AI Law Brief | July 2026: Supreme Court sets aside orders citing hallucinated authorities; Delhi HC refuses interim injunction against OpenAI

In India, the Supreme Court set aside two tribunal orders that relied on hallucinated authorities, the Delhi High Court refused an interim injunction against OpenAI on model-training and output claims, and consultation closed on draft regulations for artificial intelligence (‘AI’) use in courts. Abroad, the European Union amended the AI Act and published guidance for its 2 August transparency regime, while US courts and agencies addressed training-data provenance, algorithmic pricing and AI-accuracy claims. We cover 12 developments across four jurisdictions.

India

Supreme Court sets aside tribunal orders that relied on hallucinated authorities

The Supreme Court delivered its reportable judgment in Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd. on 2 July 2026. The National Company Law Tribunal (‘NCLT’) had cited six purported authorities. On checking them, the Court found non-existent cases, real decisions attributed with paragraphs they did not contain, and a wrong citation to an existing judgment. The National Company Law Appellate Tribunal (‘NCLAT’) had not detected the problem on appeal.

The Court held that a decision into which even an ‘iota’ of fake or hallucinated precedent enters is no decision in the eyes of law, whether or not the material ultimately changed the result. It set aside both tribunal orders, restored the underlying section 7 application and directed the parties to maintain status quo while the NCLT decides it afresh. The Court also asked the Bar Council of India (‘BCI’) to constitute a committee and frame principles and disciplinary consequences for lawyers who submit fake material.

The judgment does not prohibit AI-assisted legal work. It expressly preserves legitimate assistance while requiring a human to remain in control at every stage. Courts, law firms and in-house teams should therefore require every suggested authority, quotation and pin-cite to be checked against an authoritative source, identify the human reviewer and retain a verification record. The judgment is operative; as at 3 August 2026, the BCI had not published the requested principles or disciplinary framework.

Delhi High Court refuses ANI’s interim injunction against OpenAI

The Delhi High Court issued a 135-page interim order in ANI Media Pvt. Ltd. v OpenAI OpCo LLC on 24 July 2026. At the prima facie stage, the Court held that it had territorial jurisdiction under section 62(2) because ANI’s principal place of business and registered office were within Delhi, and under section 20 because OpenAI targeted or offered services to users there and the alleged outputs were generated there. Separately, it rejected OpenAI’s extraterritoriality argument as to the training claim, treating US storage as the terminal step in a chain beginning with access to and transmission of ANI’s works from India and declining to sever training from the output claim.

On training, the Court prima facie found that storing ANI’s original literary works fell within the fair-dealing provision in section 52(1)(a) of the Copyright Act, 1957. On outputs, it found that the retrieval-augmented-generation (‘RAG’) examples were not substantially similar to the original works and that ANI had not established memorisation or regurgitation. The Court also considered OpenAI’s blocking and opt-out measures, the absence of demonstrated subscriber or syndication loss, public interest and the fact that damages could be quantified.

The Court dismissed the interim-injunction application, not the suit. Paragraph 274 confines the observations to that application and says they will not affect the final decision. Developers should document acquisition and purpose of training copies, opt-out handling, memorisation tests and the separate operation of RAG; publishers should preserve evidence of substantial similarity, regurgitation and market harm. The order should not be described as a final ruling that all AI training in India is lawful.

Consultation closes on Supreme Court’s draft regulations for AI use in courts

The Supreme Court published draft Regulations for Use of Artificial Intelligence in Courts, 2026 in June, and the extended consultation period closed on 15 July. The draft makes AI subservient to human judgment and leaves the judicial officer as the exclusive authority on law, fact and justice.

Subject to approval, supervision and verification, proposed uses include case management, certified transcription, human-verified translation, legal research, citation verification, anonymisation, accessibility and backlog analysis. Prohibited uses include algorithmic-only adjudication, AI-only findings, sentencing without a human in the loop, risk-scoring for bail or credibility, opaque systems materially affecting rights or liberty, behavioural profiling, undisclosed AI output as independent evidence and compromise of confidential deliberation.

The draft also proposes impact assessments, incident reporting, annual audits, an AI Register, vendor controls and sovereign or on-premise treatment of sensitive judicial data. It is not yet binding and would require later notification for the relevant court system. Court-technology vendors should nevertheless map their controls for explainability, audit access, data use, deletion, incident response and indemnity against the draft because these requirements may shape court procurement before a final text is notified.

Government reports 20 sovereign-model proposals and 93 lakh GPU hours sanctioned under IndiaAI Mission

The Ministry of Electronics and Information Technology (‘MeitY’) published a progress statement on the Safe & Trusted AI pillar on 27 July 2026. The Government said it had identified 20 indigenous sovereign-model proposals for support—12 large language models and eight small language models—and had empanelled 15 compute-service providers. It reported 237 supported projects and 93 lakh graphics-processing-unit (‘GPU’) hours sanctioned.

The same parliamentary answer records 13 approved Responsible AI projects covering subjects including bias mitigation, machine unlearning, privacy-preserving AI, explainability, deepfake detection and AI-risk assessment. It says 27 India Data and AI Labs have been established, work is continuing on another 188, and 58 AI Centres of Excellence are being established with State and Union Territory governments and industry partners. The IndiaAI Mission retains its five-year outlay of Rs. 10,371 crore.

These are programme and implementation figures, not new binding AI rules, completed models or proof that every sanctioned GPU hour has been consumed. Model developers receiving compute or public support should preserve the distinction between background and foreground intellectual property, document training-data rights and evaluation results, and align grant milestones with any claim that a model is sovereign, deployed or production-ready.

India and Japan set cooperation agenda for AI governance, compute and talent mobility

India and Japan issued a Joint Statement on cooperation in AI on 2 July 2026. The two governments agreed to continue their AI Strategic Dialogue and coordinate on international governance, model evaluation, capability assessment, cybersecurity, child safety and the implementation of the Hiroshima AI Process.

The industrial agenda covers secure data centres, GPU and semiconductor resources, economic-security review of the AI technology stack, multilingual and domain-specific models, and energy-efficient compute. The statement also welcomes an IIT Bombay–BharatGen–National Institute of Informatics memorandum for joint large-language-model research, a Sarvam–Preferred Networks arrangement across the AI stack, and IndiaAI–METI cooperation for AI-development companies. The two governments reaffirmed the goal of inviting 500 highly skilled AI professionals from India to Japan by 2030.

The Joint Statement is a bilateral policy commitment, not a treaty that harmonises domestic AI law or grants market access. Businesses should examine the terms of each memorandum or arrangement, including funding, data access, export controls and immigration. Cross-border model work will still need to allocate rights in datasets, weights, evaluations and improvements, as well as responsibility for cybersecurity and local-law compliance.

European Union

AI Omnibus enters into force and extends high-risk-system timelines

The European Union’s Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It amends the EU Artificial Intelligence Act (‘AI Act’) after the simplification package proposed in November 2025 and adopted in June 2026. The changes extend certain benefits previously reserved for small and medium-sized enterprises to small mid-cap companies, widen access to regulatory sandboxes and provide for an EU-level sandbox.

Chapter III, sections 1–3 of the AI Act, except Article 6(5), will apply from 2 December 2027 to high-risk systems classified under Article 6(2) and Annex III, and from 2 August 2028 to high-risk systems classified under Article 6(1) and Annex I. New Article 4a permits exceptional processing of special-category personal data for bias detection and correction only where strictly necessary. Providers of high-risk systems may do so for compliance with Article 10(2)(f) and (g); providers and deployers of other systems or models, and deployers of high-risk systems, must additionally show that the possible biases are likely to affect health or safety, negatively affect fundamental rights or lead to discrimination prohibited by EU law. Every specified safeguard must be met. The package also changes the AI-literacy duty, expands oversight and adds prohibitions addressing non-consensual intimate material and child sexual abuse material; those new prohibitions apply from 2 December 2026.

The Omnibus is enacted law, not the Commission’s earlier proposal. It does not postpone the AI Act as a whole. Providers and deployers should revise implementation schedules provision by provision: the extended high-risk dates do not displace transparency, general-purpose-model, existing prohibited-practice or other applicable duties. Providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content and were placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2).

Commission publishes Article 50 transparency guidance ahead of 2 August application

The European Commission published guidelines on the AI Act’s transparency obligations on 20 July 2026. The Article 50 duties began to apply on 2 August. Providers of interactive AI must design systems to inform people when they are interacting with AI, unless that interaction is obvious to a reasonably well-informed, observant and circumspect person, while providers of systems that generate synthetic content must enable detection through machine-readable marking.

Deployers face separate disclosures for deepfakes, AI-generated public-interest text published without human review or editorial control, and uses of emotion-recognition or biometric-categorisation systems. On 31 July, the Commission confirmed that AI Act enforcement would begin on 2 August and published a first list of more than 180 organisations that had signed the transparency Code of Practice. It states that Article 50 compliance will be enforced mainly by national competent market-surveillance authorities, with the AI Office performing a limited role for specified systems.

The guidelines explain the Commission’s interpretation; they do not replace the statutory text or make the voluntary code itself mandatory. Providers should identify which technical marking obligations attach to their output formats, including the Omnibus transition for existing generators, and deployers should build disclosure into the point of exposure. Contracts need to pass through the metadata, documentation and change notices a downstream deployer needs to comply.

Commission orders Google to open 11 Android features to rival AI services

The European Commission adopted binding specification measures under the Digital Markets Act (‘DMA’) on 16 July 2026. Under Article 6(7), Google must provide rival AI services with free and effective interoperability on terms equally effective to those available to its own services. The final AI-services decision covers 11 Android features.

The specified access includes invocation and always-on hotwords, centralised on-device and contextual inputs, actions across apps and the operating system, screen automation, system-level on-device models, implementation of third-party on-device models and background execution. The implementation timetable requires the measures in Android 18 and no later than 1 August 2027, with concurrent hotword detection for several services in Android 19 and no later than 1 August 2028.

These are binding specification decisions addressed to Google, not advisory guidance or a general right to bypass user consent. Indian assistant, app and on-device-model developers may gain deeper Android integration in Europe, but should track eligibility, documentation and testing access while meeting privacy, security and certification safeguards. Device and distribution contracts should preserve access parity and escalation rights during implementation.

United States

Federal court approves Anthropic’s USD 1.5 billion copyright settlement

The US District Court for the Northern District of California gave final approval in Bartz v Anthropic PBC on 20 July 2026 to a USD 1.5 billion plus interest, non-reversionary class settlement. The defined class covers copyright owners whose books appeared in copies of LibGen or PiLiMi downloaded by Anthropic and met the settlement’s registration and identifier criteria.

The order records direct notice to 594,945 potential class members associated with 482,374 works, 350 timely valid opt-outs covering 1,802 works, and two additional late opt-outs allowed for excusable neglect. The release covers defined past claims arising from Anthropic’s acquisition and use of covered copies as AI inputs, up to—but not including—any AI output. It does not release claims based on past AI outputs or claims of any kind based on conduct on or after 25 August 2025. Anthropic did not admit liability.

The settlement is binding on its class, but it is not a merits ruling that AI training is generally infringing or fair use, and USD 1.5 billion is not a fine. For developers and investors, the source of each training copy is a separate diligence question from the asserted purpose of training. Source-level licences, dataset hashes, piracy-source exclusions, chain of title, indemnities and claim reserves belong in technical and transaction diligence.

FTC seeks comment on proposed policy statement concerning AI accuracy

The Federal Trade Commission (‘FTC’) opened consultation on a proposed AI-accuracy policy statement on 1 July 2026. The draft explains how undisclosed manipulation of system outputs for ideological objectives could, in the Commission’s view, conflict with express or implied representations about an AI system’s objectivity, effectiveness or suitability and amount to unfair or deceptive conduct under section 5 of the FTC Act.

The proposal also states the Commission’s view that a conflicting state law may be impliedly pre-empted, referring specifically to the Colorado Artificial Intelligence Act. Comments were due by 31 July. The Commission voted 2-0 to authorise publication of the notice.

No final FTC policy statement was adopted in July, and the proposed pre-emption analysis is the agency’s legal position rather than a judicial ruling. AI providers should nevertheless align marketing, demonstrations and benchmark claims with actual system behaviour, disclose material steering or limitations, preserve substantiation and avoid treating a model’s probabilistic output as a guaranteed statement of fact.

DOJ files proposed decree restricting algorithmic rent coordination

The US Department of Justice (‘DOJ’) filed a proposed settlement with Willow Bridge on 6 July 2026 in its wider RealPage litigation. The Government alleges that competing landlords supplied non-public, competitively sensitive information to common rent-pricing algorithms and used shared rules that aligned pricing decisions.

The proposed decree would prohibit Willow Bridge from using an algorithm that relies on competitors’ sensitive information, sharing its own competitively sensitive information and attending specified competitor meetings organised by RealPage. It would require monitoring when Willow Bridge uses a third-party pricing algorithm that has not been certified as compliant, together with cooperation in the continuing case.

The complaint contains allegations, Willow Bridge has not admitted liability, and the decree remains subject to a 60-day Tunney Act comment period and court approval. Platforms, property managers, airlines, hotels and marketplaces using dynamic pricing should bar competitor-sensitive inputs, preserve independent decision-making, record recommendation overrides and make vendors auditable. Calling a tool ‘AI’ or outsourcing it does not isolate its user from competition-law scrutiny.

United Kingdom

DSIT opens evidence call on data regulation in the age of AI

The Department for Science, Innovation and Technology (‘DSIT’) opened a call for evidence on data regulation and AI on 15 July 2026. It seeks practical examples of how rules for personal and non-personal data interact with AI and other data-intensive technologies. Responses close at 11:59 pm on 9 September 2026.

The five themes cover access and use, data quality and downstream effects, governance across organisations, transparency and individual rights, and the effectiveness of the present framework. Questions address lawful bases, purpose limitation, data minimisation, special-category data, controller and processor allocation, automated decision-making, human involvement, AI agents acting with delegated authority, erasure and other data-subject rights. Copyright is outside the exercise’s scope.

The call does not amend UK data law or select a reform option. Indian processors, software providers and agent developers serving UK customers can use it to document concrete friction around role allocation, lawful basis, deletion propagation, delegated authority and model or data accuracy. Until government acts, existing controller, processor, security and rights obligations remain the compliance baseline.