India's media and entertainment industry crossed a threshold in 2025. The FICCI-EY report released on 24 March 2026 put the sector at ₹2.78 trillion, up 9% on the year, and recorded that digital media crossed ₹1 trillion for the first time to become the largest segment. Digital advertising grew 26% to ₹947 billion and now accounts for 63% of all advertising revenue, while linear television advertising fell 10%. Digital subscription revenue rose 60% to ₹163 billion. The report projects the industry at ₹3.3 trillion by 2028, with new media above half of revenue. An industry that earns most of its money from digital distribution is an industry whose margins will increasingly be set by software, and AI for media in India is now a question of operating economics rather than experiment.
The same shift has brought regulation closer. Publishers of streaming content have operated under the Digital Media Ethics Code since 2021. In February 2026 the government amended the Information Technology Rules to define synthetically generated content and to require labels and provenance markers on it. Personalisation runs on personal data that the Digital Personal Data Protection Act will govern in full from 2027. And the question of whether AI developers may train on Indian news and film without paying is now before both the Delhi High Court and a government committee.
This piece is for media and OTT executives, product leaders, investors and policymakers. Its contribution is a use-case map that places ten common applications of AI in media by their likely revenue impact and their regulatory exposure, so that a leadership team can see which ones to build now, which to build with controls, and which to leave until the rules settle.
The economics that make AI for media in India urgent
Three numbers in the FICCI-EY data explain where AI pays.
The first is subscriptions. The report counts 216 million paid video subscriptions across 143 million households. When a household holds more than one subscription, the competition is for attention within the home, and the lever that decides retention is whether the next thing the viewer sees is worth watching. Recommendation quality is a churn problem, and churn is a revenue problem.
The second is the shift of advertising to digital. With 63% of ad revenue now digital, the value of an impression depends on how precisely it can be matched to a viewer and measured. AI is already the engine of programmatic targeting on global platforms; Indian broadcasters and streamers that want to keep advertising margin must match that capability on their own inventory, which means building on first-party data they are permitted to use.
The third is the decline of linear. A 10% fall in linear television advertising pushes broadcasters to repackage their archives for digital, and the cost of doing so, tagging, clipping, subtitling and dubbing decades of content in several languages, is precisely the kind of work AI has made cheaper. The report also notes about 40 million connected television units in 2025, up from 30 million, which means more archive content is being watched on large screens where quality expectations are higher.
There is a cautionary data point in the same report. The video games segment fell 17% after the ban on real-money gaming took effect in August 2025. Regulation can remove a revenue line in a single notification, and media companies building AI products should price that possibility into their plans.
The regulatory layers a media AI product touches
Four overlapping regimes matter.
The Digital Media Ethics Code. Part III of the IT Rules, 2021 applies to publishers of online curated content. The Ministry of Information and Broadcasting's background note describes self-classification into five age-based categories (U, U/A 7+, U/A 13+, U/A 16+ and A), parental locks for content rated U/A 13+ or higher, reliable age verification for adult content, and a three-level grievance mechanism running from the publisher to a self-regulatory body to an oversight mechanism under the ministry. Any AI that recommends content, generates previews or assigns age ratings is operating inside these obligations. A recommender that surfaces an A-rated title to a profile without age verification is a compliance failure, not only a product bug.
The 2026 amendments on synthetic content. The IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, notified on 10 February 2026 by Gazette notification G.S.R. 120(E), define synthetically generated information as audio, visual or audiovisual content artificially created or altered to appear real, with routine edits, accessibility features and minor technical corrections excluded. As summarised by Outlook Business, the rules require visible labels, audio disclosures and embedded metadata or provenance markers where technically feasible, prohibit removal of those markers, require significant social media intermediaries to obtain a user declaration at upload on whether content is AI-generated and to verify it with technical measures, and cut the time to act on takedown notices for unlawful content to three hours. For platforms that host user-generated video, this converts AI detection from a trust-and-safety aspiration into a due-diligence requirement.
Data protection. Personalisation and ad targeting rely on viewing histories and device signals. The DPDP Act, 2023 and the Rules notified on 14 November 2025 require clear, purpose-specific consent notices and, for children, verifiable parental consent. The PIB explainer lists penalties of up to ₹200 crore for breaches of obligations relating to children. Family accounts on streaming services make this practical rather than theoretical: a recommender that profiles a child's viewing within a shared household account needs a lawful basis for doing so.
Copyright. Two developments frame the question of training data. On 8 December 2025 the Department for Promotion of Industry and Internal Trade released Part 1 of a working paper on generative AI and copyright proposing a hybrid model: a mandatory blanket licence for AI developers, royalties payable once models are commercialised at government-set rates open to judicial review, and centralised collection and distribution; the committee rejected opt-out models, and Storyboard18 reports a Nasscom dissent favouring a text-and-data-mining allowance with opt-out. Separately, in July 2026 the Delhi High Court, ruling on an interim injunction application in ANI v OpenAI, found that the use of ANI's content to train ChatGPT did not amount to infringement at that stage; the European Commission's IP Helpdesk notes the ruling does not determine the case as a whole. For media owners the two point in different directions. The working paper would make their archives a source of royalty income; the interim ruling suggests that, absent new legislation, the courts may not.
A use-case map by revenue impact and regulatory exposure
The table below is an original assessment for this piece. Revenue impact is rated by the size of the revenue pool the use case touches in the FICCI-EY data and by how directly it moves that pool. Regulatory exposure is rated by how many of the four regimes above apply and how severe the consequence of failure is. Ratings are judgments, stated so they can be argued with.
| Use case | Revenue pool it moves | Revenue impact | Regimes engaged | Regulatory exposure | Placement |
| Content recommendation and personalisation | Subscriptions (churn), viewing time | High | Ethics Code (age ratings), DPDP (profiling, children) | Medium to high | Build with controls |
| Ad targeting and dynamic ad insertion | Digital advertising | High | DPDP (consent, children) | Medium to high | Build with controls |
| Localisation: subtitling, dubbing, voice | Archive monetisation, regional reach | High | Copyright and performer consent for voice; SGI labels where voices are synthetic | Medium | Build now, with consent contracts |
| Metadata tagging and archive search | Archive monetisation, ad inventory | Medium | Few | Low | Build now |
| Clip, trailer and highlight generation | Marketing, short-form inventory | Medium | Ethics Code (rating of clips) | Low to medium | Build now |
| Piracy detection and rights monitoring | Protects subscription and licensing revenue | Medium | Few | Low | Build now |
| Moderation of user- generated content | Platform viability, advertiser safety | Medium (defensive) | 2026 SGI rules, takedown timelines, Ethics Code | High (mandatory) | Build because required |
| Synthetic performers and likeness generation | Production cost | Medium | SGI rules, copyright, personality rights, DPDP | High | Wait for clearer rules and consent norms |
| Automated news writing and summaries | Newsroom cost | Low to medium | Ethics Code for news publishers, defamation, SGI where audiovisual | High | Use only with editorial sign- off |
| Customer support and billing assistants | Retention, cost to serve | Low to medium | DPDP | Low | Efficiency play |
Four groupings fall out of the map.
Build now. Localisation, metadata tagging, clip generation and piracy detection combine meaningful revenue with low exposure. They work on content the company already owns, they do not require profiling viewers, and they raise the value of archives at exactly the moment linear decline makes archives matter. Localisation carries one caveat: dubbing with synthetic voices engages performer consent and, where the result is presented as real speech, the 2026 labelling rules. The contract with voice artists should say what may be cloned and for how long.
Build with controls. Personalisation and ad targeting sit in the most valuable revenue pools and the most demanding data regime. The controls are known: consent notices specific to profiling, separate handling of child profiles within family accounts, age-rating checks inside the recommender, and erasure of viewing histories on request. A separate PanScience Research piece on AI governance for boards covers how to report these controls upward.
Build because required. Moderation of user-generated content earns little directly, but after February 2026 platforms hosting such content cannot operate without it. Three-hour takedown timelines are not achievable with human queues alone at Indian scale, and the user-declaration and verification duties assume automated detection. The commercial question is whether to build or buy; a separate PanScience Research piece on AI security for enterprises covers adversarial manipulation of classifiers, which is a live risk for moderation systems.
Wait or proceed only with sign-off. Synthetic performers and automated news carry the highest exposure for modest gains. Likeness generation touches copyright, the personal rights of the performer whose likeness is used, the new labelling rules and data protection at once. Automated news risks defamation and misinformation liability on a timeline measured in hours. Neither is forbidden, but both are better approached after the copyright position settles and with a named editor accountable for every output.
How to read the map in practice
The map's most useful property is that it separates two kinds of decision that media companies often conflate. Whether a use case is worth doing is a commercial question. Whether it can be done safely is a governance question. A leadership team that discusses both in the same meeting tends to approve exciting, high-exposure projects and neglect dull, high-return ones. The map makes the dull, high- return work visible: archive tagging and localisation will rarely make headlines but sit squarely on the revenue lines FICCI-EY shows growing.
It also suggests a sequencing. Companies that build the low-exposure capabilities first, metadata, transcripts, multilingual subtitles, accumulate the structured data that personalisation and ad targeting later need. Rich metadata makes recommendations better without requiring more personal data about viewers, which is both a product gain and a privacy gain. A separate PanScience Research piece on small language models covers why a narrowly scoped model running on a company's own infrastructure is often the better choice for such work, especially where content rights forbid sending unreleased material to an external API.
Finally, the map is a snapshot. The DPIIT working paper is a proposal, the ANI ruling is interim, and the DPDP Rules phase in through May 2027. Two cells are most likely to move: synthetic performers, which could become a build-now item if consent norms and licensing settle, and archive licensing to AI developers, which does not appear in the table because it is not yet a market but would become a material new revenue line under the hybrid model.
What media leaders should do next
Audit the recommender against the Ethics Code. Check that age ratings flow into the recommendation and preview pipeline and that adult titles never surface to unverified profiles. This is a quick, high- value test.
Write consent into production contracts. Any voice, face or performance that might be used to train or drive a synthetic system should be covered explicitly, with duration, purpose and compensation stated.
Build provenance into the pipeline. If the company generates any synthetic audio or video, embed labels and metadata at creation rather than retrofitting them at distribution. The 2026 rules make provenance an operational requirement for intermediaries, and publishers that supply them will be asked for it.
Decide a position on training data. Media owners should decide, before a developer asks, whether they want to license archives, under what terms, and what they will do if the hybrid model becomes law. Waiting for the court case to finish is a decision too.
Conclusion
AI for media in India is no longer a laboratory topic. The FICCI-EY numbers show an industry whose revenue has moved to digital distribution, where recommendation, targeting and localisation decide margins. The regulatory perimeter has moved in parallel, with the 2021 Ethics Code, the February 2026 synthetic content rules, the DPDP Act and an unsettled copyright position each touching different use cases. The use-case map offers a way to separate value from exposure: build the low-exposure capabilities that unlock archives now, build personalisation and targeting with explicit controls, treat moderation as a cost of operating, and approach synthetic likeness and automated news only when rules and editorial accountability are in place.
Methodology and sources
Market figures are taken from EY India's press release for the FICCI-EY report "Stories, scale and impact: Unlocking India's media and entertainment economy", dated 24 March 2026 (ey.com/en_in/newsroom/2026/03/india-s-media-and-entertainment-sector-grew-9-percent-to-inr-2- point-78-trillion-in-2025-driven-by-digital-and-live-experiences-ficci-ey-report); the release does not discuss AI, so all links drawn here between those figures and AI use cases are the author's analysis. Digital Media Ethics Code provisions come from the Ministry of Information and Broadcasting background note on the IT Rules, 2021 (mib.gov.in). The 2026 amendments are described from Mondaq's summary for the notification date and gazette number (mondaq.com/india/gaming/1783896/2026-it-rules-amendments-ai-content-labelling-and-intermediary- obligations) and from Outlook Business for the definition, labelling and takedown provisions (outlookbusiness.com/amp/story/news/it-rules-amendment-2026-3-hour-takedown-rule-synthetic- content-defined-new-compliance-norms); readers should consult the gazette text for exact obligations. DPDP figures come from the PIB explainer of 17 November 2025 (static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf). The DPIIT working paper is described from Storyboard18 (storyboard18.com/digital/dpiit-releases- working-paper-on-copyright-and-ai-proposes-hybrid-licensing-model-85551.htm), and the ANI v OpenAI interim ruling from the European Commission IP Helpdesk note of 28 July 2026 (intellectual- property-helpdesk.ec.europa.eu/news-events/news/delhi-high-court-rules-anis-copyright-claims- against-openai-2026-07-28_en). The use-case map is an original, illustrative assessment; its ratings are judgments, not measurements. Nothing here is legal advice.