Who this guide is for: AI platform operators serving Indian users, content platform compliance teams, cloud service providers, social media companies, and legal counsel advising on Indian digital regulation.
In effect since February 20, 2026. MeitY's 2026 amendments to the IT (Intermediary Guidelines) Rules slash the compliance window for removing synthetic content from 36 hours to 3 hours. All intermediaries must label AI-generated content. The definition of "intermediary" is broad: social media, messaging apps, cloud services, search engines, and AI tools that host or transmit user content. Global platforms serving Indian users are in scope.
Contents
1. What the Rules Require 2. Who Is Affected 3. Global Platform Exposure 4. SWT3 Procedure Mapping 5. The 3-Hour Window 6. Framework Mapping 7. Quick Reference 8. Quick Start References1. What the Rules Require
MeitY amended the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 with effect from February 20, 2026. The amendments introduce a dedicated framework for Synthetically Generated Information (SGI) and sharply tighten the obligations that intermediaries must meet to retain safe harbor protection under Section 79 of the IT Act.
Core Concept: Synthetically Generated Information (SGI)
SGI covers deepfakes, AI-generated or AI-altered images, video, voice cloning, and realistic algorithmically generated audio-visual content. The definition is intentionally broad. Any content where AI meaningfully contributed to its creation or modification may qualify as SGI.
Key Obligations
- 3-hour takedown window (Rule 3(1)(d)): Upon receipt of a lawful government order or court direction, intermediaries must act within 3 hours to remove or disable access to SGI content. The prior window was 36 hours.
- Mandatory visible AI content labeling: Platforms must "clearly and prominently label" AI-generated or synthetically generated content. The label must be visible to users at the point of consumption, not buried in metadata alone.
- Platform accountability: Failure to label SGI content or failure to comply with takedown orders can result in loss of safe harbor protection. Once safe harbor is lost, the intermediary becomes directly liable for the content it hosts or transmits.
- User reporting mechanisms: Platforms must provide accessible mechanisms for users to report SGI content they believe is harmful or misleading.
- Broad intermediary scope: The rules apply to social media platforms, video sharing services, streaming platforms, messaging applications, web hosting providers, search engines, cloud services, and AI tools that host or transmit user content.
2. Who Is Affected
| Entity Type | In Scope? | Key Obligation |
|---|---|---|
| Social media platforms | Yes | Label SGI, 3-hour takedown, user reporting mechanisms |
| Video and streaming platforms | Yes | Label SGI content, active content moderation |
| Messaging applications | Yes | Label SGI in forwarded and broadcast content |
| Cloud and hosting providers | Yes | Respond to takedown orders within 3 hours |
| Search engines | Yes | De-index flagged SGI content on order |
| AI tools (generative AI) | Yes | Label all outputs as AI-generated at point of delivery |
| Enterprise AI (internal use) | Conditional | In scope if outputs reach public-facing users or are distributed externally |
3. Global Platform Exposure
India has over 900 million internet users. Any platform serving Indian users is in scope regardless of where the company is headquartered or incorporated. The territorial scope of the IT Act extends to intermediaries whose services are accessed from Indian territory.
Loss of safe harbor under Section 79 of the IT Act is the nuclear penalty. It means the platform becomes directly liable for all user-generated content it hosts or transmits. At scale, this liability is existential. Maintaining safe harbor requires demonstrable, timely compliance with every obligation in the rules.
Compliance is not a policy matter alone. It requires technical infrastructure: content classification systems, labeling pipelines, takedown workflows, and verifiable evidence of response times. Regulators and courts will ask for proof, not assurances.
Indian courts have demonstrated willingness to order intermediaries to act regardless of the defendant's physical jurisdiction. Cross-border enforcement via mutual legal assistance and asset-level action against Indian operations is an established pattern.
4. SWT3 Procedure Mapping
The following SWT3 procedures provide cryptographic evidence of compliance with India IT Rules 2026 obligations. Each anchor is immutable, timestamped, and independently verifiable.
Transparency Record
India context: The mandatory labeling requirement means platforms must prove they labeled AI-generated content. Not just that labeling was configured in policy, but that it was applied to specific content at specific times. Configuration evidence is not sufficient; execution evidence is required.
SWT3 witnesses: AI-TRANS.1 records transparency actions at the moment they occur: what content was labeled, when the label was applied, and what disclosure was provided to users. Each anchor is cryptographically tied to the inference that produced the content, creating verifiable evidence of labeling compliance that cannot be retroactively altered.
AI-TRANS.1 anchors prove labeling was active and applied. Compare anchor timestamps against content publication timestamps to verify labels were applied before or at the moment of publication, not after a complaint was received.
Content Provenance Marking
India context: The rules require content to be "clearly and prominently" labeled. AI-WATERMARK.1 records both visible labels and embedded metadata, including C2PA provenance manifests and invisible watermarks. When a takedown request arrives, the watermark record proves whether the content was marked as AI-generated at creation time, which is the relevant moment for compliance.
SWT3 witnesses: Records the watermark type applied (visible label, C2PA manifest, steganographic embedding), the content hash at the time of marking, and the timestamp of the marking event. Creates an immutable record of provenance marking that survives any subsequent content distribution or re-encoding.
AI-WATERMARK.1 anchors prove provenance marking was applied at creation time. This distinction matters: marking applied after a complaint does not satisfy the rules. Look for anchors with timestamps preceding or matching the content publication event.
Inference Provenance
India context: When an AI system generates content that later becomes the subject of a takedown order, the inference provenance record shows exactly when the content was generated, by which model, and under what parameters. This is operationally critical for the 3-hour response window. Locating the source of specific content manually within 3 hours is not feasible at platform scale.
SWT3 witnesses: Every inference produces a Witness Anchor containing the model ID, input and output hashes, generation timestamp, and requesting context. The output hash can be matched directly against content in the takedown order, enabling rapid identification of origin.
AI-INF.1 anchors enable rapid content identification via hash lookup. Match the content hash provided in a takedown order against inference records to identify the source within minutes. This is how a 3-hour deadline becomes achievable at scale.
Consent and Disclosure
India context: The rules require user notification that they are interacting with AI-generated content. For generative AI platforms, this spans both input-side disclosure (the user understands they are using an AI system) and output-side disclosure (recipients understand the content they are viewing was AI-generated). Both disclosure moments require evidence.
SWT3 witnesses: AI-CONSENT.1 records consent and disclosure events: what was disclosed, to which user session, and at what timestamp. Creates verifiable evidence that users were informed about AI involvement at the point of generation and at the point of consumption.
AI-CONSENT.1 anchors prove disclosure was made. This is critical when a user claims they were not informed that content was AI-generated. The anchor provides a timestamped, immutable record of the disclosure event that can be produced in response to a regulatory inquiry or court proceeding.
Content Classification
India context: The SGI definition is broad. Platforms must classify content as synthetically generated or not, and the classification must be accurate. Failing to label AI content violates the rules and risks safe harbor loss. Falsely labeling authentic human-created content as AI undermines user trust and creates separate legal exposure. Accurate classification at scale requires a documented, auditable pipeline.
SWT3 witnesses: AI-MARK.1 records the classification decision for each content item: what content was evaluated, what classification was assigned (synthetic, human, or mixed), and the confidence level of the decision. Creates an audit trail of the classification pipeline's outputs over time.
AI-MARK.1 anchors demonstrate the classification pipeline was active and producing consistent decisions. Review confidence level distributions in anchor records to identify potential misclassification patterns before a regulator does.
5. The 3-Hour Window
The shift from 36 hours to 3 hours is not a modest tightening. It is a fundamental change in the operational model required to maintain compliance. Manual content identification, manual provenance tracing, and manual response workflows are not compatible with a 3-hour clock.
Without Governance Records
- Manual content search across production systems under time pressure
- Uncertain provenance: no way to prove when or how content was generated
- No verifiable evidence of response time if the deadline is met
- No audit trail if the regulator questions the response
With SWT3
- Instant content identification: AI-INF.1 output hashes enable lookup by content fingerprint, not by manual search
- Pre-existing provenance chain: AI-WATERMARK.1 records were created at generation time, not assembled in response to the order
- Auditable response timeline: AI-TRANS.1 records the takedown action with its own timestamp, creating evidence that the window was met
- Immutable record: All anchors are cryptographically sealed at creation time and cannot be backdated
SWT3 is the notary, not the enforcement mechanism. The platform's takedown infrastructure acts. SWT3 provides the cryptographic evidence record that proves the action was taken, when it was taken, and what content it addressed.
6. Framework Mapping
| Framework | Relevant Requirements | SWT3 Evidence |
|---|---|---|
| India IT Rules 2026 | Rule 3(1)(d) takedown window, SGI labeling obligation, safe harbor conditions | AI-TRANS.1, AI-WATERMARK.1, AI-INF.1 |
| EU AI Act Art. 50 | Transparency obligations for AI systems, AI-generated content marking requirements | AI-TRANS.1, AI-MARK.1 |
| California SB 942 | AI content detection tools requirement, visible and latent disclosure for AI-generated content | AI-WATERMARK.1, AI-TRANS.1 |
| DPDP Act (India) | Personal data protection obligations, consent requirements for data processing | AI-CONSENT.1, CJT jurisdiction fields |
7. Quick Reference
Questions an examiner or regulator is likely to ask, and how SWT3 anchors answer them.
| Examiner Question | SWT3 Evidence |
|---|---|
| Was this content labeled as AI-generated before it was published? | AI-TRANS.1 and AI-WATERMARK.1 timestamps vs. publication timestamp |
| When was the takedown order received and when did you act? | AI-TRANS.1 anchor records the response action with timestamp |
| Which model generated this content and when? | AI-INF.1 anchor: model ID, output hash, generation timestamp |
| Were users notified that this content was AI-generated? | AI-CONSENT.1 anchor: disclosure event, user session, timestamp |
| How does your classification pipeline decide what counts as SGI? | AI-MARK.1 anchor trail: classification decisions, confidence levels, pipeline version |
| Was the watermark or provenance tag present at creation, not added later? | AI-WATERMARK.1 timestamp predates distribution; immutable, cannot be backdated |
8. Quick Start
The jurisdiction CJT field tags every anchor with the applicable legal territory. Set it to "IN" for content subject to India IT Rules 2026. This field survives all clearing levels and is available for regulatory queries without requiring full anchor decryption.
from swt3_ai import Witnesswitness = Witness( tenant_id="your-tenant-id", api_key="axm_live_...", jurisdiction="IN", # India IT Rules 2026 scope legal_basis="legitimate_interest", purpose_class="content_moderation")# Witness the inference that produced AI-generated contentanchor = witness.observe( procedure="AI-INF.1", model_id="your-model-id", input_hash=hash_of_prompt, output_hash=hash_of_content, metadata={"content_type": "video", "sgi_classification": "synthetic"})# Record the labeling actionlabel_anchor = witness.observe( procedure="AI-TRANS.1", metadata={ "inference_anchor": anchor.fingerprint, "label_applied": "AI-generated content", "label_visible": True, "rule": "India IT Rules 2026 Rule 3(1)(d)" })witness.flush() # Send anchors to the SWT3 ledger
For TypeScript, Rust, C#, Ruby, and Swift implementations, see the SDK documentation. All SDKs support the jurisdiction CJT field with identical behavior.
References
- MeitY: Ministry of Electronics and Information Technology
- IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 and amendments
- Digital Personal Data Protection Act (DPDP Act), 2023
- EU AI Act Compliance Guide (Art. 50 transparency)
- California SB 942 AI Content Labeling Guide
- SWT3 SDK Documentation
- UCT Procedure Registry
- Create a free account to start witnessing