When a Computer Output Becomes an Expert-Like Inference
The proposed Rule 707 turns on a functional boundary: did the system record or calculate information, or did it interpret information and produce a conclusion? That question matters because the reader needs to know when an AI output should be treated as an expert-like inference rather than as an ordinary machine record. The boundary is not determined by a product name, a file format, or the fact that a person operated the software. It is determined by the work the system performed in the particular case.
That approach avoids two opposite mistakes. Treating every digital output as expert evidence would make ordinary measurement look more mysterious than it is. Treating every computer result as a neutral fact would allow a predictive or classificatory conclusion to bypass the reliability questions applied to a human expert. The source materials point toward a more disciplined inquiry: identify the operation first, then decide what kind of explanation the court needs.
“Machine-Generated” Describes the Source, Not the Reasoning
The phrase “machine-generated evidence” sounds precise until one asks what it includes. A spreadsheet may add figures. A GPS device may report coordinates. A timestamp may be stored in metadata. A sensor may register temperature, speed, pressure, or another physical condition. A digital image may preserve what a camera captured. These outputs are produced through machines, but that fact alone does not tell us whether the machine made an expert judgment.
The New York City Bar’s comment makes this point directly. Nearly all modern evidence is machine-generated in some sense, but many outputs come from deterministic and relatively transparent processes that resemble measurement more than analysis. Their ordinary failure modes may be mechanical: a device was not calibrated, a sensor malfunctioned, a clock was wrong, or a recording was mishandled. Those problems can still matter. They simply are not the same problems as asking whether a model’s assumptions, training data, or inferential method support a conclusion about a person or event.
A measurement is not automatically reliable, and an inference is not automatically unreliable. The distinction identifies the kind of reliability showing that may be needed. A thermometer can be tested against a known standard. A GPS result can be examined through the device, signal, and recording conditions. A spreadsheet calculation can be checked by reproducing the arithmetic. These checks may be difficult in a particular dispute, but they usually ask whether a defined process accurately recorded or transformed inputs.
An inferential output adds a further layer. The system may decide which features matter, compare a new item with a reference population, estimate a probability, classify a pattern, reconstruct missing content, or generate a recommendation. The conclusion is not simply present in the input. It depends on a method for moving from input to meaning. That method may involve statistical assumptions, a trained model, a set of rules, or a combination of them. The proposed Rule 707 is concerned with the reliability of that movement.
The examples in the public comment show why the category matters. Probabilistic genotyping does not merely preserve a DNA signal; it produces an interpretation of how likely a profile is under competing hypotheses. Facial or voice recognition does not merely store a face or voice; it proposes a match. Predictive analytics does not merely report historical records; it draws a conclusion about a likelihood or future event. AI-based image or audio reconstruction may create a result that appears to fill a gap in the source material. A generative-AI answer may present a factual or legal conclusion in fluent language even when the system cannot reliably support it.
These examples share a feature: a factfinder may be tempted to accept the output as though it were an answer supplied by a knowledgeable analyst. The risk is not limited to technical opacity. An output can be easy to read and still difficult to justify. A polished conclusion may conceal incomplete data, a mismatched population, an inappropriate use, or a method that performs differently outside the conditions in which it was tested. Ease of presentation is therefore not evidence of reliability.
The City Bar’s proposed phrase, “machine-generated inferential evidence,” aims at this problem more directly than the broader label. It focuses attention on the accuracy and reliability of an inferential output instead of treating the machine as the legally important feature. The committee’s revised draft took a different drafting route by focusing on artificial intelligence and using a statutory definition, but the underlying problem remains similar: a court must identify when a system is doing work analogous to expert reasoning.
This is why no simple list of devices can settle the question. The same category of software may be used for retrieval in one case and analysis in another. A system may perform a deterministic calculation at one stage and use a model to interpret the result at the next. A product marketed as an extraction tool may include optional analytics. A human may use a general-purpose program to create a conclusion that the program itself did not label as an opinion. The relevant unit is the operation that produced the offered conclusion, not the brand printed on the screen.
The Same Tool Can Measure in One Use and Infer in Another
The committee’s discussion of Cellebrite provides a useful example because it avoids a misleading software-versus-nonsoftware divide. If a tool is used simply to extract or mirror information from a phone, the output may function like a record of what was on the device. The court may still need to consider ordinary questions about accuracy, handling, and authenticity, but the operation is not necessarily an expert analysis. If the extracted information is then fed into a process that identifies a person’s key contacts or infers relationships, the system has performed a different job. It has moved from retrieval toward interpretation.
The person presenting the result may be the same technician in both situations. That person’s title does not resolve the evidentiary question. In the first use, they may be able to explain how the data were extracted and preserved. In the second, they may know how to run the analytical function without knowing whether its model is valid, what comparison population it uses, how it handles missing information, or how often it produces a mistaken relationship. The proposed framework looks through the operator to the reasoning that produced the conclusion.
The difference can be described as a change in the path from input to claim. In a simple extraction, the intended claim may be: “These records were present on this device and were copied by this process.” In an analytical use, the claim may become: “These records show that this person was a key contact,” or “This pattern indicates a relationship.” The latter claim requires an additional bridge. Someone—or something—has selected a pattern, assigned it significance, and connected it to a proposition in the case.
A useful way to locate that bridge is to separate four layers that are often compressed into one exhibit. The first is the underlying material: a voice recording, image, message, location point, or collection of records. The second is the technical transformation: copying, filtering, comparing, measuring, or reconstructing the material. The third is the system's output: a match score, classification, probability, predicted relationship, or generated passage. The fourth is the legal claim for which a party offers that output. The first two layers may be largely mechanical while the third is inferential; the fourth may go further still if a lawyer treats a probability as proof of an historical fact.
That distinction matters for numerical outputs. A number can look like a measurement while actually expressing a model's assessment. A reported probability is not necessarily a physical property discovered in the evidence. It may depend on a reference population, assumptions about competing explanations, choices about which features to compare, and the way the system translates those choices into a score. The number may be useful and well supported, but its meaning cannot be read from its formatting alone. The proponent must be able to say what the number represents and what it does not establish.
The same discipline applies to a generated narrative. A fluent answer may combine retrieved material, statistical associations, and generated language. Its readability does not show that each factual proposition was verified, that the system stayed within its intended purpose, or that the conclusion can be reproduced. The court therefore needs to know whether the offered claim is the system's direct output, a human interpretation of that output, or a combination. Otherwise a chain of technical steps can disappear behind a single sentence that sounds like ordinary testimony.
This layered view also explains why the legal question cannot be answered by asking whether the software uses machine learning. A system may draw an inferential conclusion through a statistical model, a rule-based process, a reconstruction procedure, or a hybrid pipeline. Conversely, a tool that contains sophisticated components may be used only to retrieve or calculate information in a particular instance. The relevant question remains the relationship between the operation and the claim offered to the factfinder. Classification must follow the evidence's use, not the marketing vocabulary attached to the tool.
That bridge is where function creep can occur. A system designed for one purpose may be used for another because its output appears convenient or persuasive. A tool validated for identifying a narrow feature may be used to make a broader judgment. A process designed to assist investigation may be presented to the jury as though it had established a fact. The official report’s proposed note identifies use for unintended purposes, analytical error or incompleteness, inaccurate or biased data or formulas, and lack of interpretability as reliability concerns. Those concerns are easier to see when the court asks what conclusion the system was actually used to support.
Opacity makes the classification especially important. A routine measurement may be checked through a visible calibration chain. An inferential system may not reveal which inputs drove a result, which alternatives it considered, or how it handles a case unlike its development data. Unknown error rates and a demonstrable risk of fabrication or hallucination are not proof that a particular output is wrong. They are reasons not to treat the output as self-explanatory. The party offering it may need to supply a method-based account rather than rely on the fact that the result came from software.
The proposed Rule 707 does not say that an inferential system must be transparent in the sense that every reader can understand its internal calculations. Its trigger is the more practical counterfactual: if the conclusion were offered by a human witness, would it be subject to Rule 702? If yes, the proposed rule would bring the AI output into a reliability framework even when the output is entered directly or accompanied only by lay testimony. The question is therefore not whether the machine can explain itself in ordinary language. It is whether the proponent can establish the reliability of the method that generated the conclusion.
The distinction also clarifies why a business-record label may be insufficient. A business record can establish that an organization kept or generated a record in the ordinary course, subject to the requirements that apply to that exception. But the fact that a system routinely produces a report does not answer whether an analytical conclusion inside the report rests on adequate data or reliable reasoning. Routine production and reliable inference are different properties. The proposed rule is intended to stop the second from disappearing inside the first.
The category boundary can be represented as a sequence rather than a catalogue:
The diagram does not create a bright-line exemption for measurement. It shows where the proposed Rule 707 question becomes more likely to matter. A measurement-like output may still be challenged for accuracy or authenticity. An inferential output may ultimately be shown reliable. The point is to prevent the court from asking the wrong preliminary question—whether a machine was involved—instead of the right one: what reasoning, if any, did the machine contribute?
This functional test also explains why the proposed rule’s definition and scope remain contested. The committee’s revised text uses artificial intelligence rather than the earlier broad “machine-generated” language, but the source materials acknowledge that many courtroom technologies combine AI and other processes. A statutory definition may help identify the family of systems under discussion, yet it cannot by itself resolve whether a particular output was a measurement, a calculation, or an expert-like inference. That work still depends on the facts of the offered evidence.
The classification has a direct consequence for the burden question. If the output is only a record or a straightforward calculation, the proponent may face ordinary accuracy and authenticity questions rather than the full expert-like inquiry described in the draft. If the output is a conclusion that would require expert reasoning from a human, the proposed Rule 707 treats the delivery channel as secondary. The offering party must be ready to establish why the conclusion is fit for the dispute and reliably produced.
The next step is to make that statement precise. Once an output has crossed from recording into expert-like inference, “reliable” cannot remain a general compliment. The proposed Rule 707 breaks the showing into four linked questions about usefulness, inputs, method, and application. Each question targets a different way for an apparently authoritative output to fail.