Opinion by Judge BETTY B. FLETCHER; Dissent by Judge BEA.
BETTY B. FLETCHER, Circuit Judge:Defendant Rachel Jernigan was arrested on November 10, 2000, for allegedly robbing three banks. After Jernigan was placed in custody and awaiting trial, two more area banks were robbed by a woman whose description bore an uncanny physical resemblance to hers: both women were roughly five feet tall, Hispanic,1 and had acne or pock-marked complexions. Although the prosecution knew that other nearby banks had been robbed by a diminutive, Hispanic female with poor skin after Jernigan’s arrest, the prosecution failed to relay this information to defense counsel.
Proceeding without knowledge of the second alleged bank robber, Jernigan’s counsel argued at trial simply that his client was misidentified. However, the jury was not persuaded, and Jernigan was convicted of bank robbery on March 23, 2001.
While in prison Jernigan learned that a woman fitting a similar description had been arrested for robbing several banks in the area. In January 2004, Jernigan filed a motion for a new trial asserting that (1) the government violated her due process rights under Brady v. Maryland, 373 U.S. 83, 83 S.Ct. 1194, 10 L.Ed.2d 215 (1963), by failing to disclose before trial material, exculpatory evidence known to the government, and alternatively that (2) evidence discovered after trial required that Jerni-gan receive a new trial pursuant to Federal Rule of Criminal Procedure 33. The district court denied her motion in January 2005, and Jernigan appealed. After a panel of this court affirmed the district court, we voted to rehear this case en banc. We disagree with both the original panel and the district court and hold that the suppressed evidence was material to Jerni-gan’s guilt.2 The district court’s decision is hereby reversed, and we remand for a new trial.
I.
On September 20, 2000, the Bank of America branch in Gilbert, Arizona was robbed by a short, Hispanic woman with a pock-marked face. The robber posed as a bank customer. When her turn in line came, she passed a sloppy, hand-written note to Elizabeth Chlupsa, the victim bank teller, threatening to shoot if Chlupsa did not hand over all of the money in her drawer. Chlupsa complied with the demand note, and the robber fled without having said a word.
Jernigan became a suspect after a chance conversation between the FBI *1052agent investigating the bank robbery and a postal inspector who was investigating Jer-nigan in connection with some shoplifting incidents at a local post office. During their conversation, the postal inspector noted that Jernigan fit the description of the unidentified bank robber. After reviewing photographs of Jernigan and comparing them to surveillance photographs from the September 20th robbery, the FBI agent focused his investigation on Jerni-gan. The agent created a photographic lineup that included Jernigan and showed the photos to the victim bank teller, who identified Jernigan as the woman who had robbed her. No other eyewitnesses were shown the photospread—or any other photographs of Jernigan—until five or six months later.
Police arrested Jernigan on November 10, 2000, and she has remained in custody since that time. In addition to the September 20, 2000 robbery, police also charged Jernigan with two additional bank robberies: an October 11, 2000 robbery at 906 East Baseline Road in Tempe, and an October 25, 2000 robbery at 2298 North Alma School in Chandler. Following Jer-nigan’s incarceration, but before her trial, two other bank robberies were committed—one on November 28, 2000, and another on November 30, 2000—by a person matching Jernigan’s description: a short, Hispanic woman with acne. One bank was located across the street from one of the banks allegedly robbed by Jernigan; the other bank was located approximately ten miles away.
The district court ultimately severed the charges involving the September 20th robbery from the other robberies with which she was charged. Jernigan’s trial for the September 20th robbery began on March 20, 2001. At trial, the government relied entirely on the accounts of five eyewitnesses and the bank surveillance video. The video did not provide a clean look at the bank robber’s face3 and was used primarily to bolster the eyewitnesses’ testimony. No physical evidence tied Jernigan to the robbery. Jernigan flatly denied involvement in any bank robberies and, at trial, her counsel argued that she had been misidentified by the witnesses. Counsel did not, however, suggest to the jury who might have been the robber if Jernigan were not.
The jury convicted Jernigan of both armed bank robbery and use of a firearm during an armed bank robbery. The district court sentenced her to 168 months in jail and five years of supervised release. The remaining bank robbery charges were dismissed by stipulation.
On December 11, 2001, Juanita Rodríguez-Gallegos robbed the same bank allegedly robbed by Jernigan on September 20, 2000. During that robbery the victim teller, Kathleen Golliher, placed a tracking device in the stolen money. Police stopped Rodríguez-Gallegos half an hour after the robbery and Golliher identified her as the robber.4 The police report described Rodriguez-Gallegos as a Hispanic female, 4'11" and 125 pounds, with brown eyes, black hair, and pock-marked cheeks. Police charged Rodriguez-Galle*1053gos with the November 28, 2000, November 30, 2000, and December 11, 2001 bank robberies, and with one count of brandishing a firearm during a violent crime. She pled guilty to the firearm offense, and the remaining charges were dropped.
After learning of Rodriguez-Gallegos’s arrest from fellow inmates, Jernigan moved for a new trial. In her motion, Jernigan asserted that the government had failed to meet its Brady obligations by not disclosing the existence of a phenotypi-cally similar bank robber who had been robbing banks in the same area after Jer-nigan’s incarceration. She also argued that the district court should grant her a new trial under Rule 33 based upon the post-conviction developments involving Rodriguez Gallegos. The district court denied the motion on both grounds, and Jer-nigan appealed.
II.
In Brady v. Maryland, the Supreme Court explained, “[s]ociety wins not only when the guilty are convicted but when criminal trials are fair; our system of the administration of justice suffers when any accused is treated unfairly.” 373 U.S. at 87, 83 S.Ct. 1194. The Court also noted that prosecutors are charged not only with winning trials but with seeking justice. Id. at 87-88, 83 S.Ct. 1194 & n. 2. Premised on this understanding of fairness and prosecutorial responsibility, the Court held that the suppression of evidence favorable to any accused violates the due process of law, irrespective of whether the suppression is done in good faith or bad. Id. at 87, 83 S.Ct. 1194.
Brady’s progeny have articulated three elements that defendants must prove to show a Brady violation. See Benn v. Lambert, 283 F.3d 1040, 1052 (9th Cir.2002). First, the suppressed evidence must be favorable to the accused. See United States v. Bagley, 473 U.S. 667, 676, 105 S.Ct. 3375, 87 L.Ed.2d 481 (1985) (citing Brady, 373 U.S. at 87, 83 S.Ct. 1194). Second, the evidence must have been suppressed by the government, either willfully or inadvertently.5 See United States v. Agurs, 427 U.S. 97, 110, 96 S.Ct. 2392, 49 L.Ed.2d 342 (1976). And third, the suppressed evidence must be material to the guilt or innocence of the defendant. See Bagley, 473 U.S. at 676-78, 105 S.Ct. 3375. Because the government does not dispute that the evidence of an additional bank robber matching Jernigaris description was favorable to Jernigan and that the government failed to provide it to defense counsel, the only issue in dispute is whether the evidence was material.
The touchstone of materiality review is whether admission of the suppressed evidence would have created a “ ‘reasonable probability’ of a different result.”6 Kyles v. Whitley, 514 U.S. 419, 434, 115 S.Ct. 1555, 131 L.Ed.2d 490 (1995) (citing Bagley, 473 U.S. at 682, 105 S.Ct. *10543375). And as the Kyles Court emphasized, the adjective “reasonable” is important. Id. A defendant need not show that she “would more likely than not have received a different verdict with the evidence.” Id. Instead, she must show only that “the government’s evidentiary suppression ‘undermines confidence in the outcome of the trial.’ ”7 Id. (quoting Bag-ley, 473 U.S. at 678, 105 S.Ct. 3375). In considering whether the failure to disclose exculpatory evidence undermines confidence in the outcome, judges must “ ‘undertake a careful, balanced evaluation of the nature and strength of both the evidence the defense was prevented from presenting and the evidence each side presented at trial.’ ” Bailey v. Rae, 339 F.3d 1107, 1119 (9th Cir.2003) (quoting Boss v. Pierce, 263 F.3d 734, 745 (7th Cir.2001)). In other words, “the withheld evidence must be analyzed ‘in the context of the entire record.’ ” Benn, 283 F.3d at 1053 (quoting Agurs, 427 U.S. at 112, 96 S.Ct. 2392).
Here, the suppressed evidence substantially erodes the already questionable value of the eyewitness identifications. “Centuries of experience in the administration of criminal justice have shown that convictions based solely on testimony that identifies a defendant previously unknown to the witness is highly suspect. Of all the various kinds of evidence it is the least reliable, especially where unsupported by corroborating evidence.” Jackson v. Fogg, 589 F.2d 108, 112 (2d Cir.1978); see also Department of Justice, Eyewitness Evidence: A Guide for Law Enforcement 1, 3 (Oct.1999); Gary L. Wells et al., Eyewitness Identification Procedures: Recommendations For Lineups and Photospreads, 22 L. & Hum. Behav. 603, 619-27 (1998); Gary L. Wells et al., Accuracy, Confidence, and Juror Perceptions in Eyewitness Identification, 64 J. Applied Psyohol. 440 (1979). Cross-racial identifications, such as the eyewitness accounts offered against Jernigan, are particularly suspect. See Harvey Gee, Eyewitness Testimony and Cross-Racial Identification, 35 New Eng.L. Rev. 835 (2001) (reviewing Elizabeth F. Loftus, Eyewitness Testimony (1996)); John P. Rutledge, They All Look Alike: The Inaccuracy of Cross-Racial Identifications, 28 Am. J. Crim. L. 207 (2001). In a case that turned entirely on eyewitness identifications, the presence of a second robber in the same area fitting the very same physical description was bound to “substantially reduce[ ] or destroy[ ]” the “value” of the eyewitness testimony. See Kyles, 514 U.S. at 441, 115 S.Ct. 1555.
Five eyewitnesses testified at Jernigan’s trial, describing the robber as, inter alia, “a very, very short Hispanic, what I thought was Hispanic-looking lady.” “She was very small, either five-foot or under five foot.” She was “either Hispanic or Oriental.” She was a “[v]ery small person, possibly Hispanic.” She was “a short Asian woman.” “[S]he may have had acne, kind of pocked.” She did not have “a clear complexion.” “She had very dark hair, black.”
The government knew shortly after Jer-nigan’s arrest, but did not disclose, that *1055witnesses to two additional bank robberies, which took place close to the bank she was charged with robbing, offered strikingly similar descriptions of the perpetrator. Witnesses described the November 28th robber as “Hispanic or Oriental” with “[a] little acne,” “5T'-5'2"” and “[d]ark brown hair.” The November 30th robber was described as “Female,” “Hispanic,” “5'1"— 5'2",” with a “Pot [sic] Marked Face.” She was also described as a “small Asian woman under 5 ft.” The police bulletin described a “Hispanic female, 30s, 5'02", 130 pounds, black hair, pock marked cheeks and braces.”
Moreover, the victim tellers reported similar modus operandi in all of the robberies. The victim teller testified that during the September 20th robbery the robber slid her a note that read, “Don’t make a big scene, give me all your money, don’t give me any dye packs or tracking devices and don’t press the alarm or else I will shoot.” The note was “handwritten” in a “fairly big and kind of sloppy ... printed” fashion. On November 30th the robber passed a “note” to the victim teller which said, “in very sloppy writing,” “Do not set off alarm, Give me all of your money, Please do not set off your alarm.” Similarly, on November 30th, the robber passed a note stating, “Don’t turn on any alarms, or you’re dead.”
The uncanny similarity between the descriptions of the alleged robbers would be noteworthy even if the eyewitness accounts did not describe a most unlikely bank robber. In 2000 (the year the robberies in question took place), only six percent of all bank robbery perpetrators were female. Only six percent of bank robbers overall (male and female) were Hispanic. See U.S. Fed. Bureau of Investigation, Summary and Interpretation of Bank Crime Statistics, 2000 (2002). The likelihood of two short, Hispanic female robbers with pockmarked skin holding up banks in the same area is therefore extremely low.
The existence of another bank robber for whom Jernigan may well have been mistaken also magnifies the significance of the gaps and inconsistencies in the prosecution’s case. The most obvious gap, as noted earlier, was the complete lack of physical evidence connecting Jernigan to the crime. Even after Jernigan was arrested, the police failed to produce any physical evidence connecting her to the crime: a fingerprint lifted from the victim teller’s window did not match Jernigan’s print, and, after Jernigan was arrested, the police failed to find the stolen money, the firearm used to conduct the robbery, or any clothing resembling that worn by the robber.
Furthermore, the government’s witnesses provided inaccurate or inconsistent testimony, which is troubling, but not atypical of eyewitness accounts. For instance, Chlupsa, the victim teller on September 20th, maintained both immediately after the robbery and during trial that the perpetrator had no tattoos and did not have painted fingernails. However, Jerni-gan has a number of tattoos on her hands and forearms, as evidenced in photographs submitted to the district court and com firmed by the FBI shortly after the arrest.
Chlupsa also described the perpetrator as having plucked eyebrows, wearing “very, very, very, dark eyeliner,” “a lot of makeup,” and a lot of eyeliner. By contrast, Lorraine Hawley, a customer who stood next to the perpetrator in line, described the robber as wearing “little to no make-up” and no lipstick.8 The eyewit*1056nesses’ tentativeness is equally telling. Hawley, who claimed to get a good look at the robber, had some trouble picking Jer-nigan out of a photospread, stating, “I want to say it looks like this one.... I would say this would be the one.” Golliher appears to have had trouble as well, stating, “By—I just felt that was the one that—that was whom I saw.”
These problems are not altogether surprising given that all but one of the eyewitnesses viewed the photospread five to six months after the incident. Hawley first saw the photospread two days before trial—six months after the robbery. Three others—Golliher, employee Yarjanic Nath, and customer Donovan Grierson—were all shown the photos five months after the robbery. This delay, which goes totally unexplained, also detracts from the reliability of the identifications.
Finally, the probability that Jernigan was confused with the real bank robber is bolstered further by various pieces of exculpatory evidence.9 For instance, the getaway car for the October 11, 2000 robbery—a black Toyota 4-Runner—matched the description of the November 30, 2000 get-away car used by Rodriguez-Gallegos. Additionally, the victim teller in the October 11, 2000 robbery could not identify the woman who robbed her from the photo-spread containing Jernigan; however, the victim declared that the October 11 robber was the woman in the surveillance video from the September 20, 2000 robbery, suggesting that the September 20th robber was other than Jernigan—even possibly Rodriguez-Gallegos.
When considered in isolation, the various gaps, inconsistencies, and exculpatory details in the record might be insufficient to trump the testimony of the five eyewitnesses. But when this evidence is considered collectively, it undermines confidence in the outcome of the trial. It may be that two five-foot-tall Hispanic women with pock-marked skin were robbing the same banks in the same area with the same modus operandi and the same getaway vehicles. And it may be that the police just happened to miss finding the gun used in the robbery for which Jernigan was tried, the clothes worn in the robbery, and the money stolen in the robbery, when apprehending Jernigan. But it also may be that the hesitant and conflicting identifications offered by the eyewitnesses many months after the event were incorrect, and that the witnesses simply picked Jernigan out of the photospread because she was the woman who looked the most like the real culprit.10
As we view the withheld evidence in the context of the entire record, it is apparent to us that the evidence was material and that Jernigan was prejudiced by its suppression. Withholding knowledge of a second suspect conflicts with the Supreme Court’s directive that “the criminal trial, as distinct from the prosecutor’s private deliberations, [be preserved] as the chosen forum for ascertaining the truth about criminal accusations.” Kyles, 514 U.S. at *1057440, 115 S.Ct. 1555. By suppressing this evidence, the prosecution arrogated to itself a central function belonging to the criminal jury and pursued its role as adversary to the exclusion of its role as architect of a just trial. Cf. Brady, 373 U.S. at 87-88, 83 S.Ct. 1194 & n. 2. The government has deprived Jernigan of a fair trial and placed a possibly innocent woman behind bars. Because the evidence withheld by the government was material, we reverse the decision of the panel and district court, and remand to the district court for further proceedings consistent with our opinion.
REVERSED AND REMANDED.
. "Latina” may be the more accurate term but throughout the proceedings "Hispanic” has been used.
. We therefore need not reach the question of whether the district court erred in denying Jernigan’s motion for a new trial based on Federal Rule of Criminal Procedure 33.
. We have reviewed the surveillance video and still images taken from the video and agree that this evidence does not identify Jer-nigan as the robber. The image is of poor quality and the robber is wearing a hat that obscures her face.
. Golliher asserted during the hearing on the motion for a new trial that the two bank robberies, both of which she witnessed, were conducted by different women. The robberies took place 15 months apart. Golliher was not the victim teller in the September 20, 2000 robbery. Moreover, she was at some distance from the robber and gave conflicting accounts of the woman’s complexion and age.
. The investigation of all of the robberies was conducted by an FBI agent who apparently never told the prosecutor about the November robberies. There is no indication that anyone in the U.S. Attorney's office acted in bad faith.
. The dissenting opinion presents a red herring and side steps the issue before the court. At dispute is not whether Jernigan and Rodriguez-Gallegos "look alike.” At the time of Jernigan’s trial, the government did not know who committed the November robberies. As a result, what Rodriguez-Gallegos actually looks like is completely irrelevant to the Brady analysis. The question is whether a reasonable probability existed that the jury would have arrived at a different result if provided with the excluded evidence. More specifically, the question is whether the jury, when presented with nothing more than shaky, cross-racial eyewitness identifications, unsupported by any physical evidence, would have 'arrived at a different result when informed that a woman described in uncannily similar terms—terms describing a most unlikely bank robber—was robbing banks in the same area *1054just days after Jernigan’s incarceration. Given the extraordinary unlikelihood of two such robbers existing at the same time in the same place, the majority believes that the jury may well have accepted Jernigan's "misidentification” defense had the suppressed evidence been admitted.
. Kyles specifically rejected the idea that Brady required a sufficiency of the evidence test. 514 U.S. at 434-35, 115 S.Ct. 1555. Defendants may prevail in a Brady challenge even if the evidence remains sufficient to convict following the introduction of the suppressed evidence. Id. Defendants need only show "that the favorable evidence could reasonably be taken to put the whole case in such a different light as to undermine confidence in the verdict.” Id. at 435, 115 S.Ct. 1555.
. Witnesses also had a difficult time determining whether the robber was Asian or Hispanic. Although these inconsistencies may be less significant, they highlight the uncertainty of the eyewitness identifications.
. Jernigan was questioned by a polygraph expert prior to trial. The polygrapher asked Jernigan about robbing banks both as a general matter and in regard to the three banks for which she was originally arrested. Jerni-gan denied involvement in those, or any other, bank robberies and her score of 17 indicates a truthful response (anything above a 6 is considered truthful). Jernigan does not challenge the district court’s exclusion of this evidence, and we do not consider it in reaching our decision.
. It is worth noting that Rodriguez-Gallegos was never included in any of the photo-spreads shown to the eyewitnesses. Thus, there was never an opportunity for the witnesses to compare Jernigan and Rodriguez-Gallegos.