AI Document Review in Michigan Discovery (TAR)
A serious injury case can generate hundreds of thousands of emails, text messages, claim notes, and electronic files. Reading every page by hand is slow and expensive, so lawyers on both sides increasingly turn the first pass over to software that learns from a small set of human decisions and then ranks the rest of the documents by likely relevance. That method, technology-assisted review, is a form of applied artificial intelligence, and Michigan courts measure it against the same discovery rules that govern any other production. This guide explains how TAR fits Michigan practice, what the proportionality standard requires, and where a lawyer’s own judgment still has to do the work.
What Technology-Assisted Review Actually Does
Technology-assisted review, often shortened to TAR or called predictive coding, uses a machine-learning model to sort a document collection. A lawyer who knows the case reviews a sample, marks each document relevant or not, and the model uses those decisions to score every remaining file. The review team then focuses on the documents the model ranks as most likely to matter. The goal is not to remove lawyers from the process. The goal is to point a limited amount of human attention at the records that count, and to do it in a way the producing party can defend later if the other side questions the search.
The two measures that describe how well a review performed are recall, the share of truly relevant documents the process found, and precision, the share of flagged documents that turned out to be relevant. Those terms come up whenever parties argue about whether a TAR process was adequate, so it helps to know them before a dispute starts. In meet-and-confer sessions, parties increasingly negotiate a target recall or a validation protocol rather than promise perfection, because no search method, whether manual, keyword, or TAR, finds every relevant document.
The Michigan Rule That Governs the Search
Michigan rewrote the scope of discovery effective January 1, 2020. Under MCR 2.302(B)(1), discovery now reaches any nonprivileged matter that is relevant to a claim or defense and proportional to the needs of the case. The rule directs the court to weigh the importance of the issues, the amount in controversy, the parties’ relative access to information, their resources, and whether the burden or expense of the proposed discovery outweighs its likely benefit. The older phrase that allowed anything “reasonably calculated to lead to the discovery of admissible evidence” is gone.
That proportionality test is exactly where TAR earns its place. When a document set is large enough that manual review would cost more than the case can bear, a producing party can argue that an AI-assisted search is the proportional way to meet its obligations. The same rule cuts the other direction too: a party cannot demand a hand review of millions of files when a well-run TAR process would answer the question at a fraction of the cost.
Is TAR Allowed? What the Courts Have Said
No Michigan appellate decision has yet set out a published standard for predictive coding, so Michigan lawyers look to the federal courts that built this body of law. Michigan trial courts routinely treat these federal TAR decisions as persuasive authority when they weigh whether a proposed protocol is reasonable. The first judicial opinion to approve TAR was Da Silva Moore v Publicis Groupe, 287 FRD 182 (SDNY 2012), where the court accepted predictive coding as more appropriate than keyword searching under the facts of that case. Three years later the same judge wrote in Rio Tinto PLC v Vale SA, 306 FRD 125 (SDNY 2015), that it was by then “black letter law” that courts will permit a party to use TAR, and that it is wrong to hold the method to a higher standard than keyword searching or manual review.
The practical lesson for a Michigan case is that the choice of TAR is usually defensible, but the process behind it has to be sound. A court is far less interested in which software a party used than in whether the search was reasonable, documented, and proportional. A party that can show how it trained the model, how it validated the results, and why the outcome was reasonable is in a strong position. A party that treats the tool as a black box and cannot explain its own process is not.
The Lawyer Still Owns the Result
AI does not transfer professional responsibility to a vendor or a model. The duty of competence requires a lawyer to understand enough about the available review methods to choose a reasonable one and to supervise its use. When a lawyer relies on an e-discovery vendor or a software platform to run the review, MRPC 5.3 places the obligation to supervise that nonlawyer assistance squarely on the lawyer. The output still has to be checked, because a model trained on careless decisions will reproduce careless results, and a missed production can draw the same consequences as any other discovery failure.
This is the same verification principle that runs through every responsible use of AI in a law practice. A tool can draft, sort, and suggest, but a lawyer has to confirm before relying. We walk through how that duty plays out when AI writes text rather than sorts it in our guide to AI hallucinations and court sanctions in Michigan.
Preservation Comes First
None of this matters if the data is gone. The duty to preserve relevant electronically stored information attaches once litigation is reasonably anticipated, well before any TAR protocol is designed. Auto-delete settings, overwritten phone backups, and routine email purges can erase the very records a review would have surfaced, and destroying that material can expose a party to sanctions. We cover the duty to preserve, the adverse-inference instruction, and the range of penalties in our guide to spoliation of evidence in Michigan. Lock down the data first, then decide how to review it.
| Stage | What the lawyer controls | Key authority |
|---|---|---|
| Preserve | Issue litigation hold; stop auto-delete; protect ESI sources | Common-law duty to preserve; sanctions under MCR 2.313 |
| Scope | Define what is relevant and proportional; identify inaccessible sources | MCR 2.302(B)(1); MCR 2.302(B)(6) |
| Review | Train and validate the TAR model; document the process | Da Silva Moore (2012); Rio Tinto (2015), persuasive authority on TAR protocols |
| Supervise | Oversee the vendor and verify the output before producing | MRPC 5.3; duty of competence |
How TAR Helps an Injured Client
For a plaintiff, TAR is most useful on the receiving end. When a trucking company or a hospital produces a large volume of records, a plaintiff’s team can run the same kind of AI-assisted review to find the handful of documents that decide the case: the maintenance log, the internal email, the policy that was ignored. Used well, the technology lowers the cost of holding a well-funded defendant to account, which is the point of the proportionality rule in the first place. The flip side is that AI-generated work product and AI-built exhibits are themselves discoverable and subject to authentication, a subject we address in our guide to AI-generated evidence in Michigan courts.
अक्सर पूछे जाने वाले प्रश्न
Is technology-assisted review allowed in Michigan cases?
There is no Michigan appellate rule barring it, and federal courts have approved predictive coding since Da Silva Moore v Publicis Groupe, 287 FRD 182 (SDNY 2012). What a court examines is whether the review process was reasonable, documented, and proportional under MCR 2.302(B)(1), not which brand of software produced it.
Does using AI to review documents lower my discovery obligations?
No. The duty to produce relevant, proportional material is the same. TAR is a method for meeting that duty more efficiently, and the producing party still has to validate the search and stand behind the result.
Can the other side force me to do a costly manual review instead?
Not if a well-run TAR process is proportional to the needs of the case. MCR 2.302(B)(1) weighs burden against benefit, and MCR 2.302(B)(6) protects sources that are not reasonably accessible because of undue cost, subject to the court’s power to shift costs.
Who is responsible if the software misses a document?
The lawyer. Under MRPC 5.3 a lawyer must supervise nonlawyer assistance, including software vendors, and the duty of competence requires checking the output. A model cannot absorb the consequences of a deficient production.
What should I do with my own data when a claim begins?
Preserve it. The duty to preserve relevant electronically stored information starts when litigation is reasonably anticipated. Turn off auto-delete, save text messages and backups, and tell your lawyer where the records live before anything is reviewed.
Up against a large insurer or corporation with mountains of records?
Attorney Manny Chahal uses modern review tools to find the documents that decide a case. Free statewide consultation. No fee unless we recover.
1-844-624-2425 पर कॉल करें