OPINION

NGUYEN, Circuit Judge:

Igor Zavalin appeals the district court's judgment affirming the Social Security Commissioner's denial of Supplemental Security Income disability benefits. The administrative law judge ("ALJ") found that Zavalin retains the residual functional capacity to perform simple, routine, or repetitive tasks. The ALJ further concluded that Zavalin is not disabled because he is still able to perform two occupations, namely, cashier and surveillance system monitor. Both of these occupations require the ability to perform Level 3 Reasoning on the Department of Labor's General Education Development scale, which is defined as the ability to follow written, oral, or diagrammatic instructions and to deal with problems involving several variables from a standardized situation. We hold that there is an apparent conflict between Zavalin's limitation to simple, routine, or repetitive tasks, on the one hand, and the demands of Level 3 Reasoning, on the other hand. This conflict must be reconciled by the ALJ. Because the ALJ failed to do so, we remand for further proceedings.

BACKGROUND

Zavalin, who was born in Russia, has suffered from severe impairments since childhood. His diagnoses include cerebral palsy, a learning disorder, and a speech impairment that causes him to speak in a halting manner. He also has a history of a fracture in his right knee and atrophy of the right leg, which causes balance problems.

When Zavalin was 13 years old, he moved with his family to the United States, and he began receiving Supplemental Security Income ("SSI") disability benefits that same year. Zavalin attended public schools and had an individualized education program consisting of both special education and mainstream classes with accommodations for his impairments, such as extra time so that he could work at his own pace. He did well, and eventually graduated from high school with a modified diploma in 2010.

After Zavalin turned 18 in December 2008, the Social Security Administration ("SSA") conducted a redetermination of his eligibility for benefits under the rules for determining disability for adults. Zavalin requested review before an ALJ after the SSA administratively determined that he was no longer disabled. An ALJ held a hearing on September 17, 2010, at which she received testimony from witnesses, including a vocational expert who testified about potential occupations for Zavalin.

The ALJ asked the vocational expert an extended hypothetical question in which she described a person with Zavalin's limitations, including the limitation that he "can do simple jobs," and then inquired whether there are jobs in the national economy that such a person can do. The expert opined that a person with such limitations can perform two representative occupations defined by the Department of Labor's Dictionary of Occupational Titles (the "DOT"): cashier and surveillance system monitor (an employee who monitors video surveillance footage). The DOT sets forth job requirements for both positions, including the necessary reasoning ability, which the DOT measures on a six-level scale. Both cashier and surveillance system monitor require Level 3 Reasoning, which is defined as the ability to deal with problems involving several concrete variables and apply commonsense understanding to carry out instructions furnished in written, oral, or diagrammatic form. However, the ALJ did not ask the vocational expert to explain how a person who can only "do simple jobs" because of impairments, including a learning disorder and cerebral palsy, could meet Level 3 Reasoning's requirements.

Following the hearing, the ALJ issued a written decision on October 28, 2010. The ALJ's analysis followed the well-established five-step sequential process for Social Security and SSI disability determinations. The ALJ skipped Step One, which asks whether Zavalin is presently working, because it is not relevant in age-18 disability redeterminations. At Step Two, the ALJ found that Zavalin suffers from several severe impairments: cerebral palsy, a learning disorder, a history of right knee fracture, and atrophy of the right leg. At Step Three, the ALJ found that Zavalin does not have an impairment listed in SSA regulations. Step Four, which determines whether Zavalin could return to a previous occupation, did not apply.

At Step Five — the only step at issue on appeal — the ALJ analyzed whether Zavalin has the capacity to work notwithstanding his severe impairments. The ALJ first assessed Zavalin's capability in light of his impairments. Based on medical reports from doctors and health care professionals who had evaluated Zavalin over the years, as well as Zavalin's testimony, the ALJ concluded that his ability is limited to "simple, routine tasks" and "simple, repetitive tasks." The ALJ also found that he can use his arms with some limitations, has balance problems, cannot climb ladders or ropes, and has a speech impairment. The ALJ then determined whether there are jobs in the national economy that Zavalin can perform in light of his limitations, age, education, and lack of work experience. To make this determination, she relied on the vocational expert's testimony that a person with Zavalin's limitations can perform cashier and surveillance system monitor work, and the DOT's job descriptions and requirements for these occupations. The ALJ concluded that Zavalin is not disabled because he is capable of both identified occupations. In her decision, the ALJ did not explain whether Zavalin possessed the reasoning ability required to perform these occupations, given his residual functional capacity of only simple, routine, or repetitive work.

After the SSA Appeals Council denied his request for review of the ALJ's decision, Zavalin sought judicial review in the district court. The district court summarily affirmed. This appeal followed.

JURISDICTION AND STANDARD OF REVIEW

We have jurisdiction under 28 U.S.C. § 1291. We review de novo the district court's decision. Molina v. Astrue, 674 F.3d 1104, 1110 (9th Cir.2012). We review the Commissioner's disability decision "to determine if it is supported by substantial evidence in the record as a whole and whether it is based on proper legal standards." Nyman v. Heckler, 779 F.2d 528, 530 (9th Cir.1986). Even when an ALJ commits an error of law, we must affirm if the error is harmless. Molina, 674 F.3d at 1111.

DISCUSSION

Zavalin argues that at Step Five, the ALJ failed to reconcile an apparent conflict between his residual functional capacity and the reasoning requirements of the jobs identified by the ALJ. We agree.

I

A

We begin with the legal framework for Step Five.1 At this step, the Commissioner has the burden "to identify specific jobs existing in substantial numbers in the national economy that [a] claimant can perform despite [his] identified limitations." Johnson v. Shalala, 60 F.3d 1428, 1432 (9th Cir.1995); see also 20 C.F.R. § 416.920(g). The ALJ first assesses a claimant's "residual functional capacity," defined as the most that a claimant can do despite "physical and mental limitations" caused by his impairments and related symptoms. 20 C.F.R. § 416.945(a)(1). The ALJ then considers potential occupations that the claimant may be able to perform. See 20 C.F.R. § 416.966. In making this determination, the ALJ relies on the DOT, which is the SSA's "primary source of reliable job information" regarding jobs that exist in the national economy. Terry v. Sullivan, 903 F.2d 1273, 1276 (9th Cir.1990); see also 20 C.F.R. §§ 416.969, 416.966(d)(1). The DOT describes the requirements for each listed occupation, including the necessary General Educational Development ("GED") levels; that is, "aspects of education (formal and informal)... required of the worker for satisfactory job performance." DOT, App. C, 1991 WL 688702 (4th ed.1991). The GED levels includes the reasoning ability required to perform the job, ranging from Level 1 (which requires the least reasoning ability) to Level 6 (which requires the most). See id.

In addition to the DOT, the ALJ relies on the testimony of vocational experts who testify about specific occupations that a claimant can perform in light of his residual functional capacity. 20 C.F.R. § 416.966(e); Valentine v. Comm'r Soc. Sec. Admin., 574 F.3d 685, 689 (9th Cir. 2009). Finally, to conclude the Step Five analysis, the ALJ determines "whether, given the claimant's [residual functional capacity], age, education, and work experience, he actually can find some work in the national economy." Valentine, 574 F.3d at 689; see also 20 C.F.R. § 416.920(g).

When there is an apparent conflict between the vocational expert's testimony and the DOT — for example, expert testimony that a claimant can perform an occupation involving DOT requirements that appear more than the claimant can handle — the ALJ is required to reconcile the inconsistency. Massachi v. Astrue, 486 F.3d 1149, 1153-54 (9th Cir.2007). The ALJ must ask the expert to explain the conflict and "then determine whether the vocational expert's explanation for the conflict is reasonable" before relying on the expert's testimony to reach a disability determination. Id.; see also Social Security Ruling 00-4P, 2000 WL 1898704, at *2 (Dec. 4, 2000). The ALJ's failure to resolve an apparent inconsistency may leave us with a gap in the record that precludes us from determining whether the ALJ's decision is supported by substantial evidence. See Massachi, 486 F.3d at 1154 (stating that "we cannot determine whether the ALJ properly relied on [the vocational expert's] testimony" due to unresolved occupational evidence).

B

We now turn to Zavalin's claim that the ALJ erred at Step Five. Zavalin does not contest the ALJ's finding that his residual functional capacity limits him to simple, routine, or repetitive work. He argues, however, that there is an inherent inconsistency between his limitation to simple, routine tasks, and the requirements of Level 3 Reasoning.

We have not in our circuit addressed this question, and it is one on which our sister circuits are split. For example, in Hackett v. Barnhart, the Tenth Circuit held that a claimant's limitation to "simple and routine work tasks" is "inconsistent with the demands of level-three reasoning" because the plaintiff's residual functional capacity was more consistent with Level 2 than Level 3 Reasoning. 395 F.3d 1168, 1176 (10th Cir.2005). In contrast, with little analysis, the Seventh and Eighth Circuits rejected the claim that a conflict exists. See Terry v. Astrue, 580 F.3d 471, 478 (7th Cir.2009); Renfrow v. Astrue, 496 F.3d 918, 921 (8th Cir.2007). District courts in our circuit that have confronted this issue are also divided. See, e.g., Adams v. Astrue, No. C 10-2008 DMR, 2011 WL 1833015, at *4 (N.D.Cal. May 13, 2011) (stating that "there appears to be a conflict between" a limitation to "simple, repetitive tasks" and Level 3 Reasoning); Wentz v. Astrue, CIV. No. 08-661-PK, 2009 WL 3734104, at *13-15 (D.Or. Nov. 4, 2009) (finding "no apparent conflict" between a limitation to "simple, routine, repetitive work" and Level 3).

Today, we join the Tenth Circuit and hold that there is an apparent conflict between the residual functional capacity to perform simple, repetitive tasks, and the demands of Level 3 Reasoning. We find the conflict to be plain when we consider, side-by-side, the definitions of Level 2 and Level 3 Reasoning:

DOT, App. C, 1991 WL 688702. Level 2 Reasoning — applying common sense to carry out detailed but uncomplicated instructions and dealing with problems involving a few variables — seems at least as consistent with Zavalin's limitation as Level 3 Reasoning, if not more so. See Hackett, 395 F.3d at 1176 (noting that Level 2 "appears more consistent" than Level 3 for a claimant limited to simple, routine tasks). Further, Zavalin's limitation to simple, routine tasks is at odds with Level 3's requirements because "it may be difficult for a person limited to simple, repetitive tasks to follow instructions in `diagrammatic form' as such instructions can be abstract." Adams, 2011 WL 1833015, at *4.

The Commissioner argues that the DOT's reasoning levels correspond only to a person's level of education and, therefore, Zavalin is presumptively capable of Level 3 Reasoning because he completed high school. We are unpersuaded. Contrary to the Commissioner's claim, the DOT specifically defines GED reasoning levels to include "informal" as well as "formal" education that is required for "satisfactory job performance." DOT, App. C, 1991 WL 688702. Thus, there is no rigid correlation between reasoning levels and the amount of education that a claimant has completed. While Zavalin's educational background is relevant, the DOT's reasoning levels clearly correspond to the claimant's ability because they assess whether a person can "apply" increasingly difficult principles of rational thought and "deal" with increasingly complicated problems. Id. For example, Level 1 requires the ability to "carry out simple one- or two-step instructions," whereas Level 6 requires the application of "principles of logical or scientific thinking to a wide range of intellectual and practical problems." Id. Moreover, the Commissioner's reliance on Zavalin's completion of high school ignores the fact that he was in special education classes, and succeeded in regular classes only with special accommodations that allowed him to work at his own pace. Further, while he graduated, Zavalin received a modified diploma, which is conferred on "students who have demonstrated the inability to meet the full set of academic content standards for a high school diploma even with reasonable modifications and accommodations." Or. Admin. R. 581-022-1134(2).

In sum, because the ALJ failed to recognize an inconsistency, she did not ask the expert to explain why a person with Zavalin's limitation could nevertheless meet the demands of Level 3 Reasoning. We conclude that the ALJ erred in failing to reconcile this apparent conflict.

II

Lastly, we address whether the ALJ's error is harmless. See Molina, 674 F.3d at 1111 (stating that "we may not reverse an ALJ's decision on account of an error that is harmless"). In making this determination, we are "constrained to review the reasons the ALJ asserts" and "cannot affirm the decision of an agency on a ground that the agency did not invoke in making its decision." Stout v. Comm'r, Soc. Sec. Admin., 454 F.3d 1050, 1054 (9th Cir.2006) (citations and internal quotation marks omitted). The parties dispute whether Zavalin's abilities, such as his success in school, demonstrate that he is capable of working as a cashier and surveillance system monitor, such that the ALJ's error in failing to reconcile any inconsistency would be harmless.

We first consider the DOT's descriptions of these jobs. A surveillance system monitor is a security employee responsible for monitoring security cameras in public transportation terminals. The employee must be able to "detect crimes or disturbances," "notify authorities" when required, while continuing to "maintain surveillance of [the] location where [the] incident is developing." DOT 379.367-010, 1991 WL 673244. A cashier must be able to compute bills, itemized lists, and tickets showing the amount due, reconcile the cash register's tape against cash on hand, and give cash refunds and issue credit memorandums to customers for returned merchandise. DOT 211.462-010, 1991 WL 671840.

The Commissioner relies heavily on Zavalin's success in math, which was one of his strengths while in school, and his use of computers and video games, to argue that he is capable of performing the identified jobs. However, while the record shows that Zavalin did well in math, it was in the context of a special education program. As for his use of the computer and video games, the ALJ did not rely on this evidence, and we cannot do so now to find the error harmless. Stout, 454 F.3d at 1054. Even if we were to consider this evidence, we are not persuaded that it shows Zavalin possesses the requisite reasoning ability because there is no indication of the extent or manner of his computer use, or the complexity of the video games.

Certainly, Zavalin's educational successes are not irrelevant and we agree that he appears capable of performing some of the duties required in these occupations. However, other tasks seem neither simple nor routine. As a cashier, reconciling the cash on hand against the cash register's tape and issuing credit memorandums to customers could contain situational variables that may not be simple or repetitive. Similarly, a surveillance system monitor may be called upon to use discretion and judgment in rapidly evolving scenarios, including deciding when a situation requires the authorities to be notified, all while continuing to maintain surveillance.

On this mixed record, "we cannot determine whether substantial evidence supports the ALJ's step-five finding that [Zavalin] could perform [the] work." Massachi, 486 F.3d at 1154. We therefore conclude that the ALJ's failure to reconcile the apparent conflict is not harmless.

CONCLUSION

We reverse and remand to the district court so that it may remand to the SSA for further proceedings consistent with this opinion. See Moisa v. Barnhart, 367 F.3d 882, 886 (9th Cir.2004). We do not reach the parties' remaining arguments.

REVERSED AND REMANDED.