Are AI Companies Liable for Celebrity Deepfakes? Personality Rights and Intermediary Liability in India.

Introduction

A user enters a celebrity’s name into an artificial-intelligence platform. Within seconds, the platform generates a song in the celebrity’s voice, a video showing the celebrity endorsing a product, or a chatbot that imitates the celebrity’s mannerisms and personality.

When legal proceedings follow, AI companies increasingly advance a familiar defence:

We did not create the infringing content. We merely created a general-purpose tool which the user chose to misuse.

The argument is not without force. A technology provider should not necessarily become liable for every unlawful use of a product merely because the misuse was technologically foreseeable. Cameras can invade privacy, editing software can produce defamatory images and social-media platforms can carry unlawful content. The law has traditionally distinguished between the creator of a neutral tool and the person who uses it unlawfully.

Generative AI, however, complicates that distinction.

An AI company may train its system on a celebrity’s voice or images, advertise the ability to imitate famous personalities, recommend celebrity-related prompts, provide pre-built celebrity models, process the user’s instruction and produce the final synthetic output. It may then charge the user for the generation or monetise the resulting engagement.

At that point, describing the company as merely the provider of a passive tool may conceal the commercial and technological reality of the transaction.

The Delhi High Court’s decision in Hindware Ltd. v. Google LLC provides an important framework for examining this issue. The Court did not treat Google as a passive intermediary merely because an advertiser selected the infringing keyword. Instead, it examined the architecture of Google’s advertising system: its recommendation of keywords, auction mechanism, control over advertisement placement and monetisation of the trademark.

The emerging question for AI platforms is similar:

Is the platform merely supplying a general technological capability, or is it actively facilitating, structuring and profiting from the commercial exploitation of another person’s identity?

The answer cannot depend solely on who typed the prompt.

Personality rights in the age of generative AI

Indian law does not presently contain a single comprehensive statute defining and regulating personality or publicity rights.

Instead, protection has developed through a combination of:

• the right to privacy and dignity under Article 21 of the Constitution;

• the tort of passing off;

• trademark law;

• copyright and performers’ rights;

• defamation;

• breach of confidence; and

• judicial recognition of an individual’s commercial interest in distinctive attributes of their identity.

Indian courts have protected a person’s name, image, likeness, voice, signature, gestures, manner of speaking and other distinctive attributes from unauthorised commercial exploitation.

In Anil Kapoor v. Simply Life India, the Delhi High Court restrained the unauthorised use of Anil Kapoor’s name, likeness, image, voice and other elements of his persona through artificial intelligence, machine learning, deepfakes, face-morphing and similar technological tools. The Court connected such misuse not only with commercial harm but also with privacy, dignity and the individual’s right to control the exploitation of their persona.

In Arijit Singh v. Codible Ventures LLP, the Bombay High Court considered the unauthorised use of the singer’s personality attributes, including technologies capable of imitating his voice, image and manner of performance. The case recognised the distinctive risk posed by AI tools that allow users to appropriate the identity and expressive attributes of a performer.

The Delhi High Court has subsequently passed several orders protecting celebrities from AI-generated, morphed, obscene, misleading or commercially exploitative content. These cases reflect a growing judicial understanding that synthetic media can affect both the economic value of a person’s identity and their dignity and reputation.

The difficulty is no longer establishing that an unauthorised deepfake may violate the celebrity’s rights. The more difficult question is determining who is legally responsible for producing it.

The apparent simplicity of the “tool” defence

The central defence available to an AI company is that its technology is content-neutral.

The company may argue that:

1. the model has numerous lawful applications;

2. the user independently selected the celebrity;

3. the user framed the prompt and determined the purpose of the output;

4. the company did not approve or endorse the final use;

5. the output was automatically generated without human review;

6. the company’s terms prohibit impersonation and infringement; and

7. liability should fall on the person who created, published or commercially exploited the output.

This reasoning draws from a legitimate principle: the manufacture or supply of a dual-use technology should not, by itself, result in liability for every subsequent misuse.

A word processor is not ordinarily responsible for a defamatory letter. A video-editing application is not necessarily responsible for every manipulated clip. A general-purpose AI model capable of generating fictional voices should not automatically be liable merely because a user instructs it to imitate a real person.

The distinction is also important for freedom of expression. AI may lawfully be used for satire, parody, criticism, scholarship, artistic experimentation, accessibility, dubbing, historical recreation and other forms of expression. A prohibition broad enough to prevent any synthetic representation of a public figure could suppress constitutionally protected speech.

The technology-provider defence is therefore not inherently evasive. In many cases, it may be correct.

The problem arises when the company’s involvement is materially greater than the defence acknowledges.

The lesson from Hindware: examine the system, not merely the final command

In Hindware Ltd. v. Google LLC, Google argued that advertisers, rather than Google, selected the keywords that triggered sponsored advertisements.

The Delhi High Court did not confine its analysis to the advertiser’s final selection. It examined Google’s role in the complete commercial process, including the recommendation, valuation, auctioning and monetisation of the trademark HINDWARE.

The Court held that Google was not merely a passive intermediary in relation to its advertising programme. It had created and operated the commercial architecture through which the trademark was offered to competitors and used to target consumers interested in Hindware’s products.

The judgment does not directly decide the liability of generative-AI providers. Its reasoning nevertheless supplies a useful analytical method.

In an AI-personality case, a court should not ask only:

"Who entered the celebrity’s name into the prompt?"

It should also ask:

• Who collected or licensed the training material?

• Was the celebrity’s voice, face or style deliberately represented in the training data?

• Did the company create a dedicated voice, avatar or character based on that person?

• Did the platform recommend the celebrity’s name to users?

• Did it advertise that users could generate content “in the voice of” or “as” the celebrity?

• Did it provide prompt templates featuring the celebrity?

• Did the system refuse, discourage or facilitate impersonation?

• Did the company retain control over the model, output and distribution?

• Did it charge specifically for access to the celebrity simulation?

• Did it continue enabling the same conduct after being notified?

The legal character of the platform depends on the answers to these questions.

A user may initiate the generation, but that does not necessarily mean that the AI company played no role in producing the result.

A spectrum of AI involvement

AI-platform liability should not be treated as binary. Platforms may fall at different points on a spectrum.

1. The general-purpose neutral tool

At one end is a general-purpose model that:

• is capable of producing a broad range of lawful content;

• does not contain a dedicated celebrity model;

• does not advertise celebrity imitation;

• does not recommend celebrity prompts;

• has reasonable safeguards against impersonation;

• responds to valid complaints; and

• does not materially promote or profit from a particular person’s likeness.

The user may bypass safeguards or use external data to generate an infringing output.

In such a case, the platform’s argument that it merely supplied a general tool is comparatively strong. The user who deliberately generated and published the content is likely to bear primary responsibility.

2. The platform that knowingly permits predictable misuse

Further along the spectrum is a company that knows its system is routinely used to clone real voices or faces but takes minimal steps to address the problem.

Its model may be general-purpose, but the company may:

• ignore repeated complaints;

• decline to introduce technically feasible safeguards;

• allow searchable public galleries of celebrity deepfakes;

• monetise repeated generations involving the same person; or

• continue processing substantially identical infringing prompts after notice.

Here, liability may depend on knowledge, control, the feasibility of prevention and the adequacy of the platform’s response.

3. The platform that actively facilitates imitation

The case for liability becomes stronger where the platform:

• offers a named celebrity voice or avatar;

• provides a “celebrity voices” section;

• permits users to search for public figures;

• suggests popular celebrities through autocomplete;

• supplies pre-trained voice models;

• encourages users to create celebrity endorsements or songs;

• optimises the product for realistic impersonation; or

• commercially promotes examples featuring recognisable personalities.

The company is no longer merely failing to prevent misuse. It is facilitating the use that gives rise to the complaint.

4. The platform that directly commercialises the persona

At the strongest end of the liability spectrum is a platform that sells access to a particular celebrity’s digital replica.

For example, the platform may charge users to:

• speak to an unauthorised chatbot presented as the celebrity;

• create advertisements using the celebrity’s face;

• generate songs in the celebrity’s cloned voice;

• obtain personalised messages from a synthetic celebrity; or

• license the generated character for commercial campaigns.

The celebrity’s identity is then not incidental to the tool. It is the product being sold.

In such cases, the proposition that only the user exploited the celebrity’s persona becomes difficult to sustain.

Can the output amount to a violation of personality rights?

A synthetic representation may violate personality rights where it appropriates an identifiable attribute of a person for commercial advantage or creates a misleading association.

The claimant would ordinarily need to establish:

1. Identifiability: The output is recognisable as referring to the claimant.

2. Use of a protected attribute: The output reproduces or imitates the claimant’s name, image, voice, likeness, mannerisms or other distinctive aspect of persona.

3. Absence of consent: The use was not authorised by the claimant.

4. Commercial exploitation, deception or cognisable harm: The use appropriates endorsement value, suggests association, injures dignity or reputation, or causes another legally recognised form of harm.

Not every reference to a celebrity will satisfy these requirements.

A caricature of a politician, an obvious parody of an actor or an academic discussion of a singer’s vocal style may be recognisable without unlawfully appropriating the person’s commercial identity.

Context therefore matters.

The central inquiry should be whether the defendant has used the person’s identity as a commercial asset, falsely represented endorsement or association, or created harmful synthetic content that meaningfully interferes with privacy, reputation or dignity.

Direct liability of the AI company

An AI provider may face direct liability where its own conduct constitutes the actionable use.

This could arise where the company itself:

• creates or commissions the digital replica;

• trains and markets a named celebrity model;

• uses the celebrity’s likeness to promote the platform;

• generates sample content featuring the celebrity;

• sells access to the celebrity simulation;

• falsely represents that the celebrity has authorised the service; or

• materially alters and publishes the output through its own channels.

In these circumstances, the dispute is not principally about liability for third-party content. It concerns the company’s own commercial conduct.

This is the clearest application of the Hindware principle. Just as Google’s liability was examined through its role in creating and monetising the advertising system, an AI company may be directly responsible where it builds and monetises a system specifically organised around another person’s identity.

Calling the process automated does not necessarily alter this conclusion.

Automation may explain how the conduct occurred. It does not determine who legally performed it.

Secondary or contributory responsibility

The position is more difficult where the platform did not itself choose the celebrity but knowingly enabled the user to generate the output.

Indian personality-rights jurisprudence has not yet developed a comprehensive test for contributory liability of AI developers. Courts may therefore draw from principles found in intermediary law, copyright, passing off, authorisation and common-law participation.

Relevant considerations may include:

Knowledge

Did the company know that its service was being used to imitate the particular individual?

General awareness that a technology can be misused should not, by itself, be sufficient.

The stronger case arises from specific knowledge, such as:

• repeated complaints;

• prior court orders;

• public models named after the celebrity;

• prominent infringing outputs hosted on the company’s platform; or

• internal data demonstrating extensive use of the celebrity’s name.

Material contribution

Did the platform merely provide computational capacity, or did it materially assist the imitation?

A platform contributes more directly where it supplies the voice model, improves its fidelity, recommends the celebrity, removes watermarks, provides lip-syncing, or offers tools specifically intended to make the imitation more convincing.

Control

Could the company meaningfully prevent or limit the unlawful use?

A court may examine whether the provider controls:

• access to the model;

• prompt processing;

• model fine-tuning;

• publication of outputs;

• payment mechanisms;

• user accounts;

• search and discovery; and

• removal or disabling of celebrity-specific models.

The greater the control, the weaker the claim of total neutrality.

Commercial benefit

Did the platform derive revenue from the infringing activity?

General subscription revenue is less probative than revenue specifically linked to celebrity models, voice packs, avatar marketplaces or user engagement generated by deepfakes.

Financial benefit alone may not create liability. It can, however, reinforce the conclusion that the company was an active participant rather than an incidental service provider.

Is an AI provider an “intermediary”?

Section 79 of the Information Technology Act, 2000 provides conditional protection to intermediaries in respect of third-party information, data or communication links made available or hosted by them. The protection applies only when the statutory conditions are satisfied.

An AI company may argue that:

• the user supplied the prompt;

• the output is third-party information;

• the platform did not initiate the unlawful purpose;

• the generation process was automated; and

• the company complied with applicable due-diligence requirements.

The classification is not straightforward.

Traditional intermediaries transmit, store or host information supplied by users. Generative systems do something more: they process inputs through a trained model and create new outputs that did not previously exist in the same form.

The fact that a user prompted the output does not automatically establish that the output is merely “third-party information.” The model’s training, architecture, system instructions, safety settings and probabilistic generation process all contribute to the result.

Courts may therefore distinguish between:

• a platform that hosts a deepfake uploaded by a user; and

• a platform whose model generates the deepfake in response to the user’s instruction.

The former more closely resembles a conventional intermediary. The latter may be characterised as a joint producer, service provider or direct participant, depending on the degree of control and involvement.

The Hindware judgment similarly treated safe harbour as activity-specific. Google’s general role as a search intermediary did not necessarily protect its separate participation in a paid advertising programme that it structured and monetised.

An AI company may likewise be an intermediary for user-hosted content but a direct actor in relation to outputs produced and commercialised through its proprietary generation system.

The significance of actual knowledge

In Shreya Singhal v. Union of India, the Supreme Court read down Section 79(3)(b), holding that the relevant actual knowledge for compulsory removal ordinarily arises through a court order or notification by the appropriate government or its agency, rather than from intermediaries independently adjudicating the legality of millions of complaints.

AI companies may rely on this principle to argue that they cannot be required to determine every disputed claim of personality rights upon receiving a private notice.

That defence has force, particularly where the alleged infringement is ambiguous. A platform should not be compelled to remove criticism, satire, fan art or political expression merely because a public figure objects to it.

However, three qualifications are necessary.

First, Shreya Singhal concerned the operation of statutory intermediary safe harbour and restrictions on online speech. It does not necessarily immunise a company from liability for its own direct commercial conduct.

Second, where a platform has itself created or marketed a named celebrity model, the dispute may not be about third-party content at all.

Third, even where a court order is necessary to compel takedown under Section 79, a platform’s voluntary conduct after receiving detailed notice may still be relevant to knowledge, good faith, injunctive relief and assessment of its role.

A notice should therefore identify the specific model, prompt category, URL, output, account or feature complained of. Demanding that a platform prevent every possible future reference to a celebrity may be technologically unrealistic and legally overbroad.

The strongest defences available to AI companies

A balanced legal framework must preserve legitimate technological and expressive uses. AI companies may have several substantial defences.

1. The technology is capable of substantial lawful use

A general-purpose model should not be treated as unlawful merely because it can be used for impersonation.

The company may show that the system is primarily designed for lawful creative, educational, accessibility or enterprise functions and that celebrity misuse forms only a small part of its use.

This defence is strongest where the platform does not promote, organise or monetise the offending conduct.

2. The user independently caused the violation

The company may argue that the user:

• supplied the celebrity reference;

• uploaded the source recordings or photographs;

• bypassed safeguards;

• selected the misleading context;

• published the output; and

• used it commercially.

These facts may establish that the user is the principal wrongdoer.

They do not invariably eliminate platform liability, but they affect causation, responsibility and the appropriate relief.

3. No reasonable person would believe the output was authentic

Personality-rights claims are strongest where the content suggests endorsement, association or authenticity.

An obvious parody, stylised cartoon or absurd fictional scenario may not deceive a reasonable audience. Clear labelling that content is AI-generated may substantially reduce confusion.

Disclosure is not always a complete defence. A labelled deepfake may still commercially appropriate a voice or demean the individual. It is nevertheless an important factor.

4. Freedom of expression

Public figures cannot claim complete control over every reference, imitation or representation of their identity.

AI-generated works may be protected where they constitute:

• satire;

• parody;

• political commentary;

• criticism;

• news reporting;

• education;

• research;

• biography; or

• transformative artistic expression.

The Delhi High Court’s approach in the Jackie Shroff matter illustrates the need for this balance. While the Court protected the actor against unauthorised commercial exploitation, it did not treat every humorous or expressive use of his persona as automatically unlawful.

A personality right cannot become a general right to suppress unflattering or unwanted expression.

5. Nominative or descriptive reference

An AI platform may need to refer to a celebrity in order to identify the subject of news, commentary, comparison or research.

Merely processing or displaying a person’s name does not necessarily amount to commercial appropriation.

The legal concern is stronger where the identity itself becomes the product sold by the platform.

6. Consent or licence

The use may be covered by an agreement with the celebrity, their estate, a record label, film producer, photographer or other rights holder.

However, companies must verify the scope of the licence. Permission to use a recording for distribution may not necessarily include permission to train a voice-cloning model or create new synthetic performances.

7. Lack of identifiability

A generic output inspired by broad characteristics may not be sufficiently identifiable as a particular person.

The claimant should ordinarily demonstrate that the relevant audience would recognise the output as depicting or imitating them. A general visual style, accent, genre or vocal quality should not automatically become private property.

Copyright and performers’ rights may create separate claims

Personality rights are not the only legal concern.

Training or operating an AI system may involve the reproduction of photographs, recordings, films, performances or other copyrighted material.

A cloned voice itself may raise a personality-rights claim. The source recordings used to build the clone may separately be protected by:

• copyright in the sound recording;

• copyright in the underlying musical or literary work;

• performers’ rights;

• contractual restrictions; or

• moral rights.

The rights may belong to different persons.

A singer may control aspects of their personality and performance, while a record label may own copyright in a particular recording. A photographer may own copyright in an image even though the celebrity depicted has personality and privacy interests in its use.

An AI company may therefore defeat one claim but remain exposed under another.

The statutory exceptions under Section 52 of the Copyright Act may be relevant to research, criticism, review and other permitted uses. Their application to commercial AI training and synthetic generation remains contested and fact-sensitive. The law should not assume that every act of machine learning is infringing, but neither should it assume that technological processing places all training uses outside copyright.

Deception, passing off and false endorsement

A celebrity deepfake is particularly vulnerable to legal challenge where it falsely represents that the person endorses a product, investment scheme, political message or commercial service.

The traditional elements of passing off provide a useful framework:

• the celebrity has goodwill or reputation;

• the defendant makes a misrepresentation suggesting endorsement, authorisation or association; and

• the misrepresentation is likely to cause damage.

AI dramatically increases the scale and realism of such misrepresentation.

A synthetic advertisement may reproduce the celebrity’s appearance, voice, gestures and manner of speaking, creating a more persuasive false endorsement than a conventional photograph or text reference.

Where the AI provider offers a feature expressly designed to create such endorsements, direct liability becomes substantially more plausible.

Where the provider supplies only a general tool and the user independently creates and distributes the deceptive advertisement, primary liability may rest with the user, subject to the platform’s knowledge and participation.

Privacy and dignity cannot be reduced to commercial loss

Not every deepfake is created to sell a product.

Synthetic sexual content, defamatory videos, fabricated conversations and false political statements may cause severe dignitary harm even where no one directly profits from the output.

Indian courts have recognised that misuse of personality attributes can interfere with privacy, dignity and reputation in addition to commercial publicity rights. The orders protecting celebrities from morphed and obscene content demonstrate that the law is not confined to endorsement value.

For an AI provider, the risk is especially high where:

• the platform is designed for sexually explicit face-swapping;

• it permits non-consensual intimate imagery;

• it markets tools for humiliation or impersonation;

• it does not implement safeguards despite obvious and repeated abuse; or

• it continues to host or regenerate harmful content after a specific judicial direction.

In such cases, freedom of expression and neutral-tool arguments will carry substantially less weight.

What should courts examine?

Rather than imposing automatic liability or automatic immunity, courts should undertake a fact-specific inquiry.

The following factors offer a workable framework.

1. The purpose and design of the tool

Was the system designed as a broad creative tool, or specifically for cloning identifiable persons?

2. The source of the celebrity attribute

Did the user upload the source material, or did the company supply a pre-trained celebrity model?

3. Recommendation and discoverability

Did the platform recommend the celebrity, provide autocomplete suggestions or organise a searchable catalogue of celebrity replicas?

4. Degree of technical contribution

Did the company merely process a prompt, or did it train, fine-tune, enhance and optimise the imitation?

5. Commercial relationship

Was the celebrity’s identity used to attract subscribers, sell credits, generate advertising revenue or market the service?

6. Presentation of the output

Was the output labelled as synthetic? Did it falsely suggest authenticity, endorsement or association?

7. Nature of the use

Was it commercial advertising, fraud, satire, criticism, private experimentation, news reporting or artistic expression?

8. Knowledge and notice

Did the platform have specific knowledge of the infringing model or output? Was there a court order or governmental direction?

9. Ability to control the conduct

Could the provider disable the specific model, block repeated prompts, remove public outputs or suspend the responsible account?

10. Response after knowledge

Did the platform expeditiously remove the identified content, prevent regeneration and preserve relevant evidence?

No single factor should be determinative. Together, they help distinguish a neutral technological service from an active commercial participant.

The risk of overbroad injunctions

Courts must also be cautious when framing relief.

An order restraining every use of a celebrity’s name, voice, image or likeness may inadvertently prohibit lawful speech, including:

• journalism;

• documentary work;

• biographical content;

• satire;

• criticism;

• fan expression;

• search results; and

• legitimate comparative references.

Orders should ideally identify the unlawful category with reasonable precision.

A proportionate injunction may require a platform to:

• disable a specified celebrity model;

• remove identified outputs or URLs;

• prevent commercial endorsement generations;

• block synthetic sexual content;

• suspend repeat infringers;

• disclose account information subject to law;

• preserve generation logs; and

• act against substantially identical reuploads.

A blanket prohibition on every AI-generated representation of a person may exceed what is necessary to prevent the established harm.

This is particularly important because personality rights remain judicially developed and do not have clearly defined statutory boundaries in India.

What responsible AI companies should do

The absence of complete legal certainty does not prevent companies from adopting risk-reduction measures.

AI providers offering voice, image, video or avatar generation should consider:

• requiring verified consent before creating a model of an identifiable person;

• implementing additional checks for public figures and private individuals;

• prohibiting false endorsements and non-consensual intimate imagery;

• preventing celebrity names from being used as promotional keywords;

• disabling public marketplaces for unauthorised replicas;

• labelling synthetic outputs;

• embedding provenance information or durable watermarks;

• retaining prompt and generation logs for a reasonable period;

• offering a specific personality-rights complaint process;

• implementing repeat-infringer policies;

• distinguishing satire and expressive use from impersonation and fraud; and

• responding quickly to specific court orders.

These safeguards are not merely defensive. They may become central evidence that the company acted as a responsible technology provider rather than an active participant in the misuse.

The core distinction: capacity versus commercial exploitation

The law should not impose liability merely because an AI system has the technical capacity to imitate a human being.

Capacity is not the same as unlawful participation.

At the same time, an AI company should not escape responsibility merely by placing the user at the beginning of the causal chain.

The user may provide the prompt, while the company supplies the training material, celebrity model, recommendation system, generation process, payment infrastructure and publication mechanism. In such a case, the output may be the result of a commercially coordinated system rather than the independent act of a user.

The correct distinction is therefore not between human-generated and machine-generated content.

It is between:

• a general-purpose tool incidentally misused by a user; and

• a platform that actively facilitates or commercialises the unauthorised exploitation of a person’s identity.

Conclusion

The argument that an AI company “only provides a tool” should neither be automatically accepted nor automatically rejected.

It should be tested against the actual design and operation of the service.

Where a platform provides a general-purpose technology, implements reasonable safeguards, does not promote celebrity imitation and acts expeditiously upon legally valid notice, the case for intermediary or technology-provider protection is strong.

Where the platform creates celebrity-specific models, recommends famous personalities, advertises imitation capabilities, sells access to digital replicas or continues facilitating misuse despite specific knowledge, it becomes increasingly difficult to characterise the company as a passive intermediary.

The broader lesson from Hindware Ltd. v. Google is that courts must look beyond the final click or prompt.

A company’s liability depends not only on who initiated the transaction, but also on who designed it, encouraged it, controlled it and profited from it.

Generative AI does not require the abandonment of existing principles of intermediary liability. It requires those principles to be applied to the complete technological and commercial architecture of the platform.

The future of personality rights in the age of AI will therefore depend on a careful balance: protecting individuals against the extraction and manipulation of their identity, without giving celebrities a monopoly over public discussion, artistic expression or technological innovation.

The neutral-tool defence must remain available.

But it must remain a defence of fact, not a label chosen by the platform.