A phone is seized in a federal investigation. Investigators run forensic software. It pulls call history, messages, location data and browser records.
Nothing new there. Digital evidence has moved through federal courts for decades.
Now change one thing.
The software does not just extract the data. It analyzes it. It finds patterns across thousands of records. It compares relationships. It decides which contacts matter most. The government offers that conclusion at trial.
Who is the expert?
That question sits behind proposed Federal Rule of Evidence 707. It is still a working draft. It may change. It may never become a rule.
The problem it identifies will not go away.
Rule 702 governs expert testimony. A witness with specialized knowledge cannot simply offer a conclusion. The party offering the testimony must demonstrate to the court that it is more likely than not that the testimony will help the factfinder, rests on sufficient facts or data, comes from reliable principles and methods, and reflects a reliable application of those methods to the case.
AI can break that structure. The conclusion can arrive without the expert.
A technician can run a program without knowing why its output is reliable. A police officer can receive an analysis without understanding the model behind it. A company can buy a system whose inner workings stay with the vendor. The output may still do exactly the analytical work that would trigger Rule 702 if a person had done it.
Proposed Rule 707 asks a simple question. Should the reliability requirement follow the conclusion, even when no expert comes with it?
The Department of Justice and the federal defenders disagree on whether courts need that rule.
DOJ says the problem is mostly anticipatory.
The federal defenders say it is already here.
A rule for a problem that may not exist
The Advisory Committee on Evidence Rules approved proposed Rule 707 for public comment in 2025. It would have applied Rule 702's admissibility requirements to certain machine-generated evidence offered without an expert witness.
The response was split.
The Committee received 59 written comments by the deadline and held two public hearings. Its summary counted three as unqualified support, 27 as support with revisions and 27 as opposition.
Critics went after the core of the proposal. "Machine-generated evidence" was too broad. The exception for basic scientific instruments was vague. The rule might open a new path for admitting AI evidence without an expert. And it was premature.
The Committee did not recommend final approval. It withdrew the version issued for public comment and continued studying a substantially revised proposal.
That matters. Federal courts have not decided how AI evidence should be treated under a new Rule 707. Rulemakers are still working out what the problem is, how often it occurs and whether a new rule is the right fix.
At the May 2026 meeting, DOJ argued the existing rules may be enough. Its representative said the problem of AI evidence offered without an expert remained anticipatory and had yet to materialize in court. Trial judges, DOJ argued, already have tools to require experts or additional information when necessary.
The federal defenders disagreed. Their representative told the Committee they are seeing this kind of evidence in federal cases now.
That is an unusual posture for a rule. The fight is not only over how to solve a problem. It is over whether the problem has materialized.
The Cellebrite line
The May meeting produced the clearest example of what is at stake.
Law enforcement uses Cellebrite to extract information from phones and other devices. A conventional extraction can produce call records, location data and browser history. The federal defender representative acknowledged that an extraction that simply mirrors a device's contents may not need an expert foundation.
Analysis is different.
Professor Andrea Roth of Berkeley Law drew the same line. Extracting information from a phone is one problem. Using software to decide who is a "key contact" is another. Under current law, she noted, that analytical conclusion could come in through a lay witness. It would skip the reliability scrutiny Rule 702 applies to a human expert.
Rule 707 does not formally divide machine evidence into observation and inference. The Committee has not written that test. But the distinction runs through the problem the Committee is trying to solve.
A system that retrieves information tells you what it found.
A system that analyzes information tells you what it thinks the information means.
The law knows how to challenge the person who does the second job. That person can be qualified. The method can be examined. Assumptions can be exposed. Error rates can be argued. The expert can be cross-examined.
The machine changes that.
The conclusion remains. The expert may not.
The witness who isn't there
The September working draft makes the concern explicit.
Under the version now before the Committee, Rule 707 would apply when evidence is the product of AI, is offered without an expert witness and would be subject to Rule 702 if a person testified to it. The party offering it would have to establish that the evidence helps the factfinder, rests on sufficient facts or data, comes from reliable principles and methods, and reflects a reliable application of those methods.
The rule does not ask whether the machine got the right answer in a given case. Courts do not require human experts to be infallible either. It asks whether the process is reliable enough to put the conclusion in front of the factfinder.
The new draft goes further than the version published for comment. Ordinarily, the proponent would need an expert who can explain how the AI system reliably produced the evidence. Only in exceptional circumstances could other proof, such as persuasive validation evidence, do the job.
Picture the courtroom this anticipates.
The operator can explain what went into the system. The investigator can explain how the output shaped the case. The vendor can describe the product in general terms. None of them necessarily knows why the system reached this conclusion.
There may be no human who can fully explain or defend the analysis.
That is where Rule 707 stops looking like a narrow evidentiary amendment. It is an early attempt to deal with a much larger shift.
Institutions increasingly separate the person who acts on a conclusion from the process that produced it. A fraud analyst gets a risk score. An investigator gets a relationship map. A security analyst gets an alert labeling activity malicious. Each may understand the surrounding facts and know very little about the machinery behind the conclusion.
That works until someone asks the institution to prove the conclusion deserved trust.
Litigation asks that question.
The other AI problem
Rule 707 is not the only place the Evidence Rules Committee is wrestling with artificial intelligence.
A separate working draft, Rule 901(c), deals with deepfakes. The problem there runs in the opposite direction.
With Rule 707, everyone knows the machine produced the evidence. The question is whether its analysis is reliable.
With Rule 901(c), the machine's involvement may be hidden. A video, recording, image, document or other item appears authentic. One side says generative AI fabricated it.
The Committee's proposed answer is a two-step process. A party cannot trigger the special inquiry merely by calling evidence a deepfake. The challenger must first present evidence sufficient to support a finding of fabrication. If that threshold is met, the party offering the evidence must then demonstrate to the court that the item is more likely than not authentic.
The two proposals solve different evidentiary problems. Together, they reveal the same institutional anxiety.
Courts have spent decades asking people to stand behind evidence. A witness authenticates the recording. An expert stands behind the analysis. Cross-examination tests both.
AI can disrupt either side of that arrangement.
It can create evidence without revealing that it created anything.
Or it can openly produce an analysis without leaving anyone who can fully explain or defend the conclusion.
Professor Daniel Capra, the Committee's longtime Reporter, has been working through both problems. His Rule 707 materials expressly identify the inability to cross-examine machine output as a concern. His memoranda otherwise treat the two problems separately for good reason. Deepfakes raise questions of authenticity. Machine-generated analysis raises questions of reliability.
One asks whether the thing in front of the court is really what it appears to be.
The other asks whether a conclusion deserves trust when no person actually reached it.
What did the machine actually do?
The version of Rule 707 published for public comment used the phrase "machine-generated evidence." Public comments exposed the problem fast.
Machines generate almost everything.
Excel spreadsheets. Google Maps. Digital thermometers. Emails. Cellphone extractions. Modern software also refuses to stay in neat categories. One product may contain deterministic code, statistical models, machine learning and generative AI.
The revised draft shifts to "artificial intelligence" and tries to define it.
That fixes one problem and creates another.
The working definition describes a machine-based system capable of analysis or of making predictions, recommendations or decisions. Even that language is not settled. Part of the definition remains bracketed, and the Committee has specifically asked its October expert panel whether the definition captures the technology that will actually show up in evidence.
This is more than a drafting problem.
Take one forensic program performing two functions. First, it extracts the contents of a phone. Second, it analyzes those contents and maps relationships. The software is the same. The function is not.
Take a cybersecurity platform. One component records that a process opened a network connection at a specific time. Another correlates thousands of events and concludes an account is probably compromised. A third uses a language model to summarize the incident.
Take an emergency department. One system displays lab results. Another combines those results with vital signs and history to calculate risk. A third uses that score to move a patient ahead of others waiting for care.
All of it is software. All of it may involve AI somewhere in the stack. None of it is doing the same thing.
Calling the whole platform "AI" tells us very little.
The better question is functional. What did the machine do?
Did it record an event? Retrieve information? Calculate a deterministic result? Classify risk? Rank people? Identify a pattern? Predict? Infer? Recommend?
Rule 707 does not adopt this test. The Federal Rules may draw the line somewhere else. But the Committee's struggle to define AI shows why function matters. As AI gets built into ordinary software, regulating the label gets harder. Scrutinizing the function may hold up better.
Lawyers should recognize this. Evidence doctrine has always cared less about what something is called than about what it is offered to prove.
The questions behind the questions
The draft Committee Note points to where this goes next. It asks whether inputs are sufficient and representative. It asks whether the system was validated under similar conditions. It asks whether independent evaluators can access it, and whether the opposing side has enough information to understand how it works.
Then it asks two questions that reach well past the courtroom.
What records does the system keep and delete?
Can the output be reproduced or audited?
Every organization relying on consequential AI outputs may eventually have to answer those. Many have not started.
Too early, until it isn't
DOJ has a serious argument.
Courts should be careful about technology-specific rules written before the problem matures. AI is changing fast. Definitions age badly. Trial judges already have tools, DOJ argues, to require experts or additional information when necessary. A premature rule could sweep in technology it was never meant to reach and impose new costs before the need for it is clear.
Rule 707's own history supports that caution. It has already required substantial revision after one round of public comment.
The federal defenders have a serious answer.
Waiting has costs too.
If software performs analysis that would require Rule 702 scrutiny from a person, the missing expert should not become the reason the analysis escapes scrutiny. That concern is especially acute in criminal cases. A defendant may have to challenge a conclusion from a system the defendant did not choose, may not be able to inspect and may not understand.
The two positions are closer than they look.
Neither side wants unreliable AI conclusions in federal court. They disagree on whether existing rules can catch them, and whether a new rule would help or hurt.
On October 15, the Advisory Committee takes up the question again. The working draft may advance, change or stay on the shelf.
Either way, the Committee has already exposed the deeper issue.
The modern law of expert testimony is built around a human witness who can be qualified, examined and cross-examined. The expert applies a method, reaches a conclusion and comes to court to defend it.
AI can break that chain.
The investigator may have the conclusion but not the model. The vendor may know the model but not the case. The developer may know the architecture and still be unable to explain a particular output. In some systems, nobody may be able to reconstruct exactly why a result occurred.
The law can decide whether to admit that output. It can decide what foundation it needs. It can decide whether an expert must come with it.
First it has to answer a simpler question.
When the analysis comes from a machine, who stands behind the conclusion?
That question will not stay in the courtroom.
Next in Deep Signal: Reconstruct the Machine. An AI conclusion made today may have to be defended years from now. What would that take? And why might a faithful record of what happened matter more than making the machine say it again?