I. Introduction
A defendant’s location at the time of a crime may be established either by a GPS record authenticated through human testimony or by the independent output of an autonomous AI system. Across the United States, India, and the United Kingdom, existing evidentiary doctrine often makes it easier to admit the latter.
Hearsay doctrine evaluates claims by evaluating their human declarant, the swearing of the oath, the opportunity for confrontation and cross-examination. Generative AI does not have a human declarant to evaluate. The courts did not respond by making an examination of the output of autonomous AI systems more rigorous, but rather by taking autonomous AI outputs out of the scope of hearsay doctrine. The outcome is a paradox in the core of modern evidence law: the more autonomous the system, the less rigorous the evaluation of its claims.
It is not just a thought experiment. Recent cases such as Malia LLC v. State Farm 2025, and State v. Horcasitas (May 2024) illustrate the problem. The central legal difficulty is identifying which evidentiary rule the video violates.
Although there have been studies examining machine testimony and the reliability of algorithms, there is limited research focusing on the structure and relationship between machine autonomy and hearsay doctrine. The India’s Bharatiya Sakshya Adhiniyam 2023 is one example of how legislative bodies continue to apply an updated version of the law concerning evidence from computers despite the advancement in the form of artificial intelligence. This piece will argue that the issue does not stem from artificial intelligence technology, but rather from the hearsay doctrine’s reliance on human declarants.
II. The Hearsay Doctrine and the Human Declarant Assumption
The hearsay rule is epistemic in its rationale. Assertions made out of court are not admissible since the fact-finder cannot observe the demeanour of the declarant, cannot make the declarant swear before God, and the other side cannot cross-examine the declarant to find out about Wigmore’s four ‘testimonial infirmities’: defects of perception (did the declarant observe correctly?), memory (did they recall accurately?), narration (did they express themselves precisely?), and sincerity (are they telling the truth?).
A. The Infirmities Translated to Machine Context
These infirmities do not disappear in the absence of a human declarant; they relocate. Perception becomes sensor precision and calibration. Memory becomes data integrity and chain-of-custody preservation. Narrative becomes inference according to the model’s rules of inference, be that determinism or randomness, and Sincerity becomes susceptibility to bias in the training data and hallucination by the model.
This is very much in line with Professor Andrea Roth’s concept of “black box dangers”: machines introduce risks of a kind analogous to testimonial infirmities through their black boxes in terms of design, programming, and decisions. Brian Sites makes the same argument that evidentiary doctrine has to transcend the assumptions around human witnesses. This system-focused approach is also evident in other areas. For example, the EU Artificial Intelligence Act requires transparency, technical documentation, and human oversight for high-risk AI systems, signalling an emerging regulatory preference for assessing the reliability of systems rather than attributing responsibility solely to human actors.
Yet the procedural safeguard does not translate. Cross-examination exploits fear, inconsistency, and deception in human witnesses. Source code cannot be embarrassed; training data cannot be sworn in.
B. Why the Human Declarant Assumption Is Load-bearing
The key issue here is that the lack of a human declarant has become a reason to bypass hearsay scrutiny altogether. Yet removing the declarant does not eliminate reliability concerns; it merely changes their source.
According to The U.S. Advisory Committee on the Federal Rules of Evidence 2025-2026 Report, the main problem with respect to machine outputs consists of the impossibility of cross-examination of an algorithm. Nevertheless, such a conclusion solves only half of the problem. The question arises as to why machine outputs are exempt from hearing scrutiny when there is no human declarant.
III. Three Jurisdictions, Three Inadequate Answers
There are three advanced legal systems that reach the same failure point by completely different paths of doctrinal development.
A. India: BSA 2023 and the Generative AI Silence
The Bharatiya Sakshya Adhiniyam 2023 is the biggest reform of evidence law in India since 1872. The treatment of electronic records as provided for in Section 63 (formerly Section 65B IEA), follows a similar approach to the one introduced via the amendments to the 2000 IT Act: electronic records are admissible provided there is a certificate regarding the conditions of production in the computer system.
The Supreme Court’s construction of this framework in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020) reaffirmed Anvar P.V. v. P.K. Basheer (2014), settling the longstanding uncertainty over whether a certificate under Section 63 was mandatory, it is. But that framework was built around records a computer passively stored: call logs, transaction records, server entries. It assumes a human who can attest to the computer’s ‘proper working condition’ and to the correspondence between the input and output data, verifying the hardware condition of the server, not the epistemic reliability of what a model produces.
However, the failure in this statutory framework is in the wording of ‘computer output’. As the IT Act 2000 states, this term means ‘information produced by a computer in the course of its operation’, implying that there is a deterministic correlation between input and output. A large language model has an inverse approach, where the output is the result of billions of learned parameters and stochastic sampling, not deterministic correlation between input and output, which also depends on temperature tuning, tokenization, and training data. No human can certify that the output corresponds to the input in any meaningful sense; the certificate mechanism becomes a problem. More basically, there is the issue of what ‘output’ means for a system that creates new material. As per Section 63, it refers to ‘information contained in electronic form’ and ‘computer output,’ which assumes that the information is stored and can be retrieved rather than generated or interpreted. The creation of a victim impact video by an AI, the report produced by a fraud detection algorithm, or the decision made by a deepfake detection system is not covered by the statute, leaving courts to improvise.
B. United States: The ‘No Declarant’ Safe Harbor
The courts of America have a well-established framework in relation to legacy machine outputs since only a ‘person’ can be a ‘declarant’ under FRE 801(b); hence, machine outputs cannot be considered hearsay. The decision in United States v. Washington subsequently adopted laboratories, GPS coordinates, and any form of automated reporting, all of which were understandable given the technology’s passive nature; however, when applied to generative AI, the limited hearsay rule becomes a broad exemption from hearsay.
The Malia LLC v. State Farm provides a clear example of how the reasoning applies to generative AI outputs. The Supreme Court ruled that the report generated by AI was not hearsay, given the circuit rule that ‘machine-generated information does not constitute a statement’ under Rule 801, without considering the model’s training data, architecture, or error rate.
This approach conforms to American evidence law, which now draws distinctions between authentication issues and substantive reliability issues. The human expert who analyses the same evidence and prepares the same report must fall within the business-record exception or another hearsay exception, whereas the AI program need not meet this criterion. Oversight makes it difficult for evidence to be admitted; lack of oversight makes it easy for evidence to be admitted.
C. United Kingdom: The Mechanical Device Carve-Out
Section 129 of the Criminal Justice Act 2003 makes clear that some statements mechanically generated are not hearsay where the statement made is not dependent on the assertion of the information supplier. This was in response to the particular factual situation of R v. Cochrane (1993), where it was determined that the statements made on an ATM were outside the scope of the hearsay rule as the machine recorded only objective information.
This applies well in cases involving recording devices but not in systems where the “statement” is made through inference, synthesis or probabilistic processes. The output from an Large-Language Model is not merely a mechanical transcription of input data but a new creation formed through the training parameters, sampling process and the prompt structure itself. There is no good fit between the “depends on the information supplied” test and outputs that cannot be determined from the inputs, even in theory.
In the UK Ministry of Justice’s 2025 policy paper on software-generated evidence, this gap is acknowledged but no call is made for legislation in this area, with courts being referred back to existing authentication standards. Zhou states that the Criminal Justice Act hearsay provisions offer little help in dealing with conversational AI systems capable of producing context-dependent, generative outputs.
However, authentication merely confirms that an output is what it is asserted to be and nothing more. It does not confirm the AI system’s reasoning. This way, the UK evades the hearsay issue simply by diverting attention to authentication, with the reasoning of the AI remaining unexamined.
Each jurisdiction uses the Indian certification, the lack of a declarant in the US, and mechanical generation in the UK through different doctrinal approaches. Yet, the effect across all three is the same: AI outputs that have greater autonomy than ever before are subjected to less scrutiny than human-generated evidence.
IV. The Perverse Incentive Paradox: Mapping the Problem
A. The Laundering Mechanism
Assertions made out of court are excluded from evidence under the hearsay rule due to their unreliability, which usually cannot be tested through confrontation and cross-examination. What happens in the case where the ‘declarant’ is not human?
The “No Declarant, No Hearsay” principle was supposed to be applicable to technologies that would just document objective information. The presence of generative artificial intelligence negates this assumption. While the function of the thermometer and the GPS device will be limited to recording factual information, the Large-Language Model will have the ability to synthesise, infer, summarise, and narrate information. Yet, because of the lack of a declarant who speaks the language, the output of these technologies is not subject to the hearsay rule.
Thus, there is a “doctrinal inversion”, where the more autonomous the AI system becomes, the less human intervention there is in producing its outputs, the easier it is to be exempted from the application of the hearsay rule.
Here is the laundering process: claims made by humans are input into AI programs as prompts, reconstituted as machine outputs, stripped of their hearsay tag, and introduced without being subjected to any tests applicable to human claims.
Professor Andrea Roth has noted the structure of this incoherence in the way that courts are trying to “shoehorn” the new machine evidence into old frameworks, which is intellectually incoherent and fails to provide juries with the means of assessing machine credibility. According to The Advisory Committee on Evidence Rules, there was the same recognition of this gap in January 2026, when machine-generated analysis provided by a lay witness, and not an expert, fell outside Rule 702’s reliability requirements.
B. Case Studies: Horcasitas and Malia LLC
Two recent cases make the problem concrete.
The Malia LLC v. State Farm (2025) – a case where an AI-produced wind verification report was allowed into evidence. That judgment was based on the reasoning that “the consensus among the circuit courts is that machine-generated data is not hearsay because it is not a statement under Rule 801.” In this instance, the AI-produced report escaped the hearsay objection without its reliability being examined, without being sponsored by an expert witness, and without consideration of how the algorithm reached its conclusion.
The Horcasitas (Arizona, May 2025) is even more remarkable in this respect. The victim impact statement in open court, the face and voice of murder victim Christopher Pelkey, was presented by an artificial intelligence reconstruction of the victim. The sentencing judge even commended the AI presentation and imposed a punishment exceeding that requested by the prosecution. The defence has appealed, asserting that undue weight was given to the AI presentation in influencing the judgment. Hearsay objections would be difficult to make to this evidence because the AI, not an individual, made the “statements.” However, the AI was creating statements that Pelkey’s relatives believe he would have made, which is precisely the sort of substituted assertion that hearsay rules sought to prevent.
This is not an exception either. For instance, in State v. Puloka (Wash. 2024), the court held that the video evidence was not admissible since it failed the test of reliability, while in Matter of 2024 NY Slip Op 24258 (Surr. Ct., Saratoga County, Oct. 10, 2024), expert evidence concerning Microsoft Copilot.
C. The Confrontation Clause and Natural Justice Dimensions
The problem deepens in criminal proceedings. The Sixth Amendment Confrontation Clause provides the accused person with the right to confront the witnesses who have made statements against him/her. With regard to the Crawford v. Washington (2004) case, testimonial evidence requires confrontation and only confrontation, since the right to confrontation is a procedural right and not a reliability right. Smith v. Arizona (2024) protects the fact that no expert witness could give testimonial evidence of non-testifying analysts as the basis for his/her independent opinion. It is doubtful whether the same applies to an AI system.
The case State v. Loomis (Wis. 2016) made the issue very clear. The appellant maintained that the sentence should be reversed because the COMPAS algorithm used was closed source and could not be reviewed or challenged. The Wisconsin Supreme Court rejected this argument, holding that COMPAS could be used at sentencing subject to certain disclaimers and safeguards; the U.S. Supreme Court subsequently denied certiorari. No disclaimer could lead to “meaningful judicial scepticism” because judges are unable to evaluate the software and are under pressure to use it.
V. Toward a Two-Tier Admissibility Framework
It is not a solution to graft onto machines the same processes that apply under the hearsay rule’s machinery. Rather, it is necessary to determine the function of the hearsay rule and to develop parallel protections in the context of AI output.
Rule 707 of the Federal Rules of Evidence, approved by the Advisory Committee in June 2025, is one of the most important developments in this regard. Essentially, Rule 707 states that if the output provided by the machine would have been governed by Rule 702 had the analysis been provided by a human expert witness, then the court will find that the reliability requirements of Rule 702 are satisfied despite the absence of an expert witness. Thus, the rule closes the gap in evasion noted by the Advisory Committee; it is no longer possible to evade the review of expert testimony merely by using an AI.
Nevertheless, Rule 707 is inadequate for the purposes of developing a fully-fledged system for two reasons. Firstly, as pointed out by the AAJ in its February 2026 comment letter on Rule 707, “this Rule should not require each reader to interpolate the language of Rule 702(a)-(d), the vocabulary of human expert witnesses, into the world of machines, models, and algorithms.” The rule adopts the wrong terminology. Secondly, Rule 707 does not discuss hearsay in any way. It is a reliability test, not a declarant one.
A two-tier approach addresses both issues.
Tier 1: Non-inferential outputs (data from sensors, transaction logs, GPS tracking logs, classifications): This output stays out of hearsay rule. It is not affected by that rule and will be authenticated by Rule 901 and the simple test of whether the device was working properly. This is what Rule 707’s “basic scientific instruments” exception wanted to capture.
The boundary between Tier 1 and Tier 2 turns on a single question: does the system exercise judgment, synthesis, or inference, or does it merely record? Roth’s distinction between opaque “black box” machines and those whose operation and reliability are transparent and provides a useful starting point for drawing that line.
Tier 2: Inferential/generative outputs (narratives, reconstructions, risk assessments, synthesised reports, statements from AI): These must be treated as being of equal value to the testimony of an expert witness. The proponent must reveal the particular model used, as well as the inputs/prompt and the version of the AI system. The output must meet a reliability threshold equivalent to the one set out in Rule 702(a)-(d). To the extent that the output is derived from human assertions fed into the system, those assertions must themselves meet hearsay standards or an exception.
In cases where the model used is proprietary, the proponent will be unable to satisfy Tier 2 through mere statement. Instead, the tier demands that the proponent provide external proof, such as published reports or compliance with a regulatory assessment, similar to what would be expected from an expert witness regarding the basis of his/her testimony. The proponent is not only concerned with the validity of the model (that is, the ability to accomplish the intended purpose) but also with the reliability (that is, the consistency in accomplishing that purpose).
When applied to the case of Malia LLC, this implies that the wind model’s methodology and margin of error should be revealed prior to the admissibility; when applied to the case of Horcasitas, this would imply overcoming the hearsay exception to the human claims that form the foundation of the script.
VI. Conclusion
The hearsay rule has never been about the word “declarant” in isolation. It has always been about a deeper concern when someone (or something) asserts that a party wants to use as proof, the opposing party deserves a meaningful opportunity to test its reliability. Cross-examination of human witnesses is one mechanism for that testing. It is not the only possible mechanism.
Generation AI does not render the hearsay rule obsolete. Instead, it makes the logic behind the rule even more pressing. Machines’ products “depend on the reliability of a source” because they could be incorrect “by intention, poor data training, bias in input, or hallucination.” While these dangers are very much real, it is just the method that should change.
This two-tiered approach does not deviate from the hearsay rule but merely translates it to a novel setting. The first tier will maintain the existing structure, which works fine for simpler recording systems. The second tier will translate the core of hearsay analysis into a test for evaluating the outputs that make assertions.
The same kind of translation is equally required in both India’s BSA 2023 as well as common law in the United Kingdom. Section 63‘s certification merely ensures that the computer was operating properly, but not that the output generated by the generative model is legitimate and does not launder human assertions. All three jurisdictions are facing the same issue, and the same solution is required in all of them: a special admissibility process that will not ask whether “this is made by a machine?” but “is this output making an assertion?”
Courts that deal with Horcasitas and Malia LLC today will be facing far more sophisticated AI-generated outputs tomorrow. The gap in the doctrine is wide open. If the hearsay rule continues assessing the admissibility by the presence of a human declarant rather than the reliability of an increasingly autonomous system, evidentiary law will be protecting the technology rather than court’s decision.
The question is no longer whether machines can speak. The question is whether the law can continue pretending that only humans make assertions.
(This post has been authored by Priyanshi Aggarwal and Komal Mirdha, third-year Ba.LLB (Hons.) student at Ram Manohar Lohiya National Law University, Lucknow.)
CITE AS: Priyanshi Aggarwal and Komal Mirdha, ‘Machine Outputs As Hearsay: Doctrinal Uncertainty And The Case For A Dedicated Evidentiary Framework’ (The Contemporary Law Forum 20 August 2026) <https://tclf.in/2026/08/20/machine-outputs-as-hearsay-doctrinal-uncertainty-and-the-case-for-a-dedicated-evidentiary-framework/> date of access.
______________________________________________________________________________