the future to arrive. Synthetic media has already appeared in Indian elections, celebrity impersonation, and consumer marketing. Politicians’ faces and voices have been cloned and circulated, sometimes with disclaimers and sometimes without. Influencers who do not exist in real life promote products to real audiences. Microsoft’s Digital Defense Report 2024 tracks the expansion of influence operations and synthetic content globally, noting that the barrier to entry for well-resourced disinformation campaigns continues to fall.

These cases are still early. Not every synthetic video is a crisis, and not every AI influencer is harmful. But the direction is clear: the cost of creating credible-looking people and events is falling, while the cost of verifying authenticity is not falling as fast. In a country with dozens of languages, millions of first-time internet users, and limited digital-literacy infrastructure, the verification gap is especially consequential.

The harms are also unevenly distributed. People with less experience online, less time to check sources, and fewer trusted gatekeepers are more likely to encounter and believe synthetic content. This is not a technology problem alone; it is a design, literacy, and governance problem.

Point C4 The same technology can be designed toward substance if quality, provenance, and user control become first-class goals.

The Same Tools, Different Outcomes

None of this means AI must deepen extraction. The same cheapening of production can flood a feed, or it can flood a classroom with explanations in a student’s language. The same personalization can trap a user, or it can help a learner find the exact concept they are stuck on. The difference is what the system is optimized for: time-on-site, or learning; engagement, or trust; retention, or growth.

Building toward substance requires more than good intentions. It requires provenance, so users know whether an image, voice, or video is synthetic. It requires friction, so the most manipulative content is not also the most frictionless. It requires user control, so people can choose slower, more substantive modes without being overridden by defaults. And it requires accountability, so platforms cannot externalize the costs of synthetic deception onto society.

The earlier these guardrails are built, the cheaper they are. Waiting until synthetic media is the default format of political and commercial communication means retrofitting an ecosystem that has already been optimized for extraction.

Sources and Method

This article draws on the Stanford HAI AI Index Report 2025, the World Economic Forum Global Risks Report 2024, and Microsoft’s Digital Defense Report 2024. It also relies on public reporting about synthetic media in Indian elections and marketing, while treating specific incidents as illustrative rather than statistically representative. Claims about future risk and platform design are arguments, not predictions. The article’s goal is to widen the conversation about AI costs so that extraction is treated as seriously as productivity.

Related in This Series

• Attention, Substance, and the AI Moment — the full series guide and reading paths.
• What AI Makes Cheap — the productivity side of the same AI cost curve.
• Bhashini and the Indic-Language AI Moment — how AI translation could shift content toward knowledge, not just engagement.
• The Attention Extraction — the original diagnosis of how platforms capture and monetize attention.