Pick your poison

How hard is it to refund the tariffs?

On February 20, 2026, the Supreme Court ruled, correctly in my view, that the International Emergency Economic Powers Act (IEEPA) does not authorize tariffs.

While the case was pending, the government collected billions of dollars in IEEPA tariffs. Because the government lost, it must now refund all that money, with interest.

The jury is still out on how eagerly the government will cooperate with the refund process. Early signs, however, are not promising. Among other things, after the Court of International Trade issued an order directing the government to immediately stop liquidating entries subject to the illegal tariffs, the government asserted it couldn’t comply with the order.

In the Paleolithic era of, say, 2022, we would have had no choice but to wince and trust the government. In 2026, we can ask AI whether the government’s assertions are correct. In this post, I’ll first walk through what’s happened so far in the tariff refund litigation and then address whether AI has any role to play.

It will already be slow, so might as well make it slower

Prior to the Supreme Court’s decision, the lower courts with jurisdiction over tariff disputes—the Court of International Trade and the Federal Circuit—both ruled that the tariffs were illegal.¹ Nonetheless, the government continued collecting the tariffs until the day the Supreme Court ruled.

The government was able to continue collecting the tariffs because the Federal Circuit stayed the Court of International Trade’s injunction against the tariffs. Even after the Federal Circuit ruled against the tariffs, the Federal Circuit stayed its own mandate until the Supreme Court ruled.

Why? The government successfully argued that an immediate injunction against the tariffs was unnecessary because if the government ultimately lost in the Supreme Court, the money could be refunded with interest. After the Court of International Trade struck down the tariffs, here is what the government told the Court of International Trade:

For any plaintiff who is an importer, even if a stay is entered and we do not prevail on appeal, plaintiffs will assuredly receive payment on their refund with interest. “[T]here is virtually no risk” to any importer that they “would not be made whole” should they prevail on appeal. See Sunpreme Inc. v. United States, 2017 WL 65421, at *5 (Ct. Int’l Trade Jan. 5, 2017). The most “harm” that could incur would be a delay in collecting on deposits. This harm is, by definition, not irreparable. See Hughes Network Sys., Inc. v. InterDigital Commc’ns Corp., 17 F.3d 691, 694 (4th Cir. 1994). Plaintiffs will not lose their entitlement to refund, plus interest, if the judgment is stayed, and they are guaranteed payment by defendants should the Court’s decision be upheld. And defendants do not oppose the reliquidation of any entries of goods subject to IEEPA duties paid by plaintiffs that are ultimately found to be unlawful after appeal.

The government told the Federal Circuit the same thing: “If tariffs imposed on plaintiffs during these appeals are ultimately held unlawful, then the government will issue refunds to plaintiffs, including any post-judgment interest that accrues.”

On February 20, Liberation Day II arrived. A few days later, the victorious plaintiffs asked the Federal Circuit to issue its mandate immediately so they could obtain the refunds that the government had guaranteed were forthcoming.

The government’s opposition to that motion did not bode well for the refund process.

The government first argued that “ordinary process” required 32 extra days of delay:

In a case on review from a federal court, the Supreme Court sends down its judgment 32 days after entry of judgment (unless the Court or a Justice orders otherwise or the parties stipulate that it be sooner), or forthwith upon the denial of rehearing. Sup. Ct. R. 45.3; see Sup. Ct. R. 45.2. The court of appeals ordinarily should not act on a case until the judgment is sent down. Accordingly, when (as here) a court of appeals has stayed both the lower court’s judgment and its own mandate pending certiorari, the court of appeals ordinarily should not dissolve the stay or issue the mandate until the judgment of the Supreme Court is sent down. Plaintiffs have not stated any reason for this Court to depart from the ordinary process by outpacing the Supreme Court.

To begin with a pedantic point that non-lawyers can skip, this paragraph misunderstands Supreme Court procedure. It isn’t the case that “the court of appeals ordinarily should not act on a case until the judgment is sent down.” The default rule is that courts of appeals do act on cases while they’re pending in the Supreme Court. Even a Supreme Court order granting certiorari doesn’t stay lower-court proceedings unless either a lower court or the Supreme Court expressly orders otherwise. True enough, in this case, the Federal Circuit stayed its own mandate. But that stay, by its terms, would remain in place pending “a judgment of the Supreme Court.” The “judgment of the Supreme Court” was entered on February 20, 2026.² The government’s proposed rule—that “when (as here) a court of appeals has stayed both the lower court’s judgment and its own mandate pending certiorari, the court of appeals ordinarily should not dissolve the stay or issue the mandate until the judgment of the Supreme Court is sent down”—does not exist.

The government’s arguments for delay were unpersuasive to the point of being virtually non-existent. For instance, the government argued that “a plaintiff’s bare desire to be paid immediately is not a basis to demand this Court comply with his every whim.” Even setting aside the tendentious prose, this is not an actual argument for delay. The government also argued that delay was warranted because the “coming process will take time,” citing an example of a case where “refunds took seven years to fully be provided.” If the DMV married the cable company, this argument would be their baby. “Refunds will take years, so let’s add 32 more days for no reason!”

Doubling down, the government said: “If anything, the Court should withhold issuance of the mandate for 90 days to allow the political branches an opportunity to consider options.” The “political branches” can always consider “options,” but that is not a persuasive basis for preventing importers from getting refunds to which they are entitled. And the government’s pre-Supreme-Court-decision argument that an injunction is unnecessary because importers are “guaranteed payment by defendants should the Court’s decision be upheld” is in some tension with the government’s post-Supreme-Court-decision argument that importers should be denied refunds for months so that the famously productive and collegial Congress can consider “options.”

In summary, the government’s filing gave no justification for additional delay. Its apparent purpose was to foreshadow the government’s intention to grind importers to bits in a bureaucratic mill.

Runaway train

Not surprisingly, the Federal Circuit immediately issued the mandate as the plaintiffs requested. Back in the Court of International Trade, the litigation got off to a bumpy start.

On March 3, the Court of International Trade issued an order asking the government a series of questions, to which the government responded on March 4. Here are some of the questions and answers.

(Note: “Liquidating” is customs-ese for “finalizing.” At the time of import, the importer files an entry and deposits estimated duties. When Customs and Border Protection (CBP) “liquidates” an entry, it finalizes the amount owed. If the importer overpaid, it gets a refund; if it underpaid, CBP bills them for the difference. Liquidation generally occurs 314 days after entry into the United States.)

Q: Is U.S. Customs and Border Protection (“Customs”) currently liquidating entries with duties imposed pursuant to IEEPA?

A: Yes, for those entries where IEEPA duties were declared at entry, CBP is liquidating those entries with duties imposed pursuant to IEEPA.

Q: Has Customs issued refunds of IEEPA duties as the result of liquidations that have taken place since the Supreme Court’s decision in Learning Resources, Inc. v. Trump, 2026 WL 477534 (U.S. Feb. 20, 2026)?

Response: No, CBP has not issued refunds of IEEPA duties as the result of liquidations that have taken place since the Supreme Court’s decision.

So it’s business as usual at CBP. And then there’s this:

Q: Is it Customs’ position that, with respect to entries that are liquidated without IEEPA duties, the importers of record will be sent a refund in an amount equal to the IEEPA duties plus interest?

A: There are several factors which inform the answer to this question, including but not limited to entry type. In accordance with applicable law, any validated refund of IEEPA duties would include interest. Regardless of entry type and liquidation cycle, CBP still requires a review period to ensure no violation of other Customs laws and no other duties, taxes, or fees are owed (e.g., anti-dumping, Sec. 301 duties, Sec. 232, etc.)

The government told the court a year ago that importers were “guaranteed” to get refunds, which sounds different from “There are several factors which inform the answer to this question, including but not limited to entry type.” Perhaps the government is right that refunds will sometimes be unwarranted. But saying “There are several factors which inform the answer to this question, including but not limited to entry type” leaves the government unlimited discretion to deny a refund.

The Court of International Trade evidently had the same reaction to this declaration that I did, because later that day, it issued an order³ saying that “with respect to any and all unliquidated entries that were entered subject to the IEEPA duties, U.S. Customs and Border Protection is hereby directed to liquidate those entries without regard to the IEEPA duties. Any liquidated entries for which liquidation is not final shall be reliquidated without regard to IEEPA duties.”

And then CBP came back with a declaration stating: “CBP is not able to comply with the Court of International Trade’s order.”

As to unliquidated entries, CBP said that the liquidation process is an unstoppable runaway train. E.g.:

Although CBP stopped accepting IEEPA duty deposits for informal entries on February 24, 2026, there are around 4 million informal entries that were filed before February 24, 2026, that have not yet liquidated. Many of these entries will automatically liquidate on March 16, 2026, when importers pay their March Periodic Monthly Statement. CBP does not have a process to prevent the liquidation of informal entries.

And as to refunds:

In light of the Court’s March 5, 2026 amended order, CBP is now facing an unprecedented volume of refunds. Its existing administrative procedures and technology are not well suited to a task of this scale and will require manual work that will prevent personnel from fully carrying out the agency’s trade enforcement mission. Personnel would be redirected from responsibilities that serve to mitigate imminent threats to national security and economic security. Those activities include, but are not limited to, detecting and disrupting transshipped goods, protecting against free trade agreement abuse, and detecting AD/CVD evasion all of which are illicit actions that threaten U.S. domestic industry.

Letting the government retain illegal tariffs is surely a small price to pay for “protecting against free trade agreement abuse”! On the plus side, the declaration says that CBP will “develop and implement new ACE functionality that will streamline and consolidate refunds,” and have “this new ACE functionality ready for use in 45 days.” An update submitted on March 12 suggests that the government has made substantial progress on this new functionality. Let’s hope.

It would have been nice if the government had been open about these challenges before the Supreme Court’s decision. The government successfully persuaded the Federal Circuit to stay its mandate on the theory that if the government lost, the importers would get refunds. The government did not mention that (1) it would push for a 90-day stay even after the Supreme Court’s decision; (2) even after that stay was denied, it would be forced to keep liquidating entries with IEEPA tariffs because its possibly FORTRAN-driven computer system made it impossible to stop the tariff snowball from rolling downhill; and (3) technology for processing refunds doesn’t currently exist.

There is nothing we can do about this now. I agree with Elsa that the past is in the past. But the government has announced a new set of so-called Section 122 tariffs. In my opinion, these tariffs are also illegal and the government will have to refund them too. The government will no doubt argue that any injunction against Illegal Tariffs II should be stayed pending Federal Circuit and Supreme Court review, as with Illegal Tariffs I. The stay should be denied. The government might say that a stay is unnecessary because the importers are guaranteed to get refunds, but fool me twice, shame on me. And you know what will happen: if the government ultimately loses on Illegal Tariffs II, it is going to say that it is too hard to issue refunds because it is too busy dealing with refunds for Illegal Tariffs I.

Strategic exaggeration

After the government submitted its declaration saying it couldn’t comply with the order, the Court of International Trade immediately suspended its order and instead directed the government to submit periodic progress reports on its refund software development project. It’s hard to blame the court for doing this. The court can’t order CBP to do the impossible. The declarant—a high-level official at CBP—is in the best position to know what’s possible, so neither the court nor the plaintiffs are in a position to push back. One might criticize the government for waiting until after the Supreme Court’s decision to begin work on a solution for issuing refunds, but that’s the way it goes.

Ordinarily, when a party submits a notice in litigation that it can’t or won’t comply with an order, the court doesn’t just suspend the order. It gives the opposing party the opportunity to respond, holds a hearing, and makes findings. In this case, however, the court must have understood that there was little the plaintiffs could do to undermine the credibility of the government’s declaration, particularly given that the government risked violating the order within days. After all, the plaintiffs’ lawyers didn’t have the personal knowledge or the technical ability to contest the declaration. Maybe they could try to find an expert, but it’s not clear there’s an expert out there who has the requisite technical, legal, and practical expertise to offer an informed opinion on how quickly CBP could develop refund software. Even if there was, finding such an expert would take time—and the compliance deadline was imminent. So there was no way, on a short time scale, for the plaintiffs to push back on CBP’s argument that compliance with the order was impossible.

That might have been true a few years ago. But now, you could just ask AI.

Let me refine the question we’re asking. The Court of International Trade directed that “with respect to any and all unliquidated entries that were entered subject to the IEEPA duties, U.S. Customs and Border Protection is hereby directed to liquidate those entries without regard to the IEEPA duties.” Although CBP’s declaration is a little bit slippery, I understand CBP to be departing from this order in two respects. First, if I’m understanding the declaration correctly, nothing is going to happen until the new software comes online. Before that occurs, CBP is going to keep liquidating entries as though the Supreme Court’s decision didn’t exist. (Remember that liquidation generally occurs 314 days after entry, so CBP will be liquidating entries that occurred prior to the Supreme Court’s decision for several more months.)

Second and more importantly, even after the new software comes online, importers will have to affirmatively submit a claim into this refund system to trigger the recalculation of duties. Importers lacking the time or sophistication or gumption to proactively seek a refund won’t get their money back—even for as-yet-unliquidated entries. If an importer doesn’t submit a claim, CBP will continue liquidating entries with IEEPA duties, in violation of the court’s now-suspended order that it not do so.⁴ It’s possible I’m misunderstanding this aspect of the declaration (indeed, I hope I am). But if I’m understanding it correctly, then depending on how many importers fail to submit claims, the government might retain billions of dollars worth of IEEPA tariffs.

So my question is, would it be possible for CBP to come closer to following the court’s order than it promises to do? Could CBP reasonably develop a system that deducted IEEPA tariffs prior to liquidation regardless of whether importers submitted refund claims? And if so, could it do so in less than 45 days?

Claude and ChatGPT both say that the answer is yes on both counts. They are both skeptical of the government’s declaration: Claude refers to it as an exercise in “strategic exaggeration.” Both LLMs acknowledge that, as the declaration indicates, immediate compliance—that is, cessation of liquidation within days of the order—may be unreasonably difficult. Thus, both LLMs agree that the court did the right thing in suspending the order to the extent it demanded immediate compliance. But they also insist that the government could, with reasonable effort, deduct IEEPA tariffs from most entries prior to liquidation even without affirmative action by the importer. And they think that putting a preliminary system online shouldn’t take 45 days.

Here’s Claude’s one-shot effort at writing a concise amicus brief explaining why the government could come closer to complying with the order than it is promising.⁵

Brief In Support Of Compliance With The Court's Order Regarding Unliquidated Ieepa Entries

255KB ∙ PDF file

Download

Download

6 > 3

Let’s take a step back. A high-level government official is declaring that it’s impossible for CBP to comply with the court’s order. AI partially agrees, but also insists that CBP can come closer to complying with the court’s order than it is promising. Who should we trust?

There are good reasons to trust the government’s declarant:

1. He has access to non-public information regarding CBP’s systems, whereas AI knows nothing about CBP’s code base and acquired its knowledge by reading Reddit comment threads.

2. It’s his full-time job to be an expert on this exotic topic that most Reddit commenters have never heard of.

3. He’s writing a declaration under oath, which, at least in principle, should increase our faith in the declaration’s accuracy.

But there are also good reasons not to trust him:

1. The declaration comes days after the government submitted a brief to the Federal Circuit arguing for delay on the basis that because refunds will take a long time, they may as well take a longer time. This filing casts some doubt on the government’s commitment to ensuring that the refund process moves briskly.

2. Likewise, it appears from the declaration that the government had no plan for refunds in the event it lost in the Supreme Court and did nothing on that front in the aftermath of the Supreme Court’s decision until summoned by the court, again casting doubt on the government’s exuberance about speeding the refund process along.

3. Public reporting indicates that the Administration is trying to find ways to keep the tariff money. Realistically, it is this declarant’s job, implicitly or perhaps explicitly, to assist the government in keeping as much tariff money as possible while not making the courts too angry.

And AI has its virtues:

1. Developing a scheme for processing refunds requires a mix of legal, practical, and technical knowledge. No one person is likely to have expertise in all of these fields, but AI does. CBP’s employees, in the aggregate, likely possess all these different types of knowledge, but it takes time to pull all the information together and CBP might not be motivated to do so, whereas AI can do so instantly. (Needless to say, this argument would be stronger if AI had access to CBP’s code base. Without that access, AI is left to making an educated guess on what it looks like, based on the descriptions in CBP’s declaration and its general knowledge of comparable systems. Even based on that educated guess, though, AI can devise a general strategy for what a refund scheme would look like and offer a loose estimate on how long the project would take. Humans do this all the time.)

2. In all likelihood, the most efficient way to craft a software solution is to use AI. If I had to guess, the software systems at CBP are well-documented but crufty, and developing a software solution will not require great creativity but will require tedious attention to obscure edge cases. This seems like the type of software project that is right up AI’s alley. This is especially true given the unlikelihood of the Administration putting its best and brightest coders on the mission of helping the government give away money it wants to keep. AI may have an advantage in assessing how long it will take for AI-assisted software design to be completed. (Again, I realize this argument would be stronger if AI had access to the code that needs to be modified.)

3. You will all yell at me if I say AI is “unbiased.” But AI lacks certain types of biases that might affect an Administration official bent on keeping the tariff money by any means necessary.

So that’s 6 reasons to trust AI and 3 reasons to trust the government. AI informs me that 6 is greater than 3 and that AI is therefore more trustworthy.

If you don’t accept that, surely there is some world-state in which an LLM’s conclusion would be more trustworthy than a human declarant’s conclusion—say, if you gave the LLM some discovery about the software currently in use at CBP and it continued to insist that a software solution that deducted IEEPA tariffs automatically would be straightforward or even vibe-coded it up.

AI vs. AI

Suppose we lived in a world in which a court was willing to consider the views of AI in deciding whether a software solution that deducted IEEPA tariffs automatically was feasible, and if so, how quickly it could be implemented. How might this litigation unfold?

Let’s start with the plaintiffs. The lawyers would have little choice but to use AI. We can debate all day whether AI is smarter than humans or what that even means, but AI has obvious advantages at this task. It requires the type of technical knowledge that most humans—certainly most human lawyers—lack. A human lawyer certainly could not write an amicus brief like the one above within 20 seconds, and would have trouble writing it within 20 days. And if we ratchet up the degree of difficulty—say, if the plaintiffs got some source code in discovery and sought to prove that compliance with the order is feasible based on an analysis of that source code—a human lawyer could never match AI.

Also, the plaintiffs’ lawyers are going to know that the court won’t trust them to pick apart the technical statements in CBP’s declaration. The lawyers were history majors after all. The lawyers can try submitting an expert declaration, but if the litigation is moving quickly and the expert submits the declaration within days, the court will think the expert is jumping the gun and didn’t have sufficient time to consider the issues.

The court might trust an AI-assisted filing. But it might also recognize that AI is easily manipulable - a human lawyer can prompt it to write a persuasive brief on either side of the issue.

The lawyers have another option: wash their hands of human involvement. The lawyers can submit neutral-sounding prompts to AI—“give me your best, real-world guess as to how long it will take to implement a refund solution that deducts IEEPA tariffs automatically”—and disclose the prompts to the court. If, even after unbiased prompts, the LLM rejects the government’s position—better yet, if multiple LLMs all disagree with the government upon such prompting—the lawyers can characterize the LLMs’ output as powerful evidence that the government’s declaration isn’t credible.

This strategy is not as outlandish as it sounds. As AI improves, courts will be more persuaded by LLMs’ predictive judgments than those of humans, at least in certain domains. Courts will be skittish about trusting LLMs’ recommendations on pure issues of law. But on technical, fact-heavy questions—such as whether it is reasonably possible for software with particular functionality to be created over a particular time period—it will seem entirely rational for a court to trust an LLM’s assessment. Perhaps that is rational even now.

OK, now the ball is in the government’s court. Suppose the government is served with an AI-assisted filing establishing that partial compliance with the court order is possible. The government can prepare a declaration explaining point-by-point why the LLM’s statements are false. But the plaintiffs will respond that the government’s declarant is both less technically sophisticated and less trustworthy than an LLM. And in a sufficiently complex case, a government official might not have an easy time parsing the LLM’s objections.

The government will have little choice but to fight AI with AI. The government can offer prompts of its own that establish that the plaintiffs’ proposed technical solution is illusory. Like the plaintiffs’ lawyers, the government can wash its hands of bias accusations by disclosing its prompts to the court and establishing that even an “unbiased” LLM agrees that compliance is impossible.

So now we have two AI-generated documents written by lawyers who insist the documents reflect the unvarnished views of AI. What is the judge to do?

The judge won’t be able to choose intelligently between these two complex technical submissions, particularly on a short timeline. The judge can’t hold a hearing in which he evaluates the credibility of the expert witnesses, because there are no expert witnesses. The judge will have little choice but to use AI.

So if everyone fulfills their role—if the lawyers try their best to be both persuasive and credible and the judge tries to resolve the dispute as accurately as possible—then we’ll have AI deciding between two AI-written submissions, with the human lawyers claiming that their submissions are credible precisely because humans were not involved. So much for our legal system.

On the other hand, the current situation is not much better. The government claims it has no choice but to keep billions of dollars in illegally-exacted tariffs, and there’s nothing anyone can do about it.

Pick your poison!

1. The D.C. District Court also struck down the tariffs, but the Supreme Court ruled that it lacked jurisdiction.

2. The Supreme Court would send its judgment to the lower court 32 days later, but the judgment was entered on February 20.

3. The court slightly amended the order the next day.

4. This understanding is consistent with CBP’s March 12 declaration, which reports that there will be a “CAPE Claim Portal … for importers and brokers to submit IEEPA refund requests” and that the “CAPE Mass Processing component will automatically remove any applicable IEEPA HTS numbers from the entry summaries submitted to and validated by the CAPE Claim Portal component.”

5. I didn’t spot any hallucinations, but I’m not making any guarantees.