#CaseoftheWeekCase Law

Episode 128: eDiscovery Implications of Facial Recognition Technology: Lessons from State v. Arteaga

In Episode 128, our CEO and Founder, Kelly Twigger discusses State v. Arteaga, 476 N.J.Super. 36 (N.J. Super. Ct. App. Div. 2023), a decision that addresses the issue of whether a criminal defendant is entitled to discovery about how the facial recognition technology used to identify him works and the human elements that entered into the process.


Good morning and welcome to our first Case of the Week for 2024. The Case of the Week series is brought to you by eDiscovery Assistant in partnership with ACEDS. I hope each of you had a wonderful, restful holiday break, and we’re excited to be back with you for 2024.

My name is Kelly Twigger. I am the CEO and founder at eDiscovery Assistant, your GPS for ediscovery knowledge and education. Thanks so much for joining me this morning. Each week on the Case of the Week, I choose a recent decision in eDiscovery and talk to you about the practical implications of that decision. This week’s decision discusses the role of facial recognition technology in a criminal matter, and we’re going to talk also about its implications in civil matters.

A couple of announcements. The 11th Annual University of Florida eDiscovery Conference is coming up on February 28th and 29th in Gainesville, Florida, and registration is open. If you haven’t attended this event in the past, you’ll want to check it out. It’s one of the best practical events in ediscovery, and it is free to attend virtually. You can also attend in person. There are a few in person registrations left and the fee is pretty nominal. You can register here.

Second, eDiscovery Assistant will be in New York City at LegalWeek, and we’d love to talk to you about whether the platform can provide value for your practice or organization. We’ll be sending out an email to everyone who’s on our email blog list, so keep your eyes peeled for that. Here is a link to schedule time with us at the conference: schedule a meeting or demo at LegalWeek 2024.

Finally, we’ll be rolling out our 2023 annual case law report just in time for the UF eDiscovery Conference in February, so keep your eyes peeled for that. We’ve had more than 400 downloads of the report each year since 2020, and hopefully this year will not be any different.

All right, let’s dive into this week’s case. Forgive me for a little bit of frog in my throat today coming back from break. Today’s decision comes to us from the case of State v. Arteaga in the New Jersey Superior Court, the Appellate Division. This is a decision written by Judge Hany Mawla.


We are before the New Jersey State Court of Appeals on an interlocutory appeal following the indictment of Francisco Arteaga for robbery and aggravated assault. The Court granted Arteaga’s leave to appeal from a May 13, 2022, order that denied his motion to compel that the state provide discovery related to the facial recognition technology used to develop a picture of him, which was then used to identify and charge him.


On November 29, 2019, a man entered a store that offered international wire transfers, cell phone repairs and accessories. He approached the rear counter and asked an employee, Jennifer Vasquez-Arias, who was counting money about wiring funds to South America. When she turned toward her computer, he walked toward an open door behind the counter that led to an office that contained a cash register. The man surprised Vasquez-Arius with a handgun and grabbed the money she had just counted. She tried to stop him, but he pistol whipped her and escaped with just under $9,000.

When a police detective arrived, Vasquez-Arius described the man as “a Hispanic male wearing a black skully hat” and carrying a black handgun. She recalled that he had entered the store earlier in the day, stood at the end of a line of customers and left before he could be served. The store manager, Judy Cardozo, was not present for the robbery, but when reviewing surveillance footage of the incident, thought that she recognized the perpetrator. She recalled that a man who she believed to be the same person had approached her during a prior visit, nervously asked about a cell phone case, waited in line briefly, and left the store without making a purchase. Not long afterwards, she spotted him again outside, adjusting his gloves as he walked back toward the store.

The store’s surveillance camera captured the earlier visits, in addition to the robbery. Detectives retrieved other footage from a nearby property, which showed the man walking around near the store for approximately ten minutes. They then generated a still image from the footage and sent it to the New Jersey Regional Operations Intelligence Center for facial recognition analysis. An investigator advised that there were no matches, but that he could rerun the inquiry if detectives provided him with a better image.

Instead of moving forward with the analysis from the New Jersey Regional Operations Intelligence Center, detectives sent all of the raw surveillance footage to the facial identification section of the New York Police Department Real Time Crime Center (RTCC).

A detective at the RTCC captured a still image from the footage, compared it against the Center’s databases, and offered defendant as a “possible match.” So, they used the technology, they found a possible match, and then they exported a picture of the defendant from their database to send back to these detectives.

Detectives working the case in New Jersey created two different photo arrays comprised of five filler photos and the photo that the NYPD RTCC provided of the defendant from its database. The first array was shown to Vasquez-Arias and the second to Cardozo in separate video recorded interviews. Both witnesses independently verified Arteaga’s photo as that of the perpetrator.

The grand jury then indicted Arteaga on charges for robbery, aggravated assault, and several other charges. Upon learning that the facial recognition technology — also called FRT in the decision — played a role in identifying the defendant, defense counsel sent the State a letter asking for the following discovery.  This is a long list, but it’s one that I think is really important for you to understand the detail that defense counsel wanted here:

  1. The name and manufacturer of the facial recognition software used to conduct the search in this case and the algorithm, version number, and years developed;
  2. The source code for the facial recognition algorithm;
  3. A list of what measurements, nodal points, or other unique identifying marks are used by the system in creating facial feature vectors, including if those marks are weighted differently, the scores given to each respective mark;
  4. The error rates for the facial recognition system used, including false accept and false reject rates — also called false match and false non-match rates — as well as documentation as to how the error rates were calculated, including whether they reflect test or operational conditions;
  5. The performance of the algorithms used on applicable NIST face recognition vendor tests, if applicable; (Recall that NIST is the National Institute of Standards and Technology)
  6. The original copy or the query or probe photo submitted to the real time crime center facial identification section;
  7. All edited copies of the query or probe photos submitted to the facial recognition system – noting, if applicable, which edited copy produced the candidate list that the defendant was in and a list of edits, filters, or other modifications made to that photo;
  8. A copy of the database photo matched to the query or probe photo and the percentage of the match, rank, number, or confidence score assigned to the photo by the facial recognition system in the candidate list;
  9. A list or description of the rank, number, or confidence scores produced by the system, including the scale on which the system is based;
  10. A complete candidate list returned by the face recognition, or the first 20 candidates in the candidate list if longer than 20 in rank order, and including the percentage of the match or confidence score assigned to each photo by the facial recognition system;
  11. A list of the parameters of the database used, including: (a) how many photos are in the database, (b) how the photos are obtained, (c) how long the photos are stored, (d) how often the database is purged, (e) what the process is for getting removed from the database, (f) who has access to the database, (g) how the database is maintained, and (h) the privacy policy for the database;
  12. The report produced by the analyst or technician who ran the facial recognition software, including any notes made about the possible match relative to any other individuals on the candidate list; and
  13. The name, training, certifications or qualifications of the analyst who ran the facial recognition search query.

Now, if you followed that whole list, you know that this is very extensive information regarding the FRT system that obviously had to be provided by an expert to defense counsel. The State answered the defendant’s discovery request by providing two items — those sought in the 8th and 10th requests in defense counsel’s letter — and they provided specifically the NYPD RTCC search result report that identified the defendant as a possible match to the individual in the surveillance footage, the still images that were used for comparison, the first ten possible matching candidates for each presented in rank order and accompanied by their confidence score, and second, a short series of handwritten notes by an NYPD RTCC analyst. Notably, many of the results that are attached to the report entail two independent sets of possible matches for a given probe, suggesting that either of the probes were compared against more than one database or that more than one FRT was used.

Defendant then, after receiving that information, moved to compel the State to provide the remaining eleven items in defense’s discovery request. The motion included a detailed declaration from defense’s proposed FRT expert, opining about the accuracy issues that are associated with FRT and why the defense needed the discovery. The expert stated, “understanding the [NYPD RTCC’s facial recognition] model, the data that was used to train the model, and the class-specific performance for the image(s) in this case are critical to understand the reliability of the output.” The expert concluded that without that additional information the current results of the image facial recognition software in this case cannot be considered scientifically replicable or relevant.

Defense counsel made two separate arguments regarding the reliability of the FRT.

First, defense counsel argued that it could not be assessed without the discovery sought because it was highly system dependent and dependent on choices made by the operator at every step of the process. Defense cited research and data, including that from NIST, showing that, “face recognition can be extremely poor at identifying a person in a low-resolution image” like the surveillance still photo that was used here. Also, the images are manipulated by the FRT operator to obtain a normalized face to run against the database. Defendant alleged that the FRT algorithm could have a high error rate and large databases can return an incorrect match due to the overrepresentation of racial minorities, an imbalance which stems from the fact that databases are populated by photographs taken during arrest and parole. The defense also noted that “a human must choose which photo from the candidate list [generated by the FRT] will be forwarded to the investigating agency as the ‘match’.” Defendant contended this part of the process was entirely subjective and what prompted their request for the identification of the analysts, the understanding of their credentials, as well as the process that was undertaken to choose the photo.

Defendant also argued that because the FRT could produce high levels of false positives, it’s critical for the defense to know exactly which tools were employed and how they were used in order to evaluate the reliability of the methods that were used to bring defendant into this case. The remaining eleven items the State failed to produce were necessary to provide context for and assess the reliability of the match. The defense also noted in it that its discovery requests were specific and tailored, and that this is the first known and disclosed use of facial recognition in New Jersey, requiring a full and fair examination of the technology that is in the public’s best interest.

Following these arguments, the motion judge denied the motion for the discovery and issued a written opinion concluding that the State had no obligation to produce the discovery because the FRT was not within its care, custody, or control. The motion judge also stated that the requested material was not “Brady material because the FRT produced nothing more than a photograph that resembles the photograph provided from the still shot of the video taken from the scene of the incident.”

According to the motion judge, “At best, the West New York Police Department was provided with an additional photograph to compile the six pack that was shown to the witness in this case. It was the witnesses’ identifications that formed the basis for probable cause to arrest [d]efendant.” On appeal, Arteaga argued that, without any information about the reliability of the FRT, admitting the photo into evidence was the fruit of the poisonous tree.

Now, of note here is that the defendant was joined on this appeal with amicus briefs from the Electronic Privacy Information Center, Electronic Frontier Foundation, and the National Association of Criminal Defense Lawyers. This is a big issue that’s going to have a huge impact for criminal cases. It’s also one that’s going to play over into civil matters.

As we’ve discussed multiple times here on the Case of the Week, we often see the analysis of these types of technologies first come up in criminal matters before they move to a civil context. But, as we’ll discuss later in the takeaways, we’re seeing so many applications of facial recognition technology now for everyday people that are not accused of crimes that it is probably only a matter of months, if not shorter than that, that we’ll start seeing the implications of facial recognition technology in civil matters as well.


What is the Court’s analysis on appeal? Well, of course, the Court starts with the standard of review, which is generally an abuse of discretion. However, the Court notes here that “where a matter involves novel scientific evidence in a criminal proceeding, we may exercise independent review of the relevant authorities, including judicial opinions and scientific literature.”

The Court goes on to engage in a discussion of relevance and whether or not the discovery sought here was, in fact, Brady material — i.e., evidence that is potentially favorable to the defense. In doing so, the Court distinguished past cases with the situation before it and that they are dealing with eyewitnesses who have already identified the perpetrator, and the identification was found admissible. In reversing the denial of discovery, the Appellate Court found that the facts of this case convinced it that the defendant would be deprived of due process if he does not have access to the raw materials integral to the building of an effective defense, and that the evidence sought here is directly tied to the defense’s ability to test the reliability of the facial recognition technology.

As such, the Court found that it is vital to impeach the witnesses’ identification, challenge the State’s investigation, create reasonable doubt, and demonstrate third party guilt. The Court also disagreed with the underlying ruling that the information about the FRT was not in the possession, custody, or control of the State. The prosecutor sent a request to the NYPD RTCC, which in turn complied by producing the information used to accuse the defendant. Moreover, the prosecutor obtained discovery materials from the NYPD RTCC in responding to two items in defense counsel’s discovery letter. The State did not argue it cannot obtain the additional information sought by defendant for proprietary or other reasons, only that defendants should subpoena the information.

The Court rejected that argument given that the burden lies with the State, especially given the fact that the FRT is novel and untested and the possibility that errors in the technology may exculpate the defendant, and there was very little evidence in the record regarding the software at issue.

However, the defendant’s expert report and the secondary sources decided by defense counsel and the amici provided the Court convincing evidence of FRT’s novelty, the human agency involved in generating images, and the fact that the FRT’s veracity has not been tested or found reliable on an evidentiary basis by any New Jersey court. The Court then walked through how the facial recognition technology works, and this is really just plain interesting for you.

FRT can detect a human face with an image. It uses an algorithm to make the analysis, which recognizes nodal points on the face — the peaks and valleys that make up human facial features — and measures them against corresponding features in comparison images. The underlying algorithm is developed through machine learning that is trained over time to recognize potential matches by being asked to compare several thousand images against a database. In each instance, the program estimates whether a match exists, is told whether the result is correct, and then uses this data regarding successes or failures to inform subsequent evaluations, eventually focusing on those features that have been the most reliable indicators of a match. The process begins when the person operating the software selects a probe image captured from surveillance footage that features the perpetrator’s face. Poorer quality images can be edited for lighting or color correction, to enhance detail, or even to change facial expression.

The technology then breaks down the image into component features and distills them into a face print — a map written in code that measures the distance between features, lines, and facial elements. It then compares this face print against others in the database, assigning scores to each based on the extent to which corresponding features line up, and returning a list with those with the highest scores ordered by rank.

The Court then looked at the reliability issue raised by defendant on appeal. The reliability of the results, according to the Court, may depend on many variables encountered throughout the process. Some of the literature cited by the Court here — and there’s a lot of really comprehensive articles here that you should take a look at if you’re looking at this issue — suggests that the quality and diversity of images used for training the technology may influence the technology’s ability to recognize faces different from those in the initial set, often resulting in poor performance with analysis of non-white faces. Moreover, the quality of a probe image — including its resolution, ambient lighting, and facial expression, as well as the extent of any editing performed to it — may impact the accuracy of the resulting faceprint and the software’s ability to meaningfully compare it with those in the database.

If you haven’t picked up on it so far, this is highly complex and requires a lot of human input.

The defendant argues that the reliability of the FRT is highly dependent on its design and training, as well as the parameters used by the analyst who operates it. He asserts that the discovery sought is relevant to assessing the reliability of the FRT results used to identify him.

The defendant’s expert explains that the source code and related materials are necessary to analyze the software’s design and training methods for inherent structural flaws in the software that can introduce bias and compromise the accuracy of the results generated by the software. The reliability of the technology implicates the accuracy of the eyewitness identifications, the thoroughness of the State’s investigation, and the ability to prove the existence of other viable suspects. Defendant asserts that if the discovery is not provided, the State should be barred from introducing any of the identification because they are the fruits of untested technology.

The Court found that the request for the FRT discovery was relevant to defendant’s ability to impeach the identification and the investigation and his overall ability to establish reasonable doubt at trial. According to the Court, the FRT’s reliability has obvious implications for the accuracy of the identification process because an array constructed around a mistaken potential match would leave the witness with no actual perpetrator to choose.

The reliability of the technology bears direct relevance to the quality and thoroughness of a broader criminal investigation and whether the potential matches the software returned yielded any other viable alternate suspects to establish third party guilt. Defendant’s request for the design, identity specifications, and operation of the program or programs used for analysis, and the databases or databases used for comparison are relevant to the FRT’s reliability, according to the Court:

Here the items sought by the defense have a direct link to testing FRT’s reliability and bear on the defendant’s guilt or innocence. Given FRT’s novelty, no one — including us — can reasonably conclude without the discovery whether the evidence is exculpatory or “merely potentially useful evidence.” For these reasons, it must be produced.

On that basis, the Court reversed and remanded for an entry of the order directing the State to provide the eleven remaining items of discovery requested by the defense. It also allowed the motion judge to enter any appropriate protective order for the information, to order the in camera review of the materials received from the State, and to hold a Daubert hearing, if necessary, on the entry of the potential expert evidence.


What are our takeaways from this case? Well, I picked this one because we spent all of last year discussing AI, and we’ve started 2024 doing exactly the same thing. We are living in a time when AI is the headline of every article, and facial recognition technology is one aspect of AI. It’s used to open your phone, your fingerprints and eyes are collected if you use Clear to fly, theme parks tie your tickets to your fingerprints to prevent theft. Those are just a few examples of how facial recognition technology, biometrics, lots of AI, is being used to identify consumers — not just for criminal behavior, but generally speaking.  New technologies are all around us, and they are creating evidence that can be used in both criminal and civil contexts.

This decision really highlights for us the need for counsel to be engaged in understanding how these technologies work and what needs to be understood to question their use to prove the elements of a theory of liability. And I know I spent a lot of time dragging you through that list of both what the defense asked for from its expert, but also how the system works. And that’s because this decision offers you a great explanation of how facial recognition technology works, as well as a list of sources that you can use to cite to should you have these kinds of issues.

It is counsel’s job to identify when that understanding is necessary and to go out and find an expert when needed. It’s no different than understanding when you need a forensics expert, when you want to try and use technology assisted review, and what that process includes. And we’ve discussed both of those multiple times here on the Case of the Week. 

The Court here also addressed the issue of possession, custody, or control, in that the facial recognition technology sat with a corresponding state agency. But the Court noted the prosecutor’s ease of interacting with the agency and that the information it did provide in response to the original request came directly from that agency.

Possession, custody, or control has become a very prominent issue in discovery in the last 18 months, and it’s one that you want to be up to speed on. Keep in mind that there are variances on how the courts are interpreting possession, custody, or control by jurisdiction and in terms of whether the information is within an individual employee, an individual, or within a corporation’s possession, custody, or control.

Read the case law on these new technologies. Leverage it to help your clients raise the issues. You need to know enough to issue spot and find an expert to help you work through it if you don’t have the required expertise. Both your client’s case and the future development of the law on these technologies depend on counsel identifying these issues and addressing them.


That’s our Case of the Week for this week. Thanks so much for joining me and sticking in for a little bit longer session. We’ll be back again next week with another decision from our eDiscovery Assistant database. As always, if you have suggestions for a case to be covered on the Case of the Week, drop me a line. If you’d like to receive the Case of the Week delivered directly to your inbox via our weekly newsletter, you can sign up on our blog. If you’re interested in doing a free trial of our case law and resource database, you can sign up to get started.

Thanks so much. Have a terrific week, and Happy New Year!

Privacy Settings
We use cookies to enhance your experience while using our website. If you are using our Services via a browser you can restrict, block or remove cookies through your web browser settings. We also use content and scripts from third parties that may use tracking technologies. You can selectively provide your consent below to allow such third party embeds. For complete information about the cookies we use, data we collect and how we process them, please check our Privacy Policy
Consent to display content from - Youtube
Consent to display content from - Vimeo
Google Maps
Consent to display content from - Google
Consent to display content from - Spotify
Sound Cloud
Consent to display content from - Sound