by David Kim

At the close of Derrick Ulett’s murder trial, his lawyer pointed the jury toward a gap in the People’s case. Ruben Alexandre had been shot in March 2008 in front of an apartment building at 48 St. Paul’s Place in Brooklyn, and that building had surveillance cameras in its lobby. If a camera had captured the shooting out front, wouldn’t the footage show who actually fired the gun? “We don’t have that video,” she told the jurors. The prosecutor stood and answered with common sense. Police had recovered footage from a camera around the corner, in the direction the shooter had fled. Wasn’t it common sense, she asked, that the jury would have seen a lobby video if one had existed?

One had existed. The prosecutor had watched it before trial. She later testified that she believed it was not Brady material because it did not show the shooter and the images were too “washed out” to identify anyone; she placed it in a box for “irrelevant” evidence and forgot about it. The District Attorney’s Office produced a copy years after the verdict, pursuant to a request under New York’s Freedom of Information Law.

The tape showed the importance of that decision. The People’s case had rested heavily on an eyewitness known as Cream, who had come forward 10 months after the murder while facing pending robbery and narcotics charges and who, after giving his statement, received a plea deal to reduced charges. The video appeared to depict Cream, Alexandre, and a third person together outside shortly before the shooting, along with other people entering and leaving the building. As Alexandre lay on the sidewalk, another figure hovered over him and extended an arm toward his body before leaving the frame.

More than eight years after the judgment, the New York Court of Appeals reversed Ulett’s second-degree murder conviction and ordered a new trial. People v. Ulett, 33 N.Y.3d 512 (2019). The Court held that the withheld video would have set the scene of the killing, impeached eyewitness testimony, identified other potential witnesses, including at least one person only a few feet away when the shots were fired, and supported an argument that someone else may have been responsible. The prosecutor’s summation, which denied the video’s existence, “compounded the prejudice.” Taken together, the suppression undermined confidence in the verdict. The new trial never occurred. On December 19, 2025, prosecutors dismissed the murder charge, ending a prosecution that had lasted more than 16 years and kept Ulett behind bars for roughly 11 of them.

Ulett is the visible version of an old problem. One camera, one recording, one prosecutor who set the evidence aside and assured the jury it did not exist. Brady v. Maryland, 373 U.S. 83 (1963), forbids precisely that. Since 1963, the government has been barred from suppressing evidence favorable to the accused and material to guilt or punishment, “irrespective of the good faith or bad faith of the prosecution.” A juror can grasp a hidden videotape. A reviewing court can watch it.

Digital discovery obscures that visibility. The government no longer has to bury the exculpatory item in a box marked irrelevant. Much like Poe’s purloined letter, it can hand the defense everything and conceal the item in plain sight. A mobile-device extraction can range from a limited logical collection to a far more complete file-system or physical image, producing gigabytes of data that may be rendered into lengthy reports. A cloud return, the data a provider such as Google or Apple produces under a warrant or other legal process, can be larger still. Add body-camera archives, jail calls, doorbell video, and location logs, and a case production can reach terabyte scale. The term is an abstraction to most people, so consider what it holds. A single terabyte can contain on the order of 10 million pages of documents, more than a thousand hours of video, or some mix of both. Reading one page a minute without ever stopping to sleep or eat, a person would need roughly 19 years to read 10 million pages. Watched around the clock, 1,000 hours of footage is six straight weeks in front of a screen. No prosecutor, defender, or judge reviews that end to end on ordinary litigation deadlines. In Ulett, a court could watch one tape and see the harm for itself. At terabyte scale, the harm is concealed somewhere inside a volume of evidence no single human could ever fully inspect.

That is the inversion this cover story discusses. Sometimes the failure is still classic suppression, with the government in possession of favorable evidence and declining to turn it over. Sometimes it is subtler and, for the defendant, no less damaging. The government discloses the haystack and leaves the defense to find the needle with no index, no native files, no metadata, no forensic software, no funded expert, and no time before the plea deadline. “We gave you everything” becomes a way of saying “we identified nothing.” Whether the burial is a deliberate tactic or the ordinary byproduct of volume, underfunding, and opaque software, the defendant is convicted, or pleads guilty, based on a file no one has fully viewed.

This is not a niche problem. Anyone whose case involved evidence from a phone, a cloud account, a doorbell camera, a body camera, a surveillance cache, a geofence return, a cell-site record, a license-plate reader, a social-media account, a jail call, or an algorithmic investigative tool has potentially faced some form of this. That now describes much of the criminal docket. Two failures Criminal Legal News (“CLN”) readers know well converge here, prosecutorial misconduct and ineffective assistance of counsel, and the damage runs downstream into coerced pleas and post-conviction litigation.

None of this is a forecast. The tools are in use, the volume is already here, and a disclosure doctrine built for a banker’s box has not caught up. What looks from the defense table like a single act of concealment is usually several distinct methods presenting as one. Telling them apart is where the analysis has to begin.

Five Ways Digital Evidence Disappears

Separate them, and five methods are revealed. Each looks identical from the defense table, where the favorable evidence is simply not there when counsel needs it. Each answers to a different rule of law, and the difference decides what a lawyer can do about it.

Classic suppression. The favorable evidence sits in the prosecution team’s own hands, and the government does not turn it over. This is Brady in its original form, the withheld video in Ulett. The item exists, the government controls it, and the defense never sees it. Volume is not the problem here. Concealment is.

Possession without review. The favorable material lives in police, crime-lab, forensic-examiner, or vendor files, and the trial prosecutor can say, truthfully, that they never saw it. However, Kyles v. Whitley, 514 U.S. 419 (1995), eliminates that as a justification. The individual prosecutor has “a duty to learn of any favorable evidence known to the others acting on the government’s behalf,” the police included, so the team’s knowledge is charged to the prosecutor whether or not anyone brought it to their attention. But the rule assumes a file a prosecutor can actually canvass, and the team whose knowledge is imputed has limits. Courts define it case by case, asking what role a person played in the investigation, not which agency employed them. When the favorable item sits in a terabyte extraction held by a forensic vendor or a cloud provider, two questions become hard to answer, viz., whether the vendor is part of the prosecution team at all, and how a prosecutor can be charged with knowing the contents of a file no one has read.

Disclosure without identification.The government produces everything and points to nothing. The defense receives millions of pages, or a full device extraction, with no index, no tags, no flag on the favorable file. Formally, the defense has the evidence. Functionally, it is lost in the volume, and a plea deadline can arrive before anyone opens the right folder. As Brian Chen documents in the student Note Big Data and Brady Disclosures, 99 N.Y.U. L. Rev. 1754 (2024), prosecutors now take in terabytes from police and often pass the whole case file through, knowing the defense is no better equipped to find the items of importance inside it.

The report instead of the data. The defense receives a curated output rather than the underlying record. What arrives is a PDF or spreadsheet export, a filtered extraction report, or a one-line notation that a lead came from a database search. What is not turned over is the forensic image, the native databases, the metadata, the tool logs, the search terms, and the candidates the software discarded. Brady does little, by itself, to force these to be provided. For instance, the Seventh Circuit distinguishes evidence favorable on its face from what it calls “latent exculpatory evidence,” material that “requires processing or supplementation to be recognized as exculpatory,” and it has held that Brady does not cover the latter. United States v. Gray, 648 F.3d 562 (7th Cir. 2011); Camm v. Faith, 937 F.3d 1096 (7th Cir. 2019). That rule does not answer the separate question whether existing underlying data already held by the prosecution team must be disclosed. The summary may be accurate, incomplete, or wrong, and the defense cannot tell which without the data it was based upon.

1. The last method is not suppression at all, because the evidence no longer exists. Cloud accounts auto-delete, surveillance footage overwrites on a loop, jail-call systems recycle their storage, and provider retention windows close. Brady governs evidence the government holds and withholds, but it says nothing about evidence that is gone. Two older cases fill that void, and they set the bar quite high. Under California v. Trombetta, 467 U.S. 479 (1984), the defendant must show the lost item had “an exculpatory value that was apparent before” it vanished and that no comparable evidence can be had by other reasonable means. Under Arizona v. Youngblood, 488 U.S. 51 (1988), if the material was merely “potentially useful,” the defendant must also prove the police acted in bad faith. Negligence, however costly, is not enough.

These are not one problem. Each method is governed by a different body of law and is subject to a different remedy. Classic suppression and possession without review are Brady questions, processed through the imputed-knowledge rule of Kyles and the materiality standard. Mass disclosure and the substituted report are litigated elsewhere, through Federal Rule of Criminal Procedure 16 and its state discovery-statute counterparts, motions to compel, negotiations over protocols for electronically stored information (“ESI”), demands for native files and metadata, and requests for funded forensic help. Disappearance is a preservation fight under Trombetta and Youngblood, and counsel should seek preservation before the data ages out. However, if the evidence is already lost, a pending request does not eliminate Youngblood’s bad-faith requirement. And where sheer volume defeats a defender who lacks the tools and the hours to review it, the same facts form the basis of a claim of ineffective assistance. The words “buried” and “hidden” describe how all five feel to the accused. But the doctrine that governs them draws sharper lines, and the discussion that follows keeps the categories separate.

What Brady Promised,
and What the Digital Age Broke

Brady announced a rule. Over the next half-century, the Supreme Court issued a series of decisions defining the rule’s scope and mechanics, and it is those features that digital evidence now puts under pressure.

The early additions expanded the duty. Giglio v. United States, 405 U.S. 150 (1972), held that the obligation covers impeachment evidence as well as direct proof of innocence, so an undisclosed promise of leniency to a key witness must be disclosed, and one prosecutor’s ignorance of a deal another prosecutor made is no defense. Thirteen years later, United States v. Bagley, 473 U.S. 667 (1985), supplied the trigger that governs every suppression claim. The Court held that evidence is “material” only where its suppression creates a reasonable probability of a different result, one “sufficient to undermine confidence in the outcome.” Kyles added that the prosecutor must gauge that probability across all the suppressed evidence together and is imputed to have knowledge of whatever the police and the rest of the team know.

Read together, these cases put a single judgment at the center of the Brady doctrine, and they put it before trial. The prosecutor, alone, must decide whether evidence they may never have examined helps the defense and predict whether it could change a verdict they have not yet seen returned. Two later decisions illustrate how far that structure favors the government, one operating before conviction, the other after.

In United States v. Ruiz, 536 U.S. 622 (2002), the Court held that the Constitution does not require prosecutors to disclose impeachment information before a guilty plea, reasoning that such evidence is “special in relation to the fairness of a trial, not … whether a plea is voluntary.” The logic is important because Brady, on this account, protects the fairness of trials, and a defendant who pleads guilty never has one. Since the overwhelming majority of convictions are obtained by plea, most defendants are convicted before the impeachment-disclosure duty ever attaches. And Ruiz decided only the impeachment question. It left unresolved whether due process requires disclosure of material exculpatory evidence before a guilty plea, and the federal circuits remain divided over whether, and under what circumstances, suppression of such evidence can invalidate a plea.

Turner v. United States, 582 U.S. 313 (2017), highlights the doctrine’s other accommodation, the one that operates after trial. The Turner Court affirmed convictions in the 1984 murder of Catherine Fuller even though the government conceded it had withheld favorable evidence, including a lead pointing to a different attacker. Weighed against the full trial record, the majority reasoned that the suppressed material was “too little, too weak, or too distant from the main evidentiary points” to be relevant.

The duty has not gone dormant. In Glossip v. Oklahoma, 604 U.S. 226 (2025), the Court ordered a new trial for Richard Glossip, who had spent more than two decades on Oklahoma’s death row, because prosecutors let their key witness testify falsely about his psychiatric treatment and did nothing to correct it. The witness, Justin Sneed, was the man who actually committed the murder and who testified against Glossip in exchange for his own life. That holding was based on Napue v. Illinois, 360 U.S. 264 (1959), the companion rule forbidding the government to let testimony it knows to be false stand uncorrected, rather than on Brady suppression, and it turned on a record in which Sneed supplied the only direct evidence of guilt. By the time the case reached the Supreme Court, Oklahoma’s own Attorney General had confessed error and asked that the conviction be set aside. The State has since elected to retry Glossip without seeking the death penalty. Congress has acted as well. The Due Process Protections Act of 2020 added Rule 5(f), which directs federal judges to enter an order at the start of every case confirming the prosecutor’s disclosure duty and the consequences of ignoring it. Whether those orders do more than restate Brady remains an open question.

Every element of this structure rests on a single assumption. The doctrine presumes a case file a human being can realistically read, held by a prosecutor who can know its contents and judge them before the first witness is sworn. In 1963, when a case file fit in a folder, that assumption was already a generous one. By Bagley’s time, when a complex prosecution could fill a wall of banker’s boxes, it was strained. Against a modern discovery record measured in terabytes, it is invalidated, because no prosecutor can know the contents of a file that large, and the materiality judgment Bagley demands becomes a guess about the contents of a haystack. Once the file grows past the point of human review, the act of disclosing it and the act of concealing something inside it begin to look the same.

The easy answer has always been the open file. Hand the defense everything, the reasoning goes, and fairness takes care of itself. But the Supreme Court has “never held that the Constitution demands an open file policy,” as Kyles itself put it, and in the digital age, the open file is often a dump that leaves both sides unable to find important evidence. The cost of that gap is not abstract. In 2025, the National Registry of Exonerations recorded 97 exonerations, and official misconduct occurred in at least 70 of them. For 2024, the Registry now counts 158 exonerations, 70% involving official misconduct. Across the more than 3,800 cases the Registry has recorded since 1989, the failure to disclose exculpatory evidence recurs as among the most common forms of misconduct. What these numbers do not yet capture is how much larger the file itself has grown.

The Data Explosion:
Terabytes, Not Folders

The paper case file has not shrunk. It has been replaced. The banker’s box of a generation ago held what one person could read in an afternoon. Its successor is measured in terabytes, and answering the question how much larger the case file has grown means starting with what now constitutes discovery.

Consider the range. A single criminal case can arrive with a full phone extraction, a forensic copy of everything on a device; a cloud return, the data a company such as Google or Apple hands over under warrant; hours of body-worn and dashboard camera video; doorbell and private surveillance footage; recorded jail calls; cell-site records that place a phone near a tower; geofence and tower-dump returns; license-plate-reader hits; social-media and messaging exports; device backups, app databases, deleted fragments, and system logs the ordinary user never sees; and the outputs of facial-recognition and automated filtering tools that ran before the file was ever assembled. Several of these carry their own hidden problems, discussed below. The point here is simpler and comes before all of them. The volume itself is now the problem.

Start with the phone, because nearly every case involves one. Using a class of software the industry calls mobile device forensic tools, built to copy a device whole, police reproduce a phone in full. Upturn, a technology-and-rights research group, documented in 2020 that more than 2,000 law enforcement agencies across all 50 states had bought these tools, that each of the 50 largest departments had them or ready access, and that agencies had run hundreds of thousands of extractions since 2015, in cases ranging from homicide to shoplifting. One extraction pulls a full copy, everything from emails and texts to photographs, location traces, app data, and material the user believed deleted. A single phone can hold more than the entire paper record of a comparable case a generation ago. Inside each device is a second case file, and it is usually larger than the first.

The volume is easiest to see where someone has measured it. In Colorado, where state law requires local law enforcement agencies and the Colorado State Patrol to provide body cameras to officers who interact with the public and generally requires those officers to wear and activate them during enforcement encounters, subject to statutory exceptions, a forensic investigator at the Denver District Attorney’s Office reported a 600% rise in audio and video evidence over five years and told reporters that even a minor misdemeanor can now generate a terabyte. The Jefferson County District Attorney’s Office handled more than 67,000 videos in 2025, roughly 41,000 hours of footage, an 83% jump in three years. One vehicular-homicide case there produced 362 photographs and as much as 90 hours of body and dashboard video, against 79 photographs and no video in 2017.

Merely storing all this runs to millions; the council serving the state’s district attorneys budgets some $3.5 million a year for the shared data system and its supporting infrastructure, and it expects the figure to climb. The disclosure deadlines are short. Colorado prosecutors must disclose specified evidence as soon as practicable, but no later than 21 days after the defendant’s first appearance at or following the filing of charges, so by the Jefferson County office’s own account, a felony attorney carrying about 100 cases can be responsible for reviewing roughly 400 hours of video to be sure nothing is missed. Those are the prosecution’s numbers, gathered by prosecutors who have staff, storage, and software. The same volume is typically turned over to the defense, which often has far fewer resources.

Every rule in the Brady line of cases is based on the premise that the evidence in a criminal case can be meaningfully reviewed by a human being. At terabyte scale, that premise is false. Somewhere within a terabyte, there may be exculpatory evidence such as a timestamp, a message, or a few frames of video that put the defendant somewhere else. But the haystack is now growing faster than anyone, on either side, can search it. The favorable evidence has not disappeared. It has become much harder to find.

Location data shows how quickly the governing rules change. On June 29, 2026, in Chatrie v. United States, 225 L. Ed. 2d 793 (2026), the Supreme Court held that police conduct a Fourth Amendment search when they obtain a person’s Google Location History, the company’s running, timestamped record of where a phone has been, through a geofence warrant, a demand for every device that crossed a drawn perimeter near the time of a crime. That holding governs how such data may be acquired. It says nothing about the discovery questions that come after, and those are the ones that decide cases: what the government collected, what it preserved, what it reduced to a summary, what it turned over, and whether the defense received it in a form it could use. The acquisition rule and the disclosure duty now sit side by side, and a defendant can win the first while losing the second. The law governing both continues to change. Chatrie noted that in July 2025, years after the warrant in that case, Google changed Location History to store the data on users’ own devices rather than on company servers, and that Google represents it can no longer answer geofence warrants seeking it. For the next defendant, the record that might have placed him elsewhere may not exist to disclose at all, and he may never learn what it would have shown.

None of this, by itself, is a constitutional violation. A large production is not yet a hidden one. The harder question is what happens when the government answers all this volume by handing the defense the whole haystack and claiming the duty discharged. That is where the Brady doctrine must do its work, and it is where the doctrine fails.

The Data Dump:
When Handing Over
Everything Discloses Nothing

One line of authority rejects that result. Where the government knows of specific favorable material, courts have declined to treat volume as an excuse for failing to identify it, and the haystack comparison used throughout this cover story comes from the courts themselves. More than two decades ago, in United States v. Hsia, 24 F. Supp. 2d 14 (D.D.C. 1998), a federal judge held that the government could not satisfy Brady by handing a defendant 600,000 documents and expecting her to “find the exculpatory information in the haystack.” To the extent it knew of Brady material, it had to identify that material. Later courts used the rule with digital evidence, applying it to a production of roughly 3.5 million pages in United States v. Saffarinia, 424 F. Supp. 3d 46 (D.D.C. 2020), and to terabytes of discovery in United States v. Sheppard, No. 21-cr-203-JDB, 2022 WL 17978837 (D.D.C. Dec. 28, 2022). A federal court in New York used the same comparison, warning that the government may not hide favorable proof as “an exculpatory needle in a haystack of discovery materials.”

Two features of that line are of greater significance than the holdings themselves. The first is that it runs mainly through the District Courts. Because they are district decisions, they bind no one, and a later court remains free to distinguish them on their facts. The second is the qualifier that appears in every one of them. The duty to identify attaches only to material the government already knows is favorable. That qualifier reduces the protection to almost nothing. The duty applies to the prosecutor who has spotted the exculpatory record and said nothing. It does not apply to the far larger category in which no one has identified the item because no one has read the file, and the format defeats any search that might have found it. Even the judges imposing the duty have been uneasy with it. The Sheppard Court said it would not make the government do the defense’s work or manage discovery item by item.

The qualifier also conflicts with the doctrine that is supposed to protect defendants in the first place. Brady allows no excuse for good faith or impracticability, and it imputes to the prosecutor knowledge of everything gathered by the investigative team, on the theory that favorable evidence in the file should be recognized even if overlooked. That imputation was meant to remove ignorance as a defense. But scale restores it in a different form. The knowledge that triggers the identification duty, the prosecutor’s actual awareness that a particular record helps the defense, is precisely what volume eliminates. A file too large to read is also a file whose exculpatory contents no one has knowingly passed over. The terabyte does more than bury the needle. It gives everyone who failed to find it a truthful explanation for never having seen it.

The analytical core is easy to state. When the favorable file is one item among millions, produced in a form the defense cannot search, under a deadline it cannot meet, the act of disclosing it and the act of hiding it become the same act. Intent is where care is required. Sometimes burial is deliberate, a chosen tactic. More often, it is a byproduct of scale, thin indexing, limited funding, opaque software, and the plea deadline that arrives before review can finish. From the defendant’s perspective, the two are indistinguishable, and that is the difficulty, because the doctrine asks whether evidence was suppressed while a data dump can suppress without anyone having decided to suppress anything. In an April 2025 federal filing, the defense contended that the government had produced more than 60 million pages, unsearchable and unindexed, and could not say whether anything exculpatory was contained inside. The filing challenged a magistrate judge’s ruling that credited the government’s contrary representation that the files were load-ready and searchable and denied the defense’s request. The two defendants received full and unconditional presidential pardons in January 2026, so the case is no longer pending. The broader structure the dispute exposed (mass production, contested searchability, and uncertainty about what no one has reviewed) is increasingly familiar.

None of this points to a rule that every data dump is a Brady violation. The law is too unsettled, and too fact-dependent, for that. It points instead to a gap. Courts agree that a haystack can hide a needle, but they have not settled on any duty to make the needle findable. The remedies that would close that gap are not complicated, i.e., an index, production in native and searchable form, and identification of the favorable material the government does know it holds. The modern Brady fight is over the gap between producing evidence and making it findable.

The dump concerns what the defense receives and cannot find. A subtler problem concerns what the defense never receives at all, the native data behind the report, the file behind the summary.

The Report Is Not the Data: What Never Enters the Production at All

The last section described a file too large to search. This one describes the opposite failure, a file too small. Very often the defense does not receive the phone’s data at all. It receives a document about the phone, a report that substitutes for the extraction and narrows it without saying so. The distinction is not a technicality. It decides what a defendant can test, because you cannot cross-examine a summary, and you cannot challenge an interpretation of data you were never handed.

Consider what a phone extraction actually is. Pulling data from a device is a two-step process, extraction and then analysis, and the tools do both. The extraction can be shallow or deep. A logical extraction takes the active data the operating system is willing to hand over. A full file-system extraction, the gold standard, reaches further, into system databases, application caches, and, as one forensic guide puts it, “artifacts that are invisible to the device user.” That is where deleted texts, removed photos, cleared browsing history, and location records are contained. The deepest, a physical extraction, copies the drive bit for bit, including the unallocated space the user believes is empty. Not every phone yields a full image. It depends on the device and its software, so what exists to disclose varies from case to case. Any of these can run to thousands of report pages. But the extraction and the report are not the same. In Cellebrite’s tools, the full extraction is one file, and the report the defense usually receives is another, which a forensic examiner describes as “a curated subset of the full dataset,” assembled by a person who chooses which categories and which span of time to include.

That choosing is where the problem lies. The report is a human’s selection from a machine’s copy, and the selection can be entirely innocent, leaving out what no one thinks is of any relevance, or it can leave out what turns out to be of great importance. For the defense, the two are once again indistinguishable. The more cynical version of the selection is a cherry-pick, and a forensic firm warns that a report built by choosing data points can present an incomplete or misleading picture. The forms this takes are concrete. A photograph can sit in the extraction marked as never opened. An access timestamp can be dropped. A location field can rest in an app database the report never mentioned. A line in a PDF reporting that a message was sent is not the database showing when it was sent, whether it was later edited, and whether something near it was deleted. The report can be accurate and still be incomplete, and incompleteness is enough to change a case.

To test any of it, the defense needs the underlying material rather than the summary of it. That means the native files in their original format; the metadata; the hash value used to verify that the acquired data has not changed after collection; the tool version and parsing logs; and the examiner’s notes and search terms. Without those, counsel is arguing against a conclusion it cannot verify, which is not much of an argument. The reliability of the government’s reading is a separate problem because timestamps shift with time zones and location estimates are less precise than they appear. But that challenge cannot begin without the data underlying the report. The disclosure problem and the reliability problem are inseparable. The government’s summary may be wrong, and the defense cannot prove it wrong without the file the summary was based upon.

The gap is not unique to phones. Wherever a tool produces a summary, it is not the same thing as the underlying data. When software generates an investigative lead, a facial-recognition candidate or a filtered analytic result, what reaches the defense is a conclusion, while the probe image, the ranked list of other candidates, the confidence scores, the discarded near-matches, and the audit log are not turned over. Disclosing the lead is not disclosing the process that produced it. The next section discusses how software has become the gatekeeper of what the defense ever sees. The point here is only that the difference between a report and its underlying data is a general feature of digital evidence, not a quirk of one device.

The most difficult version of the problem is the evidence that never enters the production because it was never kept. Some favorable material is not summarized away. It is gone. Surveillance video overwrites on a retention cycle; body-camera and jail-call systems purge on schedules; cloud data can disappear under user settings, provider policies, or changes in service; providers hold logs for a limited duration. And the sources themselves change, as they did when Google shifted Location History to users’ own devices in July 2025, so that a record a prior case could have obtained may no longer exist. Here the governing doctrine is different and far weaker.

This is no longer a disclosure question. It is a preservation question, and preservation law protects defendants much less. Under Trombetta, the government must preserve evidence whose exculpatory value was apparent before it was lost and for which no comparable evidence is reasonably available. Under Youngblood, when the evidence was only potentially useful, its loss violates due process only if the defendant can show the government acted in bad faith. That bad-faith requirement is hard to satisfy anywhere and close to impossible in the digital setting, where deletion is what the system does on its own, on a timer, and looks like routine rather than intent. A defendant cannot challenge, and a reviewing court cannot examine, what the government never disclosed or never saved.

The party that holds the data controls what “disclosed” comes to mean, whether by burying the favorable item in a mass, reducing it to a summary, or letting it expire before anyone asks. A further change is under way. More and more, the sorting itself is done by software, and the decisions that determine what the defense ever sees are made by a system no one is required to open and explain.

When Software Becomes
the Gatekeeper

Software enters a criminal case at two different moments, and at each, it makes a decision the defense may never see. It sorts the discovery, choosing what to put in front of a human and what to leave in a low-priority pile. And it generates the lead, pointing investigators at a suspect before there is a case at all. Both are versions of the same problem. The party who controls the process controls what counts as disclosed, and a process run by machine is the easiest kind to keep closed. When a program decides what is relevant, that decision shapes everything that follows, and it is made inside a process the defense cannot see.

Take the sorting first. Prosecutors handling the terabyte increasingly rely on technology-assisted review (“TAR”) software trained on a sample of documents a person has already coded so that it can classify the rest as relevant or not or rank them by likely importance. The software is not the problem by itself. In AI Suppression: E-Discovery Software and Brady, 27 Colum. Sci. & Tech. L. Rev. 307 (2026), Jason Hartline, Rebecca Wexler, Liren Shan, and Alec Sun ran simulations on synthetic datasets and found that the same TAR system can be set up to bury Brady material or to reveal it, and that little in current doctrine, agency guidance, or scholarship dictates which. The danger is that its settings stay invisible. What it was trained to find, what threshold it applied, and whether any person ever read the low-relevance pile where a favorable record might sit are not ordinarily disclosed. The authors propose fixes, among them running the tool separately to search for Brady evidence and favoring designs that can be reconfigured for that purpose. Those are proposals. Today, “we ran it through review” can mean a machine did the sorting and no one opened the discard pile.

That gap is a Brady problem in a specific way. The qualifier that has limited the identification duty in the data-dump cases is that the government must flag what it knows to be favorable. If a classifier drops the exculpatory file into the pile no one reads, the prosecutor may never know it is there. Whether Brady liability nonetheless attaches to favorable evidence the prosecution team possesses but no person knows about remains unresolved. The authors of the TAR study argue that it should. Either way, the sorting decision has become the disclosure decision, made by a tool configured within the prosecution’s review process but visible to no one outside it. The same volume that buries an evidentiary needle by accident can now bury it by algorithm.

The second moment comes earlier in the investigation, and facial recognition is its clearest form. A detective feeds the system a probe image, often a still pulled from surveillance video, and the software compares it against databases of mugshots and license photographs to return a ranked set of possible matches. Police routinely describe the result as a tip or an investigative aid rather than evidence and frequently do not disclose that it was used at all. But the tip is a conclusion based on a chain the defense cannot see without asking, starting with the quality of the probe image and running through the database searched, the ranked candidate list, the confidence scores, and the other faces the system returned that someone set aside. Disclosing that a suspect was identified is not disclosing how.

The Maryland case of Johnson v. State, No. 413, Sept. Term 2024, 2025 WL 2237582 (Md. App. Ct. Aug. 6, 2025), shows what is at stake. A store was robbed, an employee attacked, and a detective ran a still from the store’s cameras through facial recognition, which produced Craig Johnson as a suspect. No witness identified him before trial, and no physical evidence tied him to the crime. By the appellate court’s account, the software was the only thing that connected him to the case. The State disclosed its use of the technology eight days before trial, admitted it had violated its discovery obligations, and said it did not even know which facial-recognition program the police had used. Held nearly a year awaiting trial, Johnson declined a postponement, was convicted, and appealed. In 2025, the Appellate Court of Maryland reversed, in an unreported decision, finding that the late and empty disclosure left the defense “no way to investigate or challenge the identification” the whole case rested on.

Two features make these leads harder to contest than ordinary evidence. The first is the framing itself. Calling the search a mere tip treats the most consequential step in an investigation as beneath disclosure, even when, as in Johnson, everything that followed stemmed from it. The second is trade secrecy. The companies that build these systems claim their methods are proprietary, and courts have sometimes let that intellectual-property interest block the defense from the information needed to test a result.

Neither barrier, though, has held everywhere, and New Jersey shows what it looks like when a court refuses both. In State v. Arteaga, 476 N.J. Super. 36 (App. Div. 2023), detectives investigating a robbery got nothing from their own facial-recognition system, sent the surveillance footage to the New York Police Department, and used the possible match its software returned to build photo arrays from which two witnesses identified the defendant. When the defense asked how the search had worked, the trial court denied the request. In 2023, the Appellate Division reversed, holding that the defendant had shown a particularized need for 13 items about the search, the software’s name and manufacturer, its algorithms and error rates, the probe image and its edits, the candidate lists it returned, under the state’s discovery rule and Brady, reasoning that without the material no one, the Court included, could even say whether the evidence was exculpatory. In 2026, the state’s Supreme Court, in State v. Miles, 264 N.J. 209 (2026), largely ratified the approach. It declined to turn Arteaga’s 13 items into a rigid checklist and held that compelling the vendor’s source code was premature on the record before it, but it confirmed that the basic information, the identity and performance of the tools, their error rates, the probe photograph, and the matches returned, will in most cases be the minimum a fair trial requires.

That is one state’s answer, based on that state’s discovery rule as much as on the U.S. Constitution, and no other court of last resort has yet followed. Elsewhere, there is only a proposal. Deborah Won’s student Note The Missing Algorithm: Safeguarding Brady Against the Rise of Trade Secrecy in Policing, 120 Mich. L. Rev. 157 (2021), would meet a successful trade-secret claim with a missing-algorithm instruction, permitting jurors to draw a limited adverse inference when secrecy denies the defense access, though that remains a proposal and not the law. The scale of what goes untested is not trivial. By one account facial recognition was used as an investigative tool in more than 8,000 cases in a single year, and many of those defendants never learned it had been used in their case.

Whether the software sorts the file or generates the suspect, the result is the same. The party who controls the process controls what “disclosed” means, and an opaque process is the simplest to control. What the defense needs in order to test either one is specific and identifiable, the inputs and settings, the candidate lists and scores, the audit logs and the discarded results. But all of it still turns on a prior question. Someone has to decide what in all this data counts as material.

Who Decides What’s Material, and Why That’s the Trap

The entire system rests on one person’s pretrial judgment. Materiality does more than tell an appellate court when to reverse a conviction. It tells the prosecutor what to hand over in the first place. And the Supreme Court has been explicit about who holds that power. In Kyles, the Court reasoned that because the prosecution “alone can know what is undisclosed,” it must be the one to weigh the likely effect of the favorable evidence and disclose once the point of reasonable probability is reached. Kyles left the prosecutor real discretion in that judgment and said the prudence of the careful prosecutor should not be discouraged. But it paired the discretion with a burden. The design places a single official between the file and the defense, forecasting before trial what is both favorable and material.

That forecast is difficult in a case whose file fits in a single folder. At digital scale, it stops making sense. Kyles asks a prosecutor, convinced of the defendant’s guilt, to predict the cumulative effect of evidence on a trial that has not yet happened. Christopher Deal, in the student Note Brady Materiality Before Trial: The Scope of the Duty to Disclose and the Right to a Trial by Jury, 82 N.Y.U. L. Rev. 1780 (2007), calls the use of materiality as the measure of pretrial disclosure both impractical and unconstitutional: impractical because it requires a prosecutor sure of guilt to decide what combination of items, if turned over, would create a reasonable probability of acquittal at a trial that has yet to begin. The criticism is not a partisan one. At the argument in Kyles, Justice Scalia called the cumulative pretrial test “crazy,” observing that one could never say a single item was safely withheld, since everything depends on the aggregate. Now add the terabyte. The forecast assumes a file the forecaster has read.

The qualifier carries more weight here, the rule that the government must flag only the favorable material it actually knows about. Kyles charges the prosecutor with favorable evidence known to others acting on the government’s behalf in the case, and some courts extend that constructive knowledge to everything gathered in the investigation. Brady allows no excuse for good faith or impracticability. Those rules were built to protect defendants, by refusing to let the government plead ignorance. A file too large to read inverts their effect. The gatekeeper is charged in law with knowing the contents of a file no person has actually read. The imputation the doctrine depends on is what the volume of data defeats.

The result is a gap in which no one is accountable. No one fully reviews the haystack. The gatekeeper is presumed to know it. The defense is told it received everything. And if a favorable item is discovered later, the question is no longer whether the government hid it. The question is whether it was material, judged in hindsight against a full trial record, and often judged after a guilty plea has already foreclosed the trial, after retention windows have closed, after the algorithmic logs are gone. For most defendants, who plead rather than go to trial, that trial record never exists, and the Supreme Court in Ruiz has said the government need not hand over impeachment evidence before a plea in any event. A recent study of how these claims are actually decided finds that courts frequently reject them on the ground that the withheld evidence was not significant enough. Brandon L. Garrett & Adam M. Gershowitz, The Brady Materiality Standard, 78 Stan. L. Rev. 303 (2026). The standard meant to shield the defendant becomes the reason the claim fails.

Even a proven violation rarely costs the government much. The ordinary remedy is a new trial, which requires the evidence to be disclosed and then to be proven material, though such violations are, as Justice Ginsburg observed in dissent in Connick v. Thompson, 563 U.S. 51 (2011), “not easily detected.” Damages are rarely available. In Connick, the Orleans Parish District Attorney’s Office conceded that its prosecutors had committed a Brady violation that helped put John Thompson in prison for 18 years, 14 of them on death row, until a defense investigator’s chance discovery about a month before his scheduled execution produced the hidden report. The Supreme Court held, five to four, that the office could not be sued for failing to train its lawyers on Brady absent a pattern of similar violations. No pattern, no liability. As Justice Ginsburg’s dissent recounted the trial record, the District Attorney had never disciplined or fired a single prosecutor for violating Brady. And the concealment of favorable evidence sits within the category the exoneration record calls official misconduct, which was a factor in most of the wrongful convictions overturned in 2024. The incentive the structure creates is not subtle. Disclose everything, identify nothing, and argue later that the defense had access all along.

This is why the problem does not reduce to bad prosecutors. From the defendant’s perspective, four very different things look exactly alike: (1) deliberate concealment, (2) honest oversight, (3) a classifier that dropped the file in the pile no one read, and (4) a bureaucracy that never opened it. The two questions the Brady line of cases asks, whether evidence was suppressed and whether it was material, cannot tell those four apart. This is the same indistinguishability, now built into the remedy itself. And the structure rewards the posture that produces all of them, because the gatekeeper who identifies nothing is the gatekeeper least likely ever to be found to have known.

Thus, the materiality gatekeeper, built as the defendant’s protection, has become the mechanism that keeps the file closed. All of it has assumed something not yet examined, a defense with the time, money, and expertise to find the needle and prove it mattered. That assumption is the weakest of all.

The Asymmetry of Arms: Indigent Defense in the Terabyte Era

Every part of the Brady framework assumes an adversary on the other side. It assumes a defense that can receive the file, read it, understand it, and put its contents to use. That assumption supports the entire design, and it is rarely tested. The question is whether it holds when the file is a terabyte and the lawyer holding it is a public defender with more than a hundred other cases. The traditional phrase for what is supposed to be in place is equality of arms. Digital volume is where that description breaks down.

Start with the promise. Gideon v. Wainwright, 372 U.S. 335 (1963), guarantees an indigent defendant a lawyer, and the point of the lawyer is to test the government’s case. But two doctrines set how much testing the Constitution actually requires, and both were built for a world of paper. Under Strickland v. Washington, 466 U.S. 668 (1984), a defendant who claims ineffective assistance of counsel must show two things, that counsel’s performance fell below an objective standard of reasonableness and that there is a reasonable probability that, but for the deficiency, the result would have been different, a probability the Court defined as one “sufficient to undermine confidence in the outcome.” Review is deliberately forgiving. Courts presume counsel’s choices were sound, decline to second-guess tactics in hindsight, and treat strategic decisions as almost unchallengeable. And the prejudice standard should sound familiar, because it is the same reasonable-probability, confidence-undermining test the Supreme Court would adopt a year later to define Brady materiality in Bagley. The same retrospective standard therefore applies to a defendant harmed by buried evidence whether the claim is brought against the government under Brady or against counsel as ineffective assistance.

That deference is appropriate where counsel made a strategic choice. It is not appropriate where the review was impossible. A lawyer who has not watched every hour of body-camera video has not adopted a strategy; the lawyer has run out of hours. But Strickland asks whether counsel acted unreasonably and whether the missed item would probably have altered the verdict, and both questions presume that someone found the item and examined it. When the favorable record languishes unread inside a voluminous mass, no one can say what it would have done, and so the prejudice prong cannot be satisfied, for the very reason the Brady claim failed in the last section. You cannot prove the weight of what no one has seen. There is a narrow exception, from United States v. Cronic, 466 U.S. 648 (1984), that presumes prejudice when counsel fails to subject the prosecution’s case to “meaningful adversarial testing.” But courts grant it rarely, and they have been slow to treat a lawyer buried by volume as a lawyer constructively absent.

In any event, testing digital evidence takes more than a lawyer. It requires a forensic examiner who can open the full extraction, a specialist who can probe a facial-recognition result, and someone who can read the metadata and the logs the report left out. Ake v. Oklahoma, 470 U.S. 68 (1985), held that due process promises an indigent defendant “the basic tools of an adequate defense,” including expert help, on the reasoning that mere access to the courthouse doors is not a fair trial if the defendant lacks the raw materials to build a case. But Ake involved a psychiatrist in a capital trial, and whether its guarantee covers digital-forensics and software experts is contested and applied unevenly. Many defendants never get the examiner who might have found the needle. The right to test the evidence exists on paper but is rationed in practice. And where the tool is a vendor’s trade secret, as the facial-recognition cases show, even a funded expert can be denied access.

Most of this, though, never reaches a courtroom at all. The overwhelming majority of convictions are guilty pleas, and pleas come quickly, usually long before anyone could review a terabyte. The Supreme Court held in Ruiz that the government need not disclose even impeachment evidence before a plea. So the defendant is asked to decide whether to give up a trial while the exculpatory needle, if there is one, is still lost in an unread file, under a disclosure deadline often measured in a few weeks. The Colorado numbers from the data-explosion section illustrate the problem from the defense side. A single vehicular-homicide case there contained 362 photographs and as much as 90 hours of video, where a comparable case in 2017 had involved 79 photographs and no video at all. The lawyer assigned that case is likely handling dozens of others. In practice, the workload does not permit the review. What ends the case instead is the plea, entered under a pressure of its own, a trial penalty that makes going to trial the riskier choice, so the defendant gives up the very proceeding at which the buried evidence might have come to light.

Faced with all this, defenders have reached for the same kind of tool the government uses. The Miami-Dade Public Defender’s Office began using an artificial-intelligence assistant for research and case preparation. Startups now sell public-defense offices across the country software that transcribes body-camera footage, flags moments like a Miranda warning or a sobriety test, and finds the needle faster, saving hours of review. This helps, and it is being adopted out of necessity. But it is no cure for the asymmetry, and it may widen it. The tools cost money that indigent-defense systems mostly do not have, so access follows funding rather than need. They present the same opacity and reliability problems seen on the government’s side, since an automated flag is only as good as its configuration. And feeding a client’s confidential file into an outside system raises privilege questions the defense bar is still sorting out. The defense is now searching the same haystack with a less capable version of the same imperfect software, on a fraction of the budget.

Set the pieces side by side. The government generates the volume, chooses the format, and owns the tools. The defense inherits the output, under a competence standard that assumes a review it cannot perform, an expert-access right that is granted sparingly, and a plea deadline that expires before the file can be meaningfully reviewed. Equality of arms was always more aspiration than reality. Digital volume makes it unsustainable, because the party who builds the haystack is the party with the staff, the software, and the budget to search it, while the party told it received everything is given no realistic way to use any of it. The imbalance is not a matter of bad faith on anyone’s part. It is built into who holds the data and who does not.

That is the machinery in the abstract. What it does to particular people is plainer in the cases, where the needle was discovered late, or by accident, or never, and a conviction rested on data no one had truly reviewed. Those are discussed next, and they are not rare events. They are what the ordinary operation of the system now produces.

Case Studies:
The Needle That Was Found,
and the Ones That Weren’t

A doctrine’s practical effect shows in the cases it produces and in what it has done to particular people. The cases that follow track the same methods, the classic suppression, the disclosure that identifies nothing, the lead whose supporting process was never turned over, and the evidence that simply disappears. What they share is easy to miss. In almost none did the favorable evidence come to light because a rule required the government to identify it in time for the defense to use it. It appeared late, or by luck, or after the harm was done, and behind each of these cases is an unknown number in which it never appeared at all.

Begin before the digital problem, to see it by contrast. The kind of suppression the older law was built to catch looked like Derrick Ulett’s case, a withheld building-surveillance video, a prosecutor assuring the jury there was no such video, and a conviction reversed once the truth came out. What made that catchable was that the missing piece was discrete and its absence was noticeable. A defense lawyer looking at a thin file could readily see the hole. The digital cases remove that ability. The favorable record is no longer one item withheld from a slim folder. It is one file among millions, or a tidy summary offered in place of the underlying data, or an investigative lead that left no paper trail. The hole stops being something a person can notice.

Consider the dump from the defense’s perspective. In a 2025 federal prosecution in New York, the government produced more than 1.3 million documents. Buried somewhere in there, according to the defense’s own account in a court filing, was the sworn testimony of a key witness, taken privately in a related matter, that the defense regarded as plainly favorable and had never been directed to. The government’s response was not that it had disclosed the testimony and moved on. The government maintained that the testimony was not exculpatory at all. That exchange shows two things. The defense found the item only by reading deep enough into the pile, by labor and chance rather than by any identification the government supplied. And once the item was in hand, the dispute was no longer about whether it had been turned over. It was about whether it counted as favorable in the first place, the question of who decides materiality, argued while both sides had the document in front of them. At the extreme edge of the same problem is the case described in the data-dump section, where the defense alleged a production of more than 60 million pages that were unsearchable and unindexed. A magistrate judge credited the government’s contrary representation that the files were load-ready and searchable and denied the defense’s request. The two defendants received presidential pardons in January 2026, so the dispute no longer supplies a pending test. It still involves the same structure of a mass production, a contested search function, and no agreement about what the defense could realistically find.

The next case involves an algorithmic lead. Craig Johnson was convicted of a robbery on the strength of a single investigative step, a still pulled from a store’s security camera and run through facial-recognition software, and the Maryland appellate court that reversed his conviction said so plainly. Everything that followed depended on that step. The software’s result made Johnson a suspect. The suspicion produced the surveillance of his home. The surveillance produced the search. The search produced the arrest. The victim had not identified him before trial, and no physical evidence connected him directly to the robbery. But at trial she identified him with certainty, and police had seized sneakers they believed resembled the assailant’s. Eight days before trial, the State revealed that it had used facial recognition, could not say which program the police had run, and produced only a three-page email about the search. The email showed Johnson but said nothing about any other results. The State said he was the only suspect generated and that it had no other information, and the appellate court declined to speculate about what else might exist. The court itself pointed to what can follow when those other results are never disclosed. A few years earlier, a facial-recognition system had wrongly identified another Maryland man, who was arrested for an assault he had not committed. And this reversal occurred in a case where the State itself had conceded the discovery violation. Had the use of the software never been disclosed, as often happens when agencies treat a facial-recognition result as a lead rather than evidence, there might have been no known error to raise and no appeal on that ground. What was concealed here was not a document. It was the account of how this defendant became a suspect at all.

Sometimes the favorable evidence is not buried at all. It is gone. In a federal case arising from the events at the Capitol, the defense sought text messages from Secret Service agents that had been wiped from their phones, and the court applied the settled rule that destroyed evidence violates due process only if the defendant can show the government acted in bad faith, which the court did not find. Consider the defendant’s position. He cannot say whether the deleted messages would have helped him, because he never saw them. He cannot show bad faith, because deletion on an automatic schedule looks exactly like routine. The preservation doctrine asks him to prove both the value and the malice of something that no longer exists, and the digital setting makes each proof close to impossible. This is the disappearance category, applied to a defendant who can prove neither element.

One further case is usually offered as reassuring. The evidence that exonerated John Thompson, and stopped his execution, was found by a defense investigator simply by chance, about a month before his scheduled execution. That is the system working. Now compare it with the rest. In every one of these cases, the favorable evidence came to light late, or by accident, or had already vanished, and not one was discovered because any rule guaranteed the defense would find it. The known cases are the ones where the needle happened to turn up. The cases where it stayed in the haystack generate no reversal, no exoneration, no written opinion, because no one ever learns the needle was there to be found. The counted exonerations are real, but they are only the cases in which someone found the evidence. The number in which no one did cannot be counted, because those cases leave no record.

The cases raise a further doubt. Even when the needle is found, a second question remains: whether what the government produced is accurate. Accuracy turns on whether the summary is faithful to the data, whether the timestamp was read correctly, and whether the software’s answer is right. Volume and format decide whether the defense ever sees the evidence. Reliability decides whether the evidence means what the government says it means.

The Veneer of Objectivity: Whether the Evidence Is Right

Everything to this point has turned on a single question, i.e., whether the defense ever sees the favorable evidence. There is a second question that is independent of the first and just as able to decide a case. Is the evidence right? A defendant can win complete disclosure, receive every file, and still be convicted on a reading of the data that is mistaken, because digital evidence appears objective. It comes as numbers, timestamps, and matches, and that appearance persuades before anyone asks how the result was produced. The reliability question is not a technicality reserved for experts. It is often the whole case.

The raw copy is usually the trustworthy part. The danger is in the interpretation. Forensic practitioners say as much directly, that the interpretation of extracted data is more open to challenge than the extraction itself. Consider a timestamp, which looks like a hard fact. What time a device records depends on its time-zone setting, on whether it synced to the network, on whether someone had reset the clock, so the same event can carry different times. And an examiner who treats a stamp as literal can place a person somewhere he never was. The only way to catch that error is to examine the device’s own settings and the associated records, which is the material the report omits. Location is easier to overstate still. Data drawn from cell-tower connections describes a rough area, not a point, but it is simple to present and simple for a jury to hear as a precise pin dropped onto a map. The extraction can be perfect, but the story told about it false.

The outputs that look most authoritative are often the least understood. A facial-recognition system returns a ranked list with confidence scores, which imply rigor, but the Maryland court that reversed Craig Johnson’s conviction said there was ample reason to doubt the reliability of such evidence. The record gathered in that case bears it out. Identifications made from low-quality surveillance footage can approach chance-level accuracy. Error rates increase with poor images, with pose and age, and rise again across race, reaching, in some studies of human face-matching, as high as 60%. One widely used system, handed a probe image, offered the basketball legend Michael Jordan and a cartoon among its candidate matches. The Maryland misidentification the Johnson Court referenced, a man arrested for an assault he did not commit on the strength of a facial-recognition hit, is what that danger looks like in practice. A match is a probability presented as an identification.

Even the methods courts treat as most reliable can produce inconsistent results. Probabilistic genotyping software, used to untangle complex DNA mixtures, is treated in many courtrooms as nearly conclusive. But in one case, the same DNA evidence, run through the same program twice in two different locations, yielded different results, some helpful to the defense. The appellate court acknowledged a strong showing of arguably inconsistent results. The figure a jury hears as the weight of the evidence is partly an artifact of the tool and how it was set. The version, the settings, the validation studies, and the error rates are what would let anyone judge the number. They underlie the report and rarely accompany it. That is the disclosure problem again, in a different setting. You cannot test a reading you were never given the materials to test.

Used carefully and disclosed in full, these tools can be accurate, and sometimes they exonerate. The problem is that careful validation and full disclosure are the two steps most often omitted. The concern is not new. Courts have a long history of admitting forensic techniques that appeared scientific and were later determined to have little scientific foundation. Digital tools carry that appearance more strongly than any of their predecessors, because a computer produced the result. The system’s answer to unreliable evidence is adversarial testing, a rigorous cross-examination, and an expert able to say what the number really means. That is the capacity the previous section showed the defense most often lacks. So an unreliable reading is put before a defense that cannot test it, and the jury receives it unchallenged.

Together the two issues complete the problem. The defense may never see the evidence. If it sees the evidence, it may lack the means to test whether the evidence is true. Either failure alone can convict, and digital cases routinely produce both. No rule yet written requires the government to make the favorable evidence findable and testable, with a consequence attached when it falls short. That leaves the burden where it has been all along, on the defense lawyer in the room, who must work with the rules exactly as they are.

Practical Defense Implications: What to Do While the Law Catches Up

The lawyer in the room cannot wait for the rule that has not been written. The cases arrive now, in the format the government chooses, on the schedule the government sets, and the client sitting across the table needs an answer this month, not this decade. So the closing question is a practical one. Given everything the digital transformation has broken, what can a defense lawyer actually do about it today? The steps that follow share a single principle. Because the system judges in hindsight and forgets whatever no one examined, the answer is to push the questions forward, before the plea, before the deadline, before the data disappears, and to make the asking itself part of the record.

Start with the dump. Do not accept a drive full of files that no one can search. In 2012, a joint working group of the Department of Justice and the Administrative Office of the U.S. Courts issued recommendations for producing electronic discovery in federal criminal cases, voluntary best practices that invite the parties to meet and confer about format at the outset. Use that process at the first opportunity, and raise format in writing, native, searchable, with a load file and a table of contents. Where the government knows of specific favorable material, demand that it be identified and cite the line of cases holding that a prosecutor may not bury known Brady material in an unmanageable mass and call it disclosed. The government has real interests of its own here, in shielding genuinely sensitive material and in not rebuilding a finished production from scratch, and a protective order usually answers the first. The request is for the favorable evidence in a form that can be used, nothing more. Two things are gained at once. The defense may get usable discovery, and it builds a record that the government was asked so that a later claim of having turned over everything cannot pass unexamined.

When the government produces a report, ask what it is a report of. For a phone, that means demanding the full forensic extraction rather than the tidy reader file built from it, along with the native files, the metadata, the hash values that prove the copy is faithful, the tool and its version, the parsing logs, and the examiner’s notes and search terms. These are the materials an expert needs to test whether the summary left something out or read something wrong. If the government refuses, memorialize the refusal, because it turns a later reliability challenge from a guess into a documented denial of the very thing required to check the government’s work.

Where software did the sorting or produced the lead, ask how. For technology-assisted review, ask what the tool was trained to find, what relevance threshold it applied, and whether any human ever read the low-relevance pile. For facial recognition, demand the probe image, the database searched, the ranked list of candidates, the confidence scores, the faces the system returned and someone discarded, the audit log, the governing policy, and the known error rates. Where disclosure of facial-recognition use is required, invoke the requirement and note that it now takes two forms. Washington, Montana, and Maryland require notice by statute (Wash. Rev. Code § 43.386.070; Mont. Code Ann. § 44-15-107; Md. Code Ann., Crim. Proc. § 2-504), and New Jersey’s courts, its Supreme Court now among them, have ordered discovery into the search itself as a condition of a fair trial. Accordingly, the argument is available even where no legislature has acted. Where the vendor asserts a trade secret, propose the middle path first, examination of the source code and validation records by the defense expert under a protective order, which answers the company’s confidentiality concern without asking the defendant to take the machine’s results on faith, and which courts find easier to grant than outright disclosure. If the court will not order even that, seek exclusion and, if the evidence is admitted, press the missing-algorithm instruction Won proposes, i.e., a permissive adverse inference where the system is peculiarly available to the prosecution and would reveal noncumulative favorable information, candidly framed as a proposal rather than existing law. A denial, once on the record, is itself a claim preserved.

The most important motion is often the first one, and it concerns evidence that has not yet vanished. Send a preservation demand at the outset, before retention windows close, naming the sources by type, body-camera and dashboard video, jail calls, cloud accounts, provider logs, and device location data, and saying in a sentence why each is expected to matter to the defense. Then be honest about what the demand can and cannot do, because the law of lost evidence runs on two tracks. Where destroyed material had exculpatory value apparent before it was lost and nothing comparable remains, due process is violated without more. Where the material was merely potentially useful, the defendant must prove the government acted in bad faith, a showing that almost always fails. A dated, specific demand does not move a case from the second track to the first, and counsel should not claim that it does. What it supplies is evidence that silence never generates: a dated record that the government was told what material to preserve and why the defense believed it could matter. That notice does not itself prove that the material’s exculpatory value was apparent or that its destruction was in bad faith, but it can become part of the showing under either track, especially when the request is specific, supported by objective facts, and followed by destruction without an innocent explanation. So treat the letter as the beginning of a practice, not a single act. Move for a preservation order when the government does not answer, renew the demand as new sources emerge, and for material in third-party hands, providers, platforms, private cameras, ask the court or the prosecution to direct its preservation, since the defense often cannot compel a stranger to the case on its own. And if identified material is lost anyway, the remedies to seek run from an adverse-inference instruction to exclusion of the government’s related proof to, in the rare extreme, dismissal, with the demand letter as the exhibit that makes each of them arguable rather than aspirational.

Disclosure, though, is only half the fight. The other half is whether the government’s evidence means what the government says it means. Move early for expert funding and frame the request as the Supreme Court framed the right, as the basic resources an adequate defense needs to test the proof against it, extended by analogy from the psychiatric expert the Court actually guaranteed to the experts a digital case demands, here a forensic examiner or someone who can probe a facial-recognition result. When the budget is thin and the request is denied, make it anyway and make it on the record, because the denial preserves the issue. With an expert in hand, attack the interpretation and not only the disclosure, the timestamp read without its time zone, the cell-tower estimate offered as a dot on a map, the probability presented as an identification, through the ordinary rules on admissibility. In some courts a pretrial reliability hearing is the vehicle; in others, it is a motion in limine. These cases are frequently won or lost on the interpretation, on what someone claimed the data showed, more than on the data itself.

Most of this has to happen before a plea, because most cases end in one. The Supreme Court has held that the right to effective assistance reaches the plea stage, and the Federal Judicial Center, in Criminal e-Discovery: A Pocket Guide for Judges (2015), advises counsel to review a reasonable and targeted portion of the electronic discovery, in usable form, before advising a client whether to accept a deal. Digital volume makes that standard hard to satisfy, which is precisely why the effort should be visible, the discovery sought, the format demanded, the continuance requested, and, if denied, the denial placed on the record. The old reflex of counseling a plea at arraignment with almost nothing known about the case cannot survive the scale of what now goes unreviewed. Where the caseload makes meaningful review impossible, counsel should say so plainly and move for the time and the resources to do it so that the constraint becomes a fact in the record and the court’s problem to answer, rather than a silent lapse later charged to the client.

For the many whose convictions are already final, the same framework cuts the other way. The evidence that was buried, the facial-recognition use that was never revealed, the extraction that hid the favorable file, all of it is the raw material of post-conviction relief, on the occasions it can be found. Public-records requests, renewed discovery motions, and fresh forensic examination of the original devices sometimes uncover what a trial never saw. The obstacles are high, the materiality standard and the closed retention window foremost among them, but the reversals show that evidence found late can still matter. The lesson for trial counsel is the same one seen from the other end. The record built now is the record a post-conviction lawyer will need later.

These steps have a common feature. Not one of them assumes the defense will win the motion. Each assumes it might lose and works to make the loss visible, dated, and reviewable, because the most serious problem is not that favorable evidence is denied but that its denial leaves no trace. The practical answer to a system that forgets whatever no one examined is to refuse to let the asking go unrecorded. That does not cure the imbalance. It is what a careful lawyer can do inside it now. What all of this has cost is easiest to see in the few cases where the needle was discovered anyway.

Conclusion

Return to Derrick Ulett. His case was, by the standard of everything that followed, a simple one. A single video from that lobby camera showed the scene out front, the prosecution did not turn it over, and at trial, the jury was led to believe no such video existed. A court could see what had happened. There was one discrete thing withheld, and one plain misstatement about it, and the two together were enough to undo the conviction. Ulett was fortunate. The concealment left a hole a person could readily point to.

Everything described here shows how that hole disappears. The same exculpatory video, today, need not be withheld at all. It can be handed over together with dozens of hours of other footage, in a format that cannot be searched, under a disclosure deadline of mere weeks, and simply never watched. No one has to lie to the jury, because the government can say, and say truthfully, that it disclosed everything. That is the strange reversal at the center of all of this. What hides the needle now is not concealment but disclosure itself.

Brady rested on a simple assignment of responsibility. The government holds the evidence, so the government must not hide the favorable part of it. That logic turns against itself once the government also controls the volume, the format, and the timing, because the party that builds the haystack is the party that decides whether the needle can ever be found, and no general rule yet written requires it to make the needle findable. The promise that the government cannot conceal the truth now depends, in practice, on the good faith of the government, which is the very thing the promise existed to distrust.

The people named here are the exceptions. Derrick Ulett, Craig Johnson, John Thompson, their convictions unraveled because someone, late and against the odds, found the thing that should have been disclosed. They are not the measure of the problem. They are its survivors, the narrow band of cases in which the needle happened to be discovered. The measure is everyone else, the files disclosed and never read, the defendants who pleaded guilty without ever learning what the file contained, the names that generate no opinion because no one ever learned there was anything to look for. Brady was meant to guarantee that the government could not bury the truth. The digital age has not repealed that guarantee. It has made it possible to honor the guarantee in form and bury the truth in substance with the same act, in plain sight, in a place the law does not yet require anyone to look.  

Sources: Legal Aid Society, In His Own Words: Derrick Ulett’s Mind Is Quiet After 16 Years (Jan. 15, 2026); Brooklyn Eagle Staff, Murder Charges Dismissed After 17-Year Prosecution, Brooklyn Eagle (Dec. 19, 2025); Nat’l Inst. of Standards & Tech., Mobile Device Forensic Tool Specification, Test Assertions and Test Cases, Version 3.1 (Feb. 2021); Fed. R. Crim. P. 16; Due Process Protections Act of 2020, Pub. L. No. 116-182, 134 Stat. 894; Fed. R. Crim. P. 5(f); Nat’l Registry of Exonerations, 2025 Annual Report (Apr. 6, 2026); Nat’l Registry of Exonerations, Registry of Exonerations (last visited July 21, 2026); Okla. Off. of the Att’y Gen., Attorney General’s Office to Prosecute Richard Glossip for Non-Capital Murder (June 9, 2025); Associated Press, Former Oklahoma Death Row Inmate Has a New Trial Set for a 1997 Killing of Motel Owner (June 23, 2026); Brief for Google LLC as Amicus Curiae, Chatrie v. United States, No. 25-112 (U.S. Mar. 2, 2026); Upturn, Mass Extraction: The Widespread Power of U.S. Law Enforcement to Search Mobile Phones (2020); Allison Sherry, Way Too Much Body Camera Footage: Police, Prosecutors and Defense Attorneys Struggle With Loads of Digital Evidence, Colorado Public Radio (Jan. 2, 2026); Shaun Boyd, Massive Volume of Digital Evidence Is Crippling Colorado’s Criminal Justice System, CBS Colorado (Dec. 23, 2025); Colo. Rev. Stat. § 24-31-902 (2025); S.B. 20-217, 72d Gen. Assemb., 2d Reg. Sess. (Colo. 2020); Colo. R. Crim. P. 16(I)(b)(1); Defendants’ Objection to Discovery Order, United States v. Herrera-Velutini, No. 3:22-cr-00342-SCC-HRV, ECF No. 838 (D.P.R. Apr. 2, 2025); Memorandum and Order, United States v. Herrera-Velutini, No. 3:22-cr-00342-SCC-HRV, ECF No. 819 (D.P.R. Mar. 12, 2025); U.S. Dep’t of Just., Off. of the Pardon Att’y, Clemency Grants by President Donald J. Trump (2025–Present) (updated July 6, 2026); Nat’l Inst. of Standards & Tech., Mobile Device Forensic Tool Specification, Test Assertions and Test Cases, Version 3.3 (Jan. 2025); Nat’l Inst. of Standards & Tech., Special Publication 800-101 Rev. 1, Guidelines on Mobile Device Forensics (May 2014); Nat’l Inst. of Standards & Tech., Interagency Report 8387, Digital Evidence Preservation: Considerations for Evidence Handlers (Sept. 2022); Lars Daniel, Attorneys: Is Your Cell Phone Report Missing Data?, Envista Forensics (July 8, 2024); Lance Sloves, Understanding Cellebrite Extractions: A Guide for Legal Professionals, Computer Forensic Services (Feb. 16, 2026, updated Feb. 18, 2026); Google, How Google Retains Data We Collect; Rebecca Darin Goldberg, You Can See My Face, Why Can’t I? Facial Recognition and Brady, 5 Colum. Hum. Rts. L. Rev. Online 261 (2021); Frankie D. Pereira De Leon, The Hidden Role of Facial Recognition in Police Investigations, Colum. Sci. & Tech. L. Rev. Blog (Mar. 29, 2026); Nat’l Registry of Exonerations, 2024 Annual Report (2025); JusticeText: Bringing AI Audiovisual Analysis to the Public Defender’s Office, Thomson Reuters Institute (May 15, 2025); Reduct, AI Tools for Public Defenders (Mar. 26, 2025); Letter, United States v. Ho, No. 1:25-cr-00003-JAV, ECF No. 42 (S.D.N.Y. July 17, 2025); Brief of Amici Curiae, Johnson v. State, No. 413, Sept. Term 2024 (Md. App. Ct. Dec. 2024); People v. Johnson, No. F071640 (Cal. Ct. App. July 11, 2019); Joint Elec. Tech. Working Grp., Recommendations for Electronically Stored Information (ESI) Discovery Production in Federal Criminal Cases (2012); Nat’l Registry of Exonerations, Michael Morton; Nat’l Ass’n of Crim. Def. Laws., Frye and Lafler: Much Ado About What We Do – and What Prosecutors and Judges Should Not Do (2012).

Subscribe today

Already a subscriber? Login