Silicon witnesses and fabricated truths: Admissibility of AI-Generated evidence in Indian and American courts
Silicon witnesses and fabricated truths: Admissibility of AI-Generated evidence in Indian and American courts
Silicon witnesses and fabricated truths: Admissibility of AI-Generated evidence in Indian and American courts
Silicon witnesses and fabricated truths: Admissibility of AI-Generated evidence in Indian and American courts
► "Absent reform, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern."
When a Video No Longer Proves What Happened: Why AI-Generated Evidence Is the Most Dangerous Evidentiary Challenge of Our Time
For centuries, courts operated on a foundational premise: evidence reflects reality. A photograph captured what the camera saw. A video depicted what occurred. A voice recording preserved what was said. Generative artificial intelligence has shattered this premise. Today, a photograph may depict an event that never occurred. A video may portray a person committing a crime that exists only in the imagination of an algorithm. A voice recording may reproduce speech that was never uttered. This article undertakes a comparative doctrinal analysis of how Indian and American courts are grappling with this evidentiary rupture, examines the inadequacy of current authentication standards in both jurisdictions, surveys landmark cases and legislative developments, and proposes a four-pillar reform framework requiring a shift from container-based to algorithm-based evidentiary inquiry.
The challenge is not merely theoretical. In Washington v. Puloka (2024), one of the first formal judicial decisions on AI-processed criminal evidence, a Washington State court was asked to admit an AI-enhanced version of a smartphone video from a triple homicide scene. The court excluded it, finding the AI tool untested, the operator unqualified, and the risk of juror confusion unacceptably high. In India, courts have begun to grapple with AI in nascent but significant ways, while the newly enacted Bharatiya Sakshya Adhiniyam, 2023, which replaced the colonial-era Indian Evidence Act, modernised digital evidence rules yet left a conspicuous legislative gap on AI-generated outputs.
This article proceeds in five parts. Part II analyses the taxonomy of AI evidence. Part III examines the American legal framework. Part IV analyses the Indian framework. Part V proposes a four-pillar reform. Part VI concludes.
Part II: How AI Evidence Is Classified and Why the Distinction Between Acknowledged and Unacknowledged AI Evidence Is Legally Critical
AI-generated evidence divides into two categories that require fundamentally different legal responses. The first, acknowledged AI evidence, encompasses material about which there is no dispute that an AI system created or processed it: facial recognition outputs, forensic AI analyses, algorithmic crime predictions, and AI-enhanced surveillance footage. The inquiry here is about reliability: was this AI system tested, validated, and methodologically sound?
The second, unacknowledged AI evidence or deepfakes, involves material where one party claims authenticity and the opposing party alleges AI fabrication. The inquiry here is about authenticity: is this evidence real at all? These two inquiries are epistemologically distinct. Conflating them, as courts in both jurisdictions have tended to do, produces analytical confusion and wrong outcomes. A deepfake is not merely unreliable evidence; it is fabricated evidence, and the legal framework must treat the distinction seriously.
The table below illustrates the fundamental difference between the two categories and the legal question each raises.
Category
Definition
Core Legal Question
Examples
Acknowledged AI Evidence
Material where both parties accept that AI created or processed it
Is this AI system reliable, validated, and methodologically sound?
Facial recognition output, forensic AI analysis, algorithmic crime prediction, AI-enhanced surveillance
Unacknowledged AI Evidence (Deepfakes)
Material where one party claims authenticity and the other alleges AI fabrication
Is this evidence genuine at all, or is it a fabrication designed to deceive?
AI-generated audio of a person confessing, synthetic video of a person at a crime scene, fabricated photographic evidence
Undergirding both categories is the structural opacity of modern AI systems. Unlike fingerprinting or DNA analysis, which operate on principles that can be explained, tested, and cross-examined, deep learning models produce outputs through millions of weighted parameters that even their creators cannot fully explain. This opacity creates a fundamental tension with the evidentiary requirement of verifiability. A court system that relies on lay jurors to determine whether a video is authentic or AI-fabricated is a court system that has outsourced its fact-finding function to a technology it cannot evaluate. The right to confront and challenge evidence is rendered hollow when the evidence is generated by a process that resists interrogation.
Part III: The American Legal Framework and Why Existing Rules Are Insufficient for AI Evidence
Existing Standards and Their Inadequacy
American courts have applied existing frameworks, Rule 901 on authentication, Rule 702 on expert testimony, and the gatekeeping standard of Daubert v. Merrell Dow Pharmaceuticals (1993), to AI evidence by analogy. Washington v. Puloka illustrated the limits of this approach. The defendant sought to introduce AI-enhanced footage as exculpatory evidence. The court excluded it because the tool had no documented testing history, no peer-reviewed validation, and was operated by a filmmaker with no forensic AI training. While Daubert's flexible standard is capable of addressing the reliability of acknowledged AI outputs, it provides no clear framework for the far harder challenge of unacknowledged AI manipulation, deepfakes that are designed to deceive even trained observers.
Proposed Federal Rule 707 and State Developments
Federal rulemakers have responded with unusual speed. After an 8-1 vote by the Advisory Committee in May 2025, proposed Rule 707 was released for public comment through February 2026. The Rule would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702, requiring the proponent to show: sufficient facts or data as input; reliable AI principles and methods; and reliable application of those methods to the case facts. The Rule's purpose is to prevent parties from circumventing Rule 702 by offering machine outputs directly rather than through a cross-examinable human expert.
The table below compares the existing American framework with the proposed Rule 707.
Framework
Provision
Application to AI
Key Limitation
Authentication
Federal Rule of Evidence 901
Requires showing evidence is what it purports to be
Designed for documents that passively capture reality; cannot address AI fabrication
Expert Testimony
Federal Rule of Evidence 702
Daubert standard requires reliability, testability, peer review
Not designed for machine outputs presented without a human expert sponsor
Proposed AI Evidence Rule
Proposed Federal Rule 707
Applies Rule 702 standard to acknowledged AI evidence; requires sufficient input data, reliable methods, reliable application
Applies only to acknowledged AI evidence; does nothing for deepfakes presented as genuine
State Frameworks
California SB 11, Louisiana Act 2025
Condition admissibility on independent expert verification and documented reliability testing
Fragmented; no uniform national standard
Rule 707 is a significant step but an incomplete solution. Critically, it applies only to evidence the proponent acknowledges was created by AI. It therefore does nothing for the deepfake problem, the case where fabricated evidence is presented as genuine. Louisiana and California have also enacted or proposed state-level frameworks conditioning admissibility on independent expert verification and documented reliability testing, demonstrating that legislative action is both feasible and urgent.
Part IV: The Indian Legal Framework and the Conspicuous Gap Left by the BSA 2023
From Section 65B to BSA 2023: Modernisation Without Resolution
Indian evidence law's engagement with digital evidence was shaped by two Supreme Court decisions. In Anvar P.V. v. P.K. Basheer (2014) 10 SCC 473, the Court held that electronic evidence without a Section 65B certificate is inadmissible as secondary evidence, a position reaffirmed in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020) 7 SCC 1. The certificate requirement became the gatekeeping mechanism for all digital evidence in India. The Bharatiya Sakshya Adhiniyam, 2023 preserved this architecture in Sections 61 to 63, modernising the language around electronic records while retaining the same certificate-based authentication framework.
The problem is that this framework rests on assumptions that AI evidence does not satisfy.
The table below identifies the assumptions of the BSA framework and the ways AI evidence fails each of them.
BSA Framework Assumption
Satisfied by Traditional Digital Evidence?
Satisfied by AI-Generated Evidence?
Traceable human author
Yes
No; AI outputs have no singular human author
Identifiable originating device
Yes
No; cloud-based AI models produce outputs across distributed systems
Clear chain of custody
Yes
No; training data, model weights, and inference process resist conventional custody documentation
Record of a real event
Yes
No; a deepfake is a fabrication masquerading as a record
Certificate by a responsible official
Yes
Unclear; no provision addresses who certifies AI-generated output
The BSA contains no definition of "AI-generated evidence," no authentication framework for machine-generated outputs, and no provision addressing deepfakes or synthetic media. The question of whether a deepfake even constitutes an "electronic record" within the meaning of the BSA is itself unresolved: a deepfake is not a record of an event but a fabrication masquerading as one.
Section 39, Constitutional Angles, and Current Judicial Engagement
Section 39 of the BSA, governing expert opinion, offers a potential doctrinal foothold. Its reference to "any other field" of expertise may be interpreted to encompass validated AI forensic systems. However, this has never been judicially confirmed, and Section 39 provides no guidance on algorithmic bias, training data opacity, or the structural impossibility of cross-examining a machine. In Jaswinder Singh v. State of Punjab (CRR-2907-2023, Punjab and Haryana High Court, 2023), the court referenced ChatGPT as an analytical research tool in a bail application, taking care to clarify that it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
The constitutional dimensions are significant and underexplored. AI-powered facial recognition and voice identification create direct tension with Article 20(3), the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017) 10 SCC 1. When a court convicts on the basis of facial recognition with a documented error rate, the accused has not received a fair trial; they have received a statistical sentence. These constitutional questions remain unanswered by Indian jurisprudence.
► Key Principle: The BSA 2023 contains no definition of AI-generated evidence, no authentication framework for machine outputs, and no provision on deepfakes. This is not a minor oversight. It is a structural legislative gap that leaves Indian courts without a principled basis for evaluating the most consequential category of evidence they will face in the coming decade.
Part V: Comparative Assessment and the Four-Pillar Reform Framework
A comparative analysis reveals that both the American and Indian frameworks suffer from the same conceptual deficiency: they approach AI evidence through a container-based lens, treating it as a species of electronic record subject to authentication standards designed for documents that passively capture reality. This framework is fundamentally ill-suited to evidence that actively constructs or fabricates reality. The authentication question must change from "Is this file authentic?" to "Is this algorithm reliable, transparent, and free from adversarial manipulation?" These are epistemologically distinct inquiries, and conflating them is not a minor technical error. It is a structural failure of judicial gatekeeping.
The table below sets out the four-pillar reform framework proposed in this article and its application to both jurisdictions.
Pillar
Reform Proposed
India Application
US Application
Pillar 1: Statutory Definition
Enact statutory provisions defining "AI-generated evidence" as a distinct evidentiary category with disclosure requirements covering AI system architecture, training data, error rates, and validation history
Introduce a dedicated provision within the BSA 2023; define AI-generated evidence and synthetic media expressly
Enact Rule 707 with mandatory disclosure schedules; extend beyond acknowledged AI to cover unacknowledged fabrication
Pillar 2: Algorithm-Centred Authentication
Shift from authenticating the storage device to interrogating the scientific validity of the algorithm through Daubert-style reliability hearings examining training data integrity, model validation, false positive rates, and susceptibility to adversarial inputs
Adapt BSA Section 39 expert framework to require AI system validation evidence before any AI output is admitted
Strengthen proposed Rule 707 to mandate pre-admission reliability hearings for all AI-generated outputs
Pillar 3: Burden-Shifting for Deepfakes
For unacknowledged AI evidence, the challenger must first present evidence sufficient for a reasonable factfinder to conclude the evidence may have been AI-fabricated; the burden then shifts to the proponent to establish authenticity by a preponderance
No existing provision; requires new BSA amendment or dedicated AI evidence Act
No existing Federal Rule addresses deepfakes presented as genuine; requires extension of Rule 707 or new Rule
Pillar 4: Institutional Safeguards
Develop bench guidance for judges, AI forensic certification protocols, and mandatory metadata disclosure for digital submissions; adopt EU AI Act Article 13 transparency obligations as a model
NCLT, High Courts, and trial courts require dedicated guidance on AI evidence evaluation; forensic AI certification for expert witnesses
Federal Judicial Center guidance needed; state bar AI forensic certification programmes; mandatory metadata disclosure rules
Part VI: Conclusion
The admissibility of AI-generated evidence presents evidentiary law with its most demanding test since the introduction of DNA evidence. Unlike DNA, however, AI-generated evidence may be designed to deceive. Both the United States and India possess the doctrinal raw material for reform: Rule 707, state-level statutes, the BSA's modernised digital evidence framework, and a rich constitutional jurisprudence on privacy and fair trial rights. What both lack is a unified, algorithm-centred evidentiary framework that treats AI-generated outputs as a distinct category requiring categorical rules, not case-by-case improvisation.
Absent these reforms, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern.
Frequently Asked Questions (FAQs) on AI-Generated Evidence in Indian and American Courts
1. What is AI-generated evidence and why does it pose a unique legal challenge? AI-generated evidence refers to text, audio, video, images, or analytical outputs created or significantly altered by artificial intelligence systems. It poses a unique legal challenge because, unlike traditional evidence that passively records reality, AI can actively fabricate reality with a realism that defeats conventional authentication methods and deceives even trained observers.
2. What is the difference between acknowledged and unacknowledged AI evidence? Acknowledged AI evidence is material where both parties accept that an AI system created or processed it; the legal inquiry is about reliability. Unacknowledged AI evidence or deepfakes is material where one party claims authenticity and the other alleges fabrication; the legal inquiry is about whether the evidence is genuine at all. These are epistemologically distinct questions requiring different legal frameworks.
3. What did the Washington v. Puloka case decide about AI evidence? In Washington v. Puloka (2024), a Washington State court excluded AI-enhanced smartphone footage from a triple homicide case because the AI tool had no documented testing history, no peer-reviewed validation, and was operated by a person with no forensic AI training. The case illustrates both the limits of applying existing authentication standards to AI evidence and the importance of judicial gatekeeping.
4. What is Proposed Federal Rule 707 and what does it cover? Proposed Federal Rule of Evidence 707, released for public comment after an 8-1 Advisory Committee vote in May 2025, would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702. It requires the proponent to demonstrate sufficient input data, reliable AI methods, and reliable application of those methods to the case facts. Critically, it applies only to acknowledged AI evidence and does not address deepfakes presented as genuine.
5. How does the Bharatiya Sakshya Adhiniyam 2023 address AI-generated evidence? The BSA 2023 modernised India's digital evidence framework through Sections 61 to 63 and retained the certificate-based authentication framework from the old Indian Evidence Act. However, it contains no definition of AI-generated evidence, no authentication framework for machine outputs, no provision on deepfakes, and no guidance on the structural impossibility of cross-examining an algorithm. This is a significant legislative gap.
6. What constitutional concerns does AI evidence raise in India? AI-powered facial recognition and voice identification create direct tension with Article 20(3) of the Constitution, the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017). A conviction based on facial recognition with a documented error rate raises serious fair trial questions that Indian jurisprudence has not yet resolved.
7. What is the black-box problem in AI evidence? The black-box problem refers to the structural opacity of modern deep learning AI systems, which produce outputs through millions of weighted parameters that even their creators cannot fully explain. Unlike fingerprinting or DNA analysis, AI reasoning cannot be fully explained, tested, or cross-examined, creating a fundamental conflict with the evidentiary requirement of verifiability and the accused's right to confront and challenge evidence.
8. What are the four pillars of the reform framework proposed in this article? The four pillars are: first, statutory definition of AI-generated evidence as a distinct evidentiary category with mandatory disclosure requirements; second, algorithm-centred authentication through Daubert-style reliability hearings examining training data, model validation, and false positive rates; third, a burden-shifting mechanism for deepfakes where the challenger raises a reasonable fabrication concern and the burden shifts to the proponent to prove authenticity; and fourth, institutional safeguards including judicial guidance, AI forensic certification protocols, and mandatory metadata disclosure.
Key Takeaways
Generative AI has shattered the foundational evidentiary premise that evidence reflects reality; courts in India and the United States now face AI-generated text, audio, video, and analytical outputs for which existing evidence rules were not designed.
AI-generated evidence divides into two legally distinct categories: acknowledged AI evidence, where the inquiry is reliability, and unacknowledged AI evidence or deepfakes, where the inquiry is authenticity. Conflating these categories produces analytical confusion and wrong judicial outcomes.
The black-box opacity of modern deep learning models creates a fundamental conflict with the evidentiary requirement of verifiability, rendering the right to confront and challenge evidence hollow when the evidence was generated by a process that resists interrogation.
In Washington v. Puloka (2024), a US court excluded AI-enhanced footage because the tool was untested, the operator unqualified, and the risk of juror confusion unacceptably high, illustrating both the limits of existing authentication standards and the possibility of principled judicial exclusion.
Proposed Federal Rule of Evidence 707, released for public comment in 2025 after an 8-1 Advisory Committee vote, would apply the Rule 702 expert testimony standard to AI-generated evidence but applies only to acknowledged AI evidence and leaves the deepfake problem entirely unaddressed.
The Bharatiya Sakshya Adhiniyam, 2023 modernised India's digital evidence framework but left a conspicuous legislative gap: no definition of AI-generated evidence, no AI-specific authentication framework, and no provision addressing deepfakes or synthetic media.
The BSA's certificate-based authentication framework rests on assumptions (traceable human author, identifiable device, clear chain of custody, record of a real event) that AI-generated evidence systematically fails to satisfy.
In Jaswinder Singh v. State of Punjab (2023), the Punjab and Haryana High Court referenced ChatGPT as a research tool in a bail application while clarifying it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
AI-powered facial recognition and voice identification create direct constitutional tension with Article 20(3) and Article 21 of the Indian Constitution, raising fair trial questions that Indian jurisprudence has not yet resolved.
Both jurisdictions require the same four-pillar reform: statutory definition of AI evidence, algorithm-centred authentication replacing container-based authentication, burden-shifting for deepfakes, and institutional safeguards including judicial guidance, AI forensic certification, and mandatory metadata disclosure.
References
Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473 (India): The foundational Supreme Court decision requiring a Section 65B certificate for the admissibility of electronic evidence as secondary evidence in Indian courts.
Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1 (India): Reaffirmed and clarified the Section 65B certificate requirement, cementing the certificate-based authentication architecture that the BSA 2023 has now inherited.
Jaswinder Singh v. State of Punjab, CRR-2907-2023 (Punjab and Haryana High Court, 2023): The first significant Indian judicial reference to a generative AI tool, with the court referencing ChatGPT as a research aid while explicitly clarifying it had no influence on the bail order.
K.S. Puttaswamy v. Union of India, (2017) 10 SCC 1 (India): The nine-judge bench decision recognising the right to privacy as a fundamental right under Article 21, directly relevant to the constitutional implications of AI-powered facial recognition and voice identification as evidence.
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993): The foundational US Supreme Court decision establishing the Daubert reliability standard for expert testimony and scientific evidence under Federal Rule of Evidence 702, now being extended by analogy to AI-generated evidence.
Frye v. United States, 293 F. 1013 (D.C. Cir. 1923): The predecessor general acceptance standard for scientific evidence, still applicable in several US states and relevant to the historical development of evidentiary standards for novel scientific methods.
Washington v. Puloka, No. 22-1-03036-5 SEA (Washington Superior Court, King County, March 2024): One of the first formal judicial decisions on AI-processed criminal evidence; excluded AI-enhanced footage for lack of documented testing, qualified operator, and unacceptable juror confusion risk.
Bharatiya Sakshya Adhiniyam, 2023, No. 47 of 2023 (India), Sections 39, 61 to 63: The primary Indian evidence statute governing expert opinion and electronic records, whose certificate-based authentication framework does not address AI-generated evidence or deepfakes.
Indian Evidence Act, No. 1 of 1872 (India), Section 65B (repealed 2023): The predecessor provision on electronic evidence certificates, whose architecture was carried forward into the BSA 2023.
Information Technology Act, No. 21 of 2000 (India), Sections 79A and 85B: Relevant supplementary provisions on electronic evidence and digital examination in the Indian statutory framework.
Constitution of India, Articles 20(3) and 21: The constitutional provisions on the right against self-incrimination and the right to life and personal liberty, both directly engaged by AI-powered identification evidence in criminal proceedings.
Federal Rules of Evidence, Rules 403, 702, and 901: The existing US evidentiary framework being applied by analogy to AI-generated evidence, and the foundation upon which Proposed Rule 707 builds.
Proposed Federal Rule of Evidence 707 (comment period closed February 16, 2026): The proposed US federal rule that would apply the Rule 702 expert testimony standard to AI-generated evidence; applies only to acknowledged AI evidence.
California Senate Bill 11, 2025-26 Session (California, 2025): State-level legislative framework conditioning AI evidence admissibility on independent expert verification and documented reliability testing.
Louisiana Act No. (effective August 1, 2025): State-level AI evidence framework requiring reliability testing documentation as a condition of admissibility.
Disclaimer
This article is published by CLEAR LAW (clearlaw.online) strictly for educational and informational purposes only. It does not constitute legal advice, legal opinion, or any form of professional counsel, and must not be relied upon as a substitute for consultation with a qualified legal practitioner. Nothing contained herein shall be construed as creating a lawyer-client relationship between the reader and the author, publisher, or CLEAR LAW (clearlaw.online).
All views, interpretations, and conclusions expressed in this article are solely those of the author and represent independent academic analysis. CLEAR LAW (clearlaw.online) does not endorse, verify, or guarantee the accuracy, completeness, or reliability of the content, and expressly disclaims any responsibility for the same.
While reasonable efforts are made to ensure that the information presented is accurate and up to date, no warranties or representations, express or implied, are made regarding its correctness, adequacy, or applicability to any specific factual or legal situation. Laws, regulations, and judicial interpretations are subject to change, and the content may not reflect the most current legal developments.
To the fullest extent permitted by applicable law, CLEAR LAW (clearlaw.online), the author, editors, and publisher disclaim all liability for any direct, indirect, incidental, consequential, or special damages arising out of or in connection with the use of, or reliance upon, this article.
Readers are strongly advised to seek independent legal advice from a qualified professional before making any decisions or taking any action based on the contents of this article. Reliance on any information provided in this article is strictly at the reader's own risk.
By accessing and using this article, the reader expressly agrees to the terms of this disclaimer.
► "Absent reform, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern."
When a Video No Longer Proves What Happened: Why AI-Generated Evidence Is the Most Dangerous Evidentiary Challenge of Our Time
For centuries, courts operated on a foundational premise: evidence reflects reality. A photograph captured what the camera saw. A video depicted what occurred. A voice recording preserved what was said. Generative artificial intelligence has shattered this premise. Today, a photograph may depict an event that never occurred. A video may portray a person committing a crime that exists only in the imagination of an algorithm. A voice recording may reproduce speech that was never uttered. This article undertakes a comparative doctrinal analysis of how Indian and American courts are grappling with this evidentiary rupture, examines the inadequacy of current authentication standards in both jurisdictions, surveys landmark cases and legislative developments, and proposes a four-pillar reform framework requiring a shift from container-based to algorithm-based evidentiary inquiry.
The challenge is not merely theoretical. In Washington v. Puloka (2024), one of the first formal judicial decisions on AI-processed criminal evidence, a Washington State court was asked to admit an AI-enhanced version of a smartphone video from a triple homicide scene. The court excluded it, finding the AI tool untested, the operator unqualified, and the risk of juror confusion unacceptably high. In India, courts have begun to grapple with AI in nascent but significant ways, while the newly enacted Bharatiya Sakshya Adhiniyam, 2023, which replaced the colonial-era Indian Evidence Act, modernised digital evidence rules yet left a conspicuous legislative gap on AI-generated outputs.
This article proceeds in five parts. Part II analyses the taxonomy of AI evidence. Part III examines the American legal framework. Part IV analyses the Indian framework. Part V proposes a four-pillar reform. Part VI concludes.
Part II: How AI Evidence Is Classified and Why the Distinction Between Acknowledged and Unacknowledged AI Evidence Is Legally Critical
AI-generated evidence divides into two categories that require fundamentally different legal responses. The first, acknowledged AI evidence, encompasses material about which there is no dispute that an AI system created or processed it: facial recognition outputs, forensic AI analyses, algorithmic crime predictions, and AI-enhanced surveillance footage. The inquiry here is about reliability: was this AI system tested, validated, and methodologically sound?
The second, unacknowledged AI evidence or deepfakes, involves material where one party claims authenticity and the opposing party alleges AI fabrication. The inquiry here is about authenticity: is this evidence real at all? These two inquiries are epistemologically distinct. Conflating them, as courts in both jurisdictions have tended to do, produces analytical confusion and wrong outcomes. A deepfake is not merely unreliable evidence; it is fabricated evidence, and the legal framework must treat the distinction seriously.
The table below illustrates the fundamental difference between the two categories and the legal question each raises.
Category
Definition
Core Legal Question
Examples
Acknowledged AI Evidence
Material where both parties accept that AI created or processed it
Is this AI system reliable, validated, and methodologically sound?
Facial recognition output, forensic AI analysis, algorithmic crime prediction, AI-enhanced surveillance
Unacknowledged AI Evidence (Deepfakes)
Material where one party claims authenticity and the other alleges AI fabrication
Is this evidence genuine at all, or is it a fabrication designed to deceive?
AI-generated audio of a person confessing, synthetic video of a person at a crime scene, fabricated photographic evidence
Undergirding both categories is the structural opacity of modern AI systems. Unlike fingerprinting or DNA analysis, which operate on principles that can be explained, tested, and cross-examined, deep learning models produce outputs through millions of weighted parameters that even their creators cannot fully explain. This opacity creates a fundamental tension with the evidentiary requirement of verifiability. A court system that relies on lay jurors to determine whether a video is authentic or AI-fabricated is a court system that has outsourced its fact-finding function to a technology it cannot evaluate. The right to confront and challenge evidence is rendered hollow when the evidence is generated by a process that resists interrogation.
Part III: The American Legal Framework and Why Existing Rules Are Insufficient for AI Evidence
Existing Standards and Their Inadequacy
American courts have applied existing frameworks, Rule 901 on authentication, Rule 702 on expert testimony, and the gatekeeping standard of Daubert v. Merrell Dow Pharmaceuticals (1993), to AI evidence by analogy. Washington v. Puloka illustrated the limits of this approach. The defendant sought to introduce AI-enhanced footage as exculpatory evidence. The court excluded it because the tool had no documented testing history, no peer-reviewed validation, and was operated by a filmmaker with no forensic AI training. While Daubert's flexible standard is capable of addressing the reliability of acknowledged AI outputs, it provides no clear framework for the far harder challenge of unacknowledged AI manipulation, deepfakes that are designed to deceive even trained observers.
Proposed Federal Rule 707 and State Developments
Federal rulemakers have responded with unusual speed. After an 8-1 vote by the Advisory Committee in May 2025, proposed Rule 707 was released for public comment through February 2026. The Rule would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702, requiring the proponent to show: sufficient facts or data as input; reliable AI principles and methods; and reliable application of those methods to the case facts. The Rule's purpose is to prevent parties from circumventing Rule 702 by offering machine outputs directly rather than through a cross-examinable human expert.
The table below compares the existing American framework with the proposed Rule 707.
Framework
Provision
Application to AI
Key Limitation
Authentication
Federal Rule of Evidence 901
Requires showing evidence is what it purports to be
Designed for documents that passively capture reality; cannot address AI fabrication
Expert Testimony
Federal Rule of Evidence 702
Daubert standard requires reliability, testability, peer review
Not designed for machine outputs presented without a human expert sponsor
Proposed AI Evidence Rule
Proposed Federal Rule 707
Applies Rule 702 standard to acknowledged AI evidence; requires sufficient input data, reliable methods, reliable application
Applies only to acknowledged AI evidence; does nothing for deepfakes presented as genuine
State Frameworks
California SB 11, Louisiana Act 2025
Condition admissibility on independent expert verification and documented reliability testing
Fragmented; no uniform national standard
Rule 707 is a significant step but an incomplete solution. Critically, it applies only to evidence the proponent acknowledges was created by AI. It therefore does nothing for the deepfake problem, the case where fabricated evidence is presented as genuine. Louisiana and California have also enacted or proposed state-level frameworks conditioning admissibility on independent expert verification and documented reliability testing, demonstrating that legislative action is both feasible and urgent.
Part IV: The Indian Legal Framework and the Conspicuous Gap Left by the BSA 2023
From Section 65B to BSA 2023: Modernisation Without Resolution
Indian evidence law's engagement with digital evidence was shaped by two Supreme Court decisions. In Anvar P.V. v. P.K. Basheer (2014) 10 SCC 473, the Court held that electronic evidence without a Section 65B certificate is inadmissible as secondary evidence, a position reaffirmed in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020) 7 SCC 1. The certificate requirement became the gatekeeping mechanism for all digital evidence in India. The Bharatiya Sakshya Adhiniyam, 2023 preserved this architecture in Sections 61 to 63, modernising the language around electronic records while retaining the same certificate-based authentication framework.
The problem is that this framework rests on assumptions that AI evidence does not satisfy.
The table below identifies the assumptions of the BSA framework and the ways AI evidence fails each of them.
BSA Framework Assumption
Satisfied by Traditional Digital Evidence?
Satisfied by AI-Generated Evidence?
Traceable human author
Yes
No; AI outputs have no singular human author
Identifiable originating device
Yes
No; cloud-based AI models produce outputs across distributed systems
Clear chain of custody
Yes
No; training data, model weights, and inference process resist conventional custody documentation
Record of a real event
Yes
No; a deepfake is a fabrication masquerading as a record
Certificate by a responsible official
Yes
Unclear; no provision addresses who certifies AI-generated output
The BSA contains no definition of "AI-generated evidence," no authentication framework for machine-generated outputs, and no provision addressing deepfakes or synthetic media. The question of whether a deepfake even constitutes an "electronic record" within the meaning of the BSA is itself unresolved: a deepfake is not a record of an event but a fabrication masquerading as one.
Section 39, Constitutional Angles, and Current Judicial Engagement
Section 39 of the BSA, governing expert opinion, offers a potential doctrinal foothold. Its reference to "any other field" of expertise may be interpreted to encompass validated AI forensic systems. However, this has never been judicially confirmed, and Section 39 provides no guidance on algorithmic bias, training data opacity, or the structural impossibility of cross-examining a machine. In Jaswinder Singh v. State of Punjab (CRR-2907-2023, Punjab and Haryana High Court, 2023), the court referenced ChatGPT as an analytical research tool in a bail application, taking care to clarify that it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
The constitutional dimensions are significant and underexplored. AI-powered facial recognition and voice identification create direct tension with Article 20(3), the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017) 10 SCC 1. When a court convicts on the basis of facial recognition with a documented error rate, the accused has not received a fair trial; they have received a statistical sentence. These constitutional questions remain unanswered by Indian jurisprudence.
► Key Principle: The BSA 2023 contains no definition of AI-generated evidence, no authentication framework for machine outputs, and no provision on deepfakes. This is not a minor oversight. It is a structural legislative gap that leaves Indian courts without a principled basis for evaluating the most consequential category of evidence they will face in the coming decade.
Part V: Comparative Assessment and the Four-Pillar Reform Framework
A comparative analysis reveals that both the American and Indian frameworks suffer from the same conceptual deficiency: they approach AI evidence through a container-based lens, treating it as a species of electronic record subject to authentication standards designed for documents that passively capture reality. This framework is fundamentally ill-suited to evidence that actively constructs or fabricates reality. The authentication question must change from "Is this file authentic?" to "Is this algorithm reliable, transparent, and free from adversarial manipulation?" These are epistemologically distinct inquiries, and conflating them is not a minor technical error. It is a structural failure of judicial gatekeeping.
The table below sets out the four-pillar reform framework proposed in this article and its application to both jurisdictions.
Pillar
Reform Proposed
India Application
US Application
Pillar 1: Statutory Definition
Enact statutory provisions defining "AI-generated evidence" as a distinct evidentiary category with disclosure requirements covering AI system architecture, training data, error rates, and validation history
Introduce a dedicated provision within the BSA 2023; define AI-generated evidence and synthetic media expressly
Enact Rule 707 with mandatory disclosure schedules; extend beyond acknowledged AI to cover unacknowledged fabrication
Pillar 2: Algorithm-Centred Authentication
Shift from authenticating the storage device to interrogating the scientific validity of the algorithm through Daubert-style reliability hearings examining training data integrity, model validation, false positive rates, and susceptibility to adversarial inputs
Adapt BSA Section 39 expert framework to require AI system validation evidence before any AI output is admitted
Strengthen proposed Rule 707 to mandate pre-admission reliability hearings for all AI-generated outputs
Pillar 3: Burden-Shifting for Deepfakes
For unacknowledged AI evidence, the challenger must first present evidence sufficient for a reasonable factfinder to conclude the evidence may have been AI-fabricated; the burden then shifts to the proponent to establish authenticity by a preponderance
No existing provision; requires new BSA amendment or dedicated AI evidence Act
No existing Federal Rule addresses deepfakes presented as genuine; requires extension of Rule 707 or new Rule
Pillar 4: Institutional Safeguards
Develop bench guidance for judges, AI forensic certification protocols, and mandatory metadata disclosure for digital submissions; adopt EU AI Act Article 13 transparency obligations as a model
NCLT, High Courts, and trial courts require dedicated guidance on AI evidence evaluation; forensic AI certification for expert witnesses
Federal Judicial Center guidance needed; state bar AI forensic certification programmes; mandatory metadata disclosure rules
Part VI: Conclusion
The admissibility of AI-generated evidence presents evidentiary law with its most demanding test since the introduction of DNA evidence. Unlike DNA, however, AI-generated evidence may be designed to deceive. Both the United States and India possess the doctrinal raw material for reform: Rule 707, state-level statutes, the BSA's modernised digital evidence framework, and a rich constitutional jurisprudence on privacy and fair trial rights. What both lack is a unified, algorithm-centred evidentiary framework that treats AI-generated outputs as a distinct category requiring categorical rules, not case-by-case improvisation.
Absent these reforms, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern.
Frequently Asked Questions (FAQs) on AI-Generated Evidence in Indian and American Courts
1. What is AI-generated evidence and why does it pose a unique legal challenge? AI-generated evidence refers to text, audio, video, images, or analytical outputs created or significantly altered by artificial intelligence systems. It poses a unique legal challenge because, unlike traditional evidence that passively records reality, AI can actively fabricate reality with a realism that defeats conventional authentication methods and deceives even trained observers.
2. What is the difference between acknowledged and unacknowledged AI evidence? Acknowledged AI evidence is material where both parties accept that an AI system created or processed it; the legal inquiry is about reliability. Unacknowledged AI evidence or deepfakes is material where one party claims authenticity and the other alleges fabrication; the legal inquiry is about whether the evidence is genuine at all. These are epistemologically distinct questions requiring different legal frameworks.
3. What did the Washington v. Puloka case decide about AI evidence? In Washington v. Puloka (2024), a Washington State court excluded AI-enhanced smartphone footage from a triple homicide case because the AI tool had no documented testing history, no peer-reviewed validation, and was operated by a person with no forensic AI training. The case illustrates both the limits of applying existing authentication standards to AI evidence and the importance of judicial gatekeeping.
4. What is Proposed Federal Rule 707 and what does it cover? Proposed Federal Rule of Evidence 707, released for public comment after an 8-1 Advisory Committee vote in May 2025, would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702. It requires the proponent to demonstrate sufficient input data, reliable AI methods, and reliable application of those methods to the case facts. Critically, it applies only to acknowledged AI evidence and does not address deepfakes presented as genuine.
5. How does the Bharatiya Sakshya Adhiniyam 2023 address AI-generated evidence? The BSA 2023 modernised India's digital evidence framework through Sections 61 to 63 and retained the certificate-based authentication framework from the old Indian Evidence Act. However, it contains no definition of AI-generated evidence, no authentication framework for machine outputs, no provision on deepfakes, and no guidance on the structural impossibility of cross-examining an algorithm. This is a significant legislative gap.
6. What constitutional concerns does AI evidence raise in India? AI-powered facial recognition and voice identification create direct tension with Article 20(3) of the Constitution, the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017). A conviction based on facial recognition with a documented error rate raises serious fair trial questions that Indian jurisprudence has not yet resolved.
7. What is the black-box problem in AI evidence? The black-box problem refers to the structural opacity of modern deep learning AI systems, which produce outputs through millions of weighted parameters that even their creators cannot fully explain. Unlike fingerprinting or DNA analysis, AI reasoning cannot be fully explained, tested, or cross-examined, creating a fundamental conflict with the evidentiary requirement of verifiability and the accused's right to confront and challenge evidence.
8. What are the four pillars of the reform framework proposed in this article? The four pillars are: first, statutory definition of AI-generated evidence as a distinct evidentiary category with mandatory disclosure requirements; second, algorithm-centred authentication through Daubert-style reliability hearings examining training data, model validation, and false positive rates; third, a burden-shifting mechanism for deepfakes where the challenger raises a reasonable fabrication concern and the burden shifts to the proponent to prove authenticity; and fourth, institutional safeguards including judicial guidance, AI forensic certification protocols, and mandatory metadata disclosure.
Key Takeaways
Generative AI has shattered the foundational evidentiary premise that evidence reflects reality; courts in India and the United States now face AI-generated text, audio, video, and analytical outputs for which existing evidence rules were not designed.
AI-generated evidence divides into two legally distinct categories: acknowledged AI evidence, where the inquiry is reliability, and unacknowledged AI evidence or deepfakes, where the inquiry is authenticity. Conflating these categories produces analytical confusion and wrong judicial outcomes.
The black-box opacity of modern deep learning models creates a fundamental conflict with the evidentiary requirement of verifiability, rendering the right to confront and challenge evidence hollow when the evidence was generated by a process that resists interrogation.
In Washington v. Puloka (2024), a US court excluded AI-enhanced footage because the tool was untested, the operator unqualified, and the risk of juror confusion unacceptably high, illustrating both the limits of existing authentication standards and the possibility of principled judicial exclusion.
Proposed Federal Rule of Evidence 707, released for public comment in 2025 after an 8-1 Advisory Committee vote, would apply the Rule 702 expert testimony standard to AI-generated evidence but applies only to acknowledged AI evidence and leaves the deepfake problem entirely unaddressed.
The Bharatiya Sakshya Adhiniyam, 2023 modernised India's digital evidence framework but left a conspicuous legislative gap: no definition of AI-generated evidence, no AI-specific authentication framework, and no provision addressing deepfakes or synthetic media.
The BSA's certificate-based authentication framework rests on assumptions (traceable human author, identifiable device, clear chain of custody, record of a real event) that AI-generated evidence systematically fails to satisfy.
In Jaswinder Singh v. State of Punjab (2023), the Punjab and Haryana High Court referenced ChatGPT as a research tool in a bail application while clarifying it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
AI-powered facial recognition and voice identification create direct constitutional tension with Article 20(3) and Article 21 of the Indian Constitution, raising fair trial questions that Indian jurisprudence has not yet resolved.
Both jurisdictions require the same four-pillar reform: statutory definition of AI evidence, algorithm-centred authentication replacing container-based authentication, burden-shifting for deepfakes, and institutional safeguards including judicial guidance, AI forensic certification, and mandatory metadata disclosure.
References
Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473 (India): The foundational Supreme Court decision requiring a Section 65B certificate for the admissibility of electronic evidence as secondary evidence in Indian courts.
Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1 (India): Reaffirmed and clarified the Section 65B certificate requirement, cementing the certificate-based authentication architecture that the BSA 2023 has now inherited.
Jaswinder Singh v. State of Punjab, CRR-2907-2023 (Punjab and Haryana High Court, 2023): The first significant Indian judicial reference to a generative AI tool, with the court referencing ChatGPT as a research aid while explicitly clarifying it had no influence on the bail order.
K.S. Puttaswamy v. Union of India, (2017) 10 SCC 1 (India): The nine-judge bench decision recognising the right to privacy as a fundamental right under Article 21, directly relevant to the constitutional implications of AI-powered facial recognition and voice identification as evidence.
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993): The foundational US Supreme Court decision establishing the Daubert reliability standard for expert testimony and scientific evidence under Federal Rule of Evidence 702, now being extended by analogy to AI-generated evidence.
Frye v. United States, 293 F. 1013 (D.C. Cir. 1923): The predecessor general acceptance standard for scientific evidence, still applicable in several US states and relevant to the historical development of evidentiary standards for novel scientific methods.
Washington v. Puloka, No. 22-1-03036-5 SEA (Washington Superior Court, King County, March 2024): One of the first formal judicial decisions on AI-processed criminal evidence; excluded AI-enhanced footage for lack of documented testing, qualified operator, and unacceptable juror confusion risk.
Bharatiya Sakshya Adhiniyam, 2023, No. 47 of 2023 (India), Sections 39, 61 to 63: The primary Indian evidence statute governing expert opinion and electronic records, whose certificate-based authentication framework does not address AI-generated evidence or deepfakes.
Indian Evidence Act, No. 1 of 1872 (India), Section 65B (repealed 2023): The predecessor provision on electronic evidence certificates, whose architecture was carried forward into the BSA 2023.
Information Technology Act, No. 21 of 2000 (India), Sections 79A and 85B: Relevant supplementary provisions on electronic evidence and digital examination in the Indian statutory framework.
Constitution of India, Articles 20(3) and 21: The constitutional provisions on the right against self-incrimination and the right to life and personal liberty, both directly engaged by AI-powered identification evidence in criminal proceedings.
Federal Rules of Evidence, Rules 403, 702, and 901: The existing US evidentiary framework being applied by analogy to AI-generated evidence, and the foundation upon which Proposed Rule 707 builds.
Proposed Federal Rule of Evidence 707 (comment period closed February 16, 2026): The proposed US federal rule that would apply the Rule 702 expert testimony standard to AI-generated evidence; applies only to acknowledged AI evidence.
California Senate Bill 11, 2025-26 Session (California, 2025): State-level legislative framework conditioning AI evidence admissibility on independent expert verification and documented reliability testing.
Louisiana Act No. (effective August 1, 2025): State-level AI evidence framework requiring reliability testing documentation as a condition of admissibility.
Disclaimer
This article is published by CLEAR LAW (clearlaw.online) strictly for educational and informational purposes only. It does not constitute legal advice, legal opinion, or any form of professional counsel, and must not be relied upon as a substitute for consultation with a qualified legal practitioner. Nothing contained herein shall be construed as creating a lawyer-client relationship between the reader and the author, publisher, or CLEAR LAW (clearlaw.online).
All views, interpretations, and conclusions expressed in this article are solely those of the author and represent independent academic analysis. CLEAR LAW (clearlaw.online) does not endorse, verify, or guarantee the accuracy, completeness, or reliability of the content, and expressly disclaims any responsibility for the same.
While reasonable efforts are made to ensure that the information presented is accurate and up to date, no warranties or representations, express or implied, are made regarding its correctness, adequacy, or applicability to any specific factual or legal situation. Laws, regulations, and judicial interpretations are subject to change, and the content may not reflect the most current legal developments.
To the fullest extent permitted by applicable law, CLEAR LAW (clearlaw.online), the author, editors, and publisher disclaim all liability for any direct, indirect, incidental, consequential, or special damages arising out of or in connection with the use of, or reliance upon, this article.
Readers are strongly advised to seek independent legal advice from a qualified professional before making any decisions or taking any action based on the contents of this article. Reliance on any information provided in this article is strictly at the reader's own risk.
By accessing and using this article, the reader expressly agrees to the terms of this disclaimer.
► "Absent reform, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern."
When a Video No Longer Proves What Happened: Why AI-Generated Evidence Is the Most Dangerous Evidentiary Challenge of Our Time
For centuries, courts operated on a foundational premise: evidence reflects reality. A photograph captured what the camera saw. A video depicted what occurred. A voice recording preserved what was said. Generative artificial intelligence has shattered this premise. Today, a photograph may depict an event that never occurred. A video may portray a person committing a crime that exists only in the imagination of an algorithm. A voice recording may reproduce speech that was never uttered. This article undertakes a comparative doctrinal analysis of how Indian and American courts are grappling with this evidentiary rupture, examines the inadequacy of current authentication standards in both jurisdictions, surveys landmark cases and legislative developments, and proposes a four-pillar reform framework requiring a shift from container-based to algorithm-based evidentiary inquiry.
The challenge is not merely theoretical. In Washington v. Puloka (2024), one of the first formal judicial decisions on AI-processed criminal evidence, a Washington State court was asked to admit an AI-enhanced version of a smartphone video from a triple homicide scene. The court excluded it, finding the AI tool untested, the operator unqualified, and the risk of juror confusion unacceptably high. In India, courts have begun to grapple with AI in nascent but significant ways, while the newly enacted Bharatiya Sakshya Adhiniyam, 2023, which replaced the colonial-era Indian Evidence Act, modernised digital evidence rules yet left a conspicuous legislative gap on AI-generated outputs.
This article proceeds in five parts. Part II analyses the taxonomy of AI evidence. Part III examines the American legal framework. Part IV analyses the Indian framework. Part V proposes a four-pillar reform. Part VI concludes.
Part II: How AI Evidence Is Classified and Why the Distinction Between Acknowledged and Unacknowledged AI Evidence Is Legally Critical
AI-generated evidence divides into two categories that require fundamentally different legal responses. The first, acknowledged AI evidence, encompasses material about which there is no dispute that an AI system created or processed it: facial recognition outputs, forensic AI analyses, algorithmic crime predictions, and AI-enhanced surveillance footage. The inquiry here is about reliability: was this AI system tested, validated, and methodologically sound?
The second, unacknowledged AI evidence or deepfakes, involves material where one party claims authenticity and the opposing party alleges AI fabrication. The inquiry here is about authenticity: is this evidence real at all? These two inquiries are epistemologically distinct. Conflating them, as courts in both jurisdictions have tended to do, produces analytical confusion and wrong outcomes. A deepfake is not merely unreliable evidence; it is fabricated evidence, and the legal framework must treat the distinction seriously.
The table below illustrates the fundamental difference between the two categories and the legal question each raises.
Category
Definition
Core Legal Question
Examples
Acknowledged AI Evidence
Material where both parties accept that AI created or processed it
Is this AI system reliable, validated, and methodologically sound?
Facial recognition output, forensic AI analysis, algorithmic crime prediction, AI-enhanced surveillance
Unacknowledged AI Evidence (Deepfakes)
Material where one party claims authenticity and the other alleges AI fabrication
Is this evidence genuine at all, or is it a fabrication designed to deceive?
AI-generated audio of a person confessing, synthetic video of a person at a crime scene, fabricated photographic evidence
Undergirding both categories is the structural opacity of modern AI systems. Unlike fingerprinting or DNA analysis, which operate on principles that can be explained, tested, and cross-examined, deep learning models produce outputs through millions of weighted parameters that even their creators cannot fully explain. This opacity creates a fundamental tension with the evidentiary requirement of verifiability. A court system that relies on lay jurors to determine whether a video is authentic or AI-fabricated is a court system that has outsourced its fact-finding function to a technology it cannot evaluate. The right to confront and challenge evidence is rendered hollow when the evidence is generated by a process that resists interrogation.
Part III: The American Legal Framework and Why Existing Rules Are Insufficient for AI Evidence
Existing Standards and Their Inadequacy
American courts have applied existing frameworks, Rule 901 on authentication, Rule 702 on expert testimony, and the gatekeeping standard of Daubert v. Merrell Dow Pharmaceuticals (1993), to AI evidence by analogy. Washington v. Puloka illustrated the limits of this approach. The defendant sought to introduce AI-enhanced footage as exculpatory evidence. The court excluded it because the tool had no documented testing history, no peer-reviewed validation, and was operated by a filmmaker with no forensic AI training. While Daubert's flexible standard is capable of addressing the reliability of acknowledged AI outputs, it provides no clear framework for the far harder challenge of unacknowledged AI manipulation, deepfakes that are designed to deceive even trained observers.
Proposed Federal Rule 707 and State Developments
Federal rulemakers have responded with unusual speed. After an 8-1 vote by the Advisory Committee in May 2025, proposed Rule 707 was released for public comment through February 2026. The Rule would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702, requiring the proponent to show: sufficient facts or data as input; reliable AI principles and methods; and reliable application of those methods to the case facts. The Rule's purpose is to prevent parties from circumventing Rule 702 by offering machine outputs directly rather than through a cross-examinable human expert.
The table below compares the existing American framework with the proposed Rule 707.
Framework
Provision
Application to AI
Key Limitation
Authentication
Federal Rule of Evidence 901
Requires showing evidence is what it purports to be
Designed for documents that passively capture reality; cannot address AI fabrication
Expert Testimony
Federal Rule of Evidence 702
Daubert standard requires reliability, testability, peer review
Not designed for machine outputs presented without a human expert sponsor
Proposed AI Evidence Rule
Proposed Federal Rule 707
Applies Rule 702 standard to acknowledged AI evidence; requires sufficient input data, reliable methods, reliable application
Applies only to acknowledged AI evidence; does nothing for deepfakes presented as genuine
State Frameworks
California SB 11, Louisiana Act 2025
Condition admissibility on independent expert verification and documented reliability testing
Fragmented; no uniform national standard
Rule 707 is a significant step but an incomplete solution. Critically, it applies only to evidence the proponent acknowledges was created by AI. It therefore does nothing for the deepfake problem, the case where fabricated evidence is presented as genuine. Louisiana and California have also enacted or proposed state-level frameworks conditioning admissibility on independent expert verification and documented reliability testing, demonstrating that legislative action is both feasible and urgent.
Part IV: The Indian Legal Framework and the Conspicuous Gap Left by the BSA 2023
From Section 65B to BSA 2023: Modernisation Without Resolution
Indian evidence law's engagement with digital evidence was shaped by two Supreme Court decisions. In Anvar P.V. v. P.K. Basheer (2014) 10 SCC 473, the Court held that electronic evidence without a Section 65B certificate is inadmissible as secondary evidence, a position reaffirmed in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020) 7 SCC 1. The certificate requirement became the gatekeeping mechanism for all digital evidence in India. The Bharatiya Sakshya Adhiniyam, 2023 preserved this architecture in Sections 61 to 63, modernising the language around electronic records while retaining the same certificate-based authentication framework.
The problem is that this framework rests on assumptions that AI evidence does not satisfy.
The table below identifies the assumptions of the BSA framework and the ways AI evidence fails each of them.
BSA Framework Assumption
Satisfied by Traditional Digital Evidence?
Satisfied by AI-Generated Evidence?
Traceable human author
Yes
No; AI outputs have no singular human author
Identifiable originating device
Yes
No; cloud-based AI models produce outputs across distributed systems
Clear chain of custody
Yes
No; training data, model weights, and inference process resist conventional custody documentation
Record of a real event
Yes
No; a deepfake is a fabrication masquerading as a record
Certificate by a responsible official
Yes
Unclear; no provision addresses who certifies AI-generated output
The BSA contains no definition of "AI-generated evidence," no authentication framework for machine-generated outputs, and no provision addressing deepfakes or synthetic media. The question of whether a deepfake even constitutes an "electronic record" within the meaning of the BSA is itself unresolved: a deepfake is not a record of an event but a fabrication masquerading as one.
Section 39, Constitutional Angles, and Current Judicial Engagement
Section 39 of the BSA, governing expert opinion, offers a potential doctrinal foothold. Its reference to "any other field" of expertise may be interpreted to encompass validated AI forensic systems. However, this has never been judicially confirmed, and Section 39 provides no guidance on algorithmic bias, training data opacity, or the structural impossibility of cross-examining a machine. In Jaswinder Singh v. State of Punjab (CRR-2907-2023, Punjab and Haryana High Court, 2023), the court referenced ChatGPT as an analytical research tool in a bail application, taking care to clarify that it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
The constitutional dimensions are significant and underexplored. AI-powered facial recognition and voice identification create direct tension with Article 20(3), the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017) 10 SCC 1. When a court convicts on the basis of facial recognition with a documented error rate, the accused has not received a fair trial; they have received a statistical sentence. These constitutional questions remain unanswered by Indian jurisprudence.
► Key Principle: The BSA 2023 contains no definition of AI-generated evidence, no authentication framework for machine outputs, and no provision on deepfakes. This is not a minor oversight. It is a structural legislative gap that leaves Indian courts without a principled basis for evaluating the most consequential category of evidence they will face in the coming decade.
Part V: Comparative Assessment and the Four-Pillar Reform Framework
A comparative analysis reveals that both the American and Indian frameworks suffer from the same conceptual deficiency: they approach AI evidence through a container-based lens, treating it as a species of electronic record subject to authentication standards designed for documents that passively capture reality. This framework is fundamentally ill-suited to evidence that actively constructs or fabricates reality. The authentication question must change from "Is this file authentic?" to "Is this algorithm reliable, transparent, and free from adversarial manipulation?" These are epistemologically distinct inquiries, and conflating them is not a minor technical error. It is a structural failure of judicial gatekeeping.
The table below sets out the four-pillar reform framework proposed in this article and its application to both jurisdictions.
Pillar
Reform Proposed
India Application
US Application
Pillar 1: Statutory Definition
Enact statutory provisions defining "AI-generated evidence" as a distinct evidentiary category with disclosure requirements covering AI system architecture, training data, error rates, and validation history
Introduce a dedicated provision within the BSA 2023; define AI-generated evidence and synthetic media expressly
Enact Rule 707 with mandatory disclosure schedules; extend beyond acknowledged AI to cover unacknowledged fabrication
Pillar 2: Algorithm-Centred Authentication
Shift from authenticating the storage device to interrogating the scientific validity of the algorithm through Daubert-style reliability hearings examining training data integrity, model validation, false positive rates, and susceptibility to adversarial inputs
Adapt BSA Section 39 expert framework to require AI system validation evidence before any AI output is admitted
Strengthen proposed Rule 707 to mandate pre-admission reliability hearings for all AI-generated outputs
Pillar 3: Burden-Shifting for Deepfakes
For unacknowledged AI evidence, the challenger must first present evidence sufficient for a reasonable factfinder to conclude the evidence may have been AI-fabricated; the burden then shifts to the proponent to establish authenticity by a preponderance
No existing provision; requires new BSA amendment or dedicated AI evidence Act
No existing Federal Rule addresses deepfakes presented as genuine; requires extension of Rule 707 or new Rule
Pillar 4: Institutional Safeguards
Develop bench guidance for judges, AI forensic certification protocols, and mandatory metadata disclosure for digital submissions; adopt EU AI Act Article 13 transparency obligations as a model
NCLT, High Courts, and trial courts require dedicated guidance on AI evidence evaluation; forensic AI certification for expert witnesses
Federal Judicial Center guidance needed; state bar AI forensic certification programmes; mandatory metadata disclosure rules
Part VI: Conclusion
The admissibility of AI-generated evidence presents evidentiary law with its most demanding test since the introduction of DNA evidence. Unlike DNA, however, AI-generated evidence may be designed to deceive. Both the United States and India possess the doctrinal raw material for reform: Rule 707, state-level statutes, the BSA's modernised digital evidence framework, and a rich constitutional jurisprudence on privacy and fair trial rights. What both lack is a unified, algorithm-centred evidentiary framework that treats AI-generated outputs as a distinct category requiring categorical rules, not case-by-case improvisation.
Absent these reforms, courts will face the silicon witness unprepared and the integrity of the adversarial fact-finding process will bear the cost. The law must now move as quickly as the technology it is failing to govern.
Frequently Asked Questions (FAQs) on AI-Generated Evidence in Indian and American Courts
1. What is AI-generated evidence and why does it pose a unique legal challenge? AI-generated evidence refers to text, audio, video, images, or analytical outputs created or significantly altered by artificial intelligence systems. It poses a unique legal challenge because, unlike traditional evidence that passively records reality, AI can actively fabricate reality with a realism that defeats conventional authentication methods and deceives even trained observers.
2. What is the difference between acknowledged and unacknowledged AI evidence? Acknowledged AI evidence is material where both parties accept that an AI system created or processed it; the legal inquiry is about reliability. Unacknowledged AI evidence or deepfakes is material where one party claims authenticity and the other alleges fabrication; the legal inquiry is about whether the evidence is genuine at all. These are epistemologically distinct questions requiring different legal frameworks.
3. What did the Washington v. Puloka case decide about AI evidence? In Washington v. Puloka (2024), a Washington State court excluded AI-enhanced smartphone footage from a triple homicide case because the AI tool had no documented testing history, no peer-reviewed validation, and was operated by a person with no forensic AI training. The case illustrates both the limits of applying existing authentication standards to AI evidence and the importance of judicial gatekeeping.
4. What is Proposed Federal Rule 707 and what does it cover? Proposed Federal Rule of Evidence 707, released for public comment after an 8-1 Advisory Committee vote in May 2025, would subject AI-generated evidence to the same admissibility standard as expert testimony under Rule 702. It requires the proponent to demonstrate sufficient input data, reliable AI methods, and reliable application of those methods to the case facts. Critically, it applies only to acknowledged AI evidence and does not address deepfakes presented as genuine.
5. How does the Bharatiya Sakshya Adhiniyam 2023 address AI-generated evidence? The BSA 2023 modernised India's digital evidence framework through Sections 61 to 63 and retained the certificate-based authentication framework from the old Indian Evidence Act. However, it contains no definition of AI-generated evidence, no authentication framework for machine outputs, no provision on deepfakes, and no guidance on the structural impossibility of cross-examining an algorithm. This is a significant legislative gap.
6. What constitutional concerns does AI evidence raise in India? AI-powered facial recognition and voice identification create direct tension with Article 20(3) of the Constitution, the right against self-incrimination, and Article 21's protection of personal liberty and privacy as elaborated in K.S. Puttaswamy v. Union of India (2017). A conviction based on facial recognition with a documented error rate raises serious fair trial questions that Indian jurisprudence has not yet resolved.
7. What is the black-box problem in AI evidence? The black-box problem refers to the structural opacity of modern deep learning AI systems, which produce outputs through millions of weighted parameters that even their creators cannot fully explain. Unlike fingerprinting or DNA analysis, AI reasoning cannot be fully explained, tested, or cross-examined, creating a fundamental conflict with the evidentiary requirement of verifiability and the accused's right to confront and challenge evidence.
8. What are the four pillars of the reform framework proposed in this article? The four pillars are: first, statutory definition of AI-generated evidence as a distinct evidentiary category with mandatory disclosure requirements; second, algorithm-centred authentication through Daubert-style reliability hearings examining training data, model validation, and false positive rates; third, a burden-shifting mechanism for deepfakes where the challenger raises a reasonable fabrication concern and the burden shifts to the proponent to prove authenticity; and fourth, institutional safeguards including judicial guidance, AI forensic certification protocols, and mandatory metadata disclosure.
Key Takeaways
Generative AI has shattered the foundational evidentiary premise that evidence reflects reality; courts in India and the United States now face AI-generated text, audio, video, and analytical outputs for which existing evidence rules were not designed.
AI-generated evidence divides into two legally distinct categories: acknowledged AI evidence, where the inquiry is reliability, and unacknowledged AI evidence or deepfakes, where the inquiry is authenticity. Conflating these categories produces analytical confusion and wrong judicial outcomes.
The black-box opacity of modern deep learning models creates a fundamental conflict with the evidentiary requirement of verifiability, rendering the right to confront and challenge evidence hollow when the evidence was generated by a process that resists interrogation.
In Washington v. Puloka (2024), a US court excluded AI-enhanced footage because the tool was untested, the operator unqualified, and the risk of juror confusion unacceptably high, illustrating both the limits of existing authentication standards and the possibility of principled judicial exclusion.
Proposed Federal Rule of Evidence 707, released for public comment in 2025 after an 8-1 Advisory Committee vote, would apply the Rule 702 expert testimony standard to AI-generated evidence but applies only to acknowledged AI evidence and leaves the deepfake problem entirely unaddressed.
The Bharatiya Sakshya Adhiniyam, 2023 modernised India's digital evidence framework but left a conspicuous legislative gap: no definition of AI-generated evidence, no AI-specific authentication framework, and no provision addressing deepfakes or synthetic media.
The BSA's certificate-based authentication framework rests on assumptions (traceable human author, identifiable device, clear chain of custody, record of a real event) that AI-generated evidence systematically fails to satisfy.
In Jaswinder Singh v. State of Punjab (2023), the Punjab and Haryana High Court referenced ChatGPT as a research tool in a bail application while clarifying it had no influence on the order, representing the current outer boundary of Indian judicial engagement with AI.
AI-powered facial recognition and voice identification create direct constitutional tension with Article 20(3) and Article 21 of the Indian Constitution, raising fair trial questions that Indian jurisprudence has not yet resolved.
Both jurisdictions require the same four-pillar reform: statutory definition of AI evidence, algorithm-centred authentication replacing container-based authentication, burden-shifting for deepfakes, and institutional safeguards including judicial guidance, AI forensic certification, and mandatory metadata disclosure.
References
Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473 (India): The foundational Supreme Court decision requiring a Section 65B certificate for the admissibility of electronic evidence as secondary evidence in Indian courts.
Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1 (India): Reaffirmed and clarified the Section 65B certificate requirement, cementing the certificate-based authentication architecture that the BSA 2023 has now inherited.
Jaswinder Singh v. State of Punjab, CRR-2907-2023 (Punjab and Haryana High Court, 2023): The first significant Indian judicial reference to a generative AI tool, with the court referencing ChatGPT as a research aid while explicitly clarifying it had no influence on the bail order.
K.S. Puttaswamy v. Union of India, (2017) 10 SCC 1 (India): The nine-judge bench decision recognising the right to privacy as a fundamental right under Article 21, directly relevant to the constitutional implications of AI-powered facial recognition and voice identification as evidence.
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993): The foundational US Supreme Court decision establishing the Daubert reliability standard for expert testimony and scientific evidence under Federal Rule of Evidence 702, now being extended by analogy to AI-generated evidence.
Frye v. United States, 293 F. 1013 (D.C. Cir. 1923): The predecessor general acceptance standard for scientific evidence, still applicable in several US states and relevant to the historical development of evidentiary standards for novel scientific methods.
Washington v. Puloka, No. 22-1-03036-5 SEA (Washington Superior Court, King County, March 2024): One of the first formal judicial decisions on AI-processed criminal evidence; excluded AI-enhanced footage for lack of documented testing, qualified operator, and unacceptable juror confusion risk.
Bharatiya Sakshya Adhiniyam, 2023, No. 47 of 2023 (India), Sections 39, 61 to 63: The primary Indian evidence statute governing expert opinion and electronic records, whose certificate-based authentication framework does not address AI-generated evidence or deepfakes.
Indian Evidence Act, No. 1 of 1872 (India), Section 65B (repealed 2023): The predecessor provision on electronic evidence certificates, whose architecture was carried forward into the BSA 2023.
Information Technology Act, No. 21 of 2000 (India), Sections 79A and 85B: Relevant supplementary provisions on electronic evidence and digital examination in the Indian statutory framework.
Constitution of India, Articles 20(3) and 21: The constitutional provisions on the right against self-incrimination and the right to life and personal liberty, both directly engaged by AI-powered identification evidence in criminal proceedings.
Federal Rules of Evidence, Rules 403, 702, and 901: The existing US evidentiary framework being applied by analogy to AI-generated evidence, and the foundation upon which Proposed Rule 707 builds.
Proposed Federal Rule of Evidence 707 (comment period closed February 16, 2026): The proposed US federal rule that would apply the Rule 702 expert testimony standard to AI-generated evidence; applies only to acknowledged AI evidence.
California Senate Bill 11, 2025-26 Session (California, 2025): State-level legislative framework conditioning AI evidence admissibility on independent expert verification and documented reliability testing.
Louisiana Act No. (effective August 1, 2025): State-level AI evidence framework requiring reliability testing documentation as a condition of admissibility.
Disclaimer
This article is published by CLEAR LAW (clearlaw.online) strictly for educational and informational purposes only. It does not constitute legal advice, legal opinion, or any form of professional counsel, and must not be relied upon as a substitute for consultation with a qualified legal practitioner. Nothing contained herein shall be construed as creating a lawyer-client relationship between the reader and the author, publisher, or CLEAR LAW (clearlaw.online).
All views, interpretations, and conclusions expressed in this article are solely those of the author and represent independent academic analysis. CLEAR LAW (clearlaw.online) does not endorse, verify, or guarantee the accuracy, completeness, or reliability of the content, and expressly disclaims any responsibility for the same.
While reasonable efforts are made to ensure that the information presented is accurate and up to date, no warranties or representations, express or implied, are made regarding its correctness, adequacy, or applicability to any specific factual or legal situation. Laws, regulations, and judicial interpretations are subject to change, and the content may not reflect the most current legal developments.
To the fullest extent permitted by applicable law, CLEAR LAW (clearlaw.online), the author, editors, and publisher disclaim all liability for any direct, indirect, incidental, consequential, or special damages arising out of or in connection with the use of, or reliance upon, this article.
Readers are strongly advised to seek independent legal advice from a qualified professional before making any decisions or taking any action based on the contents of this article. Reliance on any information provided in this article is strictly at the reader's own risk.
By accessing and using this article, the reader expressly agrees to the terms of this disclaimer.