Introduction

This article arose from a legal consultation. The author has long practised in customs and export control and in smuggling defence. Recently, an enterprise consulted about the classification of equipment. As usual, while beginning the research, the author also searched publicly available enforcement positions, and during this noticed an article on a certain public account listing three customs administrative penalty decisions involving artificial intelligence and commodity classification.

However, when verifying these decisions, the author found that the full text of two of them contained no mention of artificial intelligence or similar expressions. The author therefore adopted a not-very-clever method: searching the administrative penalty columns on the official websites of major customs districts nationwide one by one, and reading through the decisions made public in recent years, which took about seven hours in total. The result was unexpected: among all the public decisions retrieved, only one genuinely recorded that the party had used artificial intelligence software to determine the commodity code and declared on that basis, namely the administrative penalty decision Wanding Guan Ji Wei Zi [2026] No. 17 recently published by Kunming Customs [1]. Although it is currently an isolated case in customs public information, the core of the iceberg theory is that what is visible is only the surface, and what is invisible is the essence. The weight of this decision is therefore somewhat different.

The facts of the case are summarised as follows. In November 2025, You purchased 3 lathes intending to export them to Myanmar. Being unfamiliar with export procedures, he entrusted Luo to handle customs declaration as his agent. Luo was likewise unsure whether the goods were export-controlled items, and so sought an answer through artificial intelligence software. The software replied that the goods were not export-controlled items. Luo accepted this without verification and, following the commodity code given by the software, declared the 3 lathes as 8461909000.

On 12 February 2026, officers of Wanding Customs, when reviewing this declaration, found that the commodity code corresponding to the machine tools at issue should be 8458110010, and that the export of goods under that code requires a dual-use items and technologies export licence. A case of false declaration thus arose. The case involves four subjects: besides You and Luo, there are two companies, namely Taiyuan Maixiangfeng Trading Co., Ltd., the domestic consignor on the declaration, and Jinan Fuhuali Customs Declaration Co., Ltd., the customs declaration unit. Customs ascertained that the two companies had lent their legal person names to You and Luo for consideration and were unaware of the matters relating to the unlawful export. Customs, pursuant to Article 24, paragraph 1 and Article 10, paragraph 3 of the Customs Law, determined that the conduct of the two companies constituted a violation of customs supervision provisions under Article 12 of the Implementing Regulations of the Customs Administrative Penalty, and imposed a fine of 10,500 yuan on each.

The penalty is not heavy. But this decision raises a question rarely seen before in customs enforcement documents: an enterprise, being uncertain about the nature of the goods, turned to artificial intelligence software for an answer and declared on that basis, and ultimately bore legal liability. A question not yet widely recorded often means the risk is not yet widely recognised, and this is the question this article attempts to explore.

I. Within the Customs Law Perspective: The Judgmental Nature of Commodity Classification and the Capability Boundaries of Artificial Intelligence

To understand the weight of this case, it is first necessary to explain what kind of work commodity classification is.

In the eyes of many import and export enterprises, classification is a search task, namely matching the name, material and use of goods one by one against the Import and Export Tariff to find a code. The emergence of artificial intelligence software has reinforced this impression, because it can indeed produce an eight- or even ten-digit code within seconds, accompanied by an explanation that looks quite professional.

But classification is in fact a legal judgment. The Import and Export Tariff contains six General Rules for the Interpretation of the Harmonized System, of which Rule 3 sets out the order of application of the three criteria of "most specific description", "essential character" and "last in numerical order", and Rule 6 requires that subheadings be compared only at the same level. The rules themselves are not abstruse, but their application depends heavily on an understanding of the goods. The same plastic connector, whether classified under heading 3926 as other articles of plastics or under heading 8487 as parts of machinery, depends on whether the part is solely or principally used with a particular machine and whether it already has the character of a complete machine. For the lathes in this case, a difference in any one parameter—model, processing method, control method, positioning accuracy—may change the commodity code.

The facts on which these judgments rest cannot be read from the text of the Tariff; they can only come from technical knowledge of the goods themselves. Artificial intelligence software can recite the provisions of the rules and can list several possible codes, but it cannot answer on the user's behalf a prior question: what should the goods at issue legally be evaluated as? Until that question is resolved, the code it gives is at most a possibility.

The difficulty of classification does not end there. Within the customs enforcement system itself, this judgment also involves a process of unifying positions. When the General Administration of Customs replied to an administrative reconsideration case in 2008, it stated: "Customs pre-classification review, classification review at the declaration stage and classification review in subsequent audits are respectively the responsibility of different customs departments; enforcement officers of different departments, based on different understandings of the goods and different interpretations of the Import and Export Tariff and its notes, may reach inconsistent results in the classification review and determination of the same commodity." [2] That reply mentioned that Customs A had made different classification decisions for the same batch of tool magazines imported by Ruian Company at the document review stage and at the pre-classification stage. The customs system has specifically established classification sub-centres for this purpose, to which the tariff departments of the directly subordinate customs submit classification applications when disputes arise, and which make classification conclusions. That a dedicated institution is established to unify positions on a technical judgment shows the degree of its professionalism.

Once classification goes wrong, the consequences fall into three levels. The lightest is recovery of duties plus late payment surcharges. The heavier is administrative penalty: under Article 15 of the Implementing Regulations of the Customs Administrative Penalty, where declaration items such as the tariff heading are falsely declared and this affects the collection of state duties, a fine of not less than 30 per cent and not more than twice the amount of duties evaded is imposed; this case falls into this category. The heaviest is the crime of smuggling, and the key to distinguishing administrative illegality from crime lies in subjective intent. The publicly released practical trial points of the Shanghai High People's Court state, as published in Shanghai Legal Daily on 18 August 2026: "Strictly distinguish classification misunderstandings from smuggling intent; where a classification divergence is caused by objective reasons such as industry practice, differences in understanding of customs policy, or the technical complexity of classification, and the actor subjectively has no smuggling intent, it should not be determined to be the crime of smuggling." [3]

The other side of this statement also deserves attention. If the evidence in the case shows that the actor had already recognised that the code was doubtful and still continued to declare, he may be found to have smuggling intent. And the fact that "artificial intelligence software was once used for a query" may have evidentiary effect in both scenarios. It can be used to prove that the actor lacked understanding of the nature of the goods, and it can also be used to prove that he had already realised that a doubt existed here. The evidentiary effect often depends on other circumstances of the case, not on the act itself.

The third level—the serious consequence of suspected smuggling crime—deserves separate explanation. If, because of a deviation in a commodity code, a controlled item that legally requires a licence is declared as ordinary goods requiring no licence, the nature of the matter changes. Take this case: goods under 8458110010 require a dual-use items and technologies export licence for export, while those under 8461909000 do not. Declaring the former as the latter has the same effect as exporting controlled items without a licence. Once it is found that there was intent to evade customs supervision, what the actor faces is no longer a fine but the crime of smuggling goods and articles prohibited from import and export by the state under Article 151, paragraph 3 of the Criminal Law; upon conviction, the punishment is fixed-term imprisonment of not more than five years or criminal detention, and/or a fine. Where the circumstances are serious, the punishment is fixed-term imprisonment of not less than five years and a fine.

These consequences are worth listing in detail in this article because whether they occur depends on an extremely difficult boundary point: whether the actor subjectively knew. And proof of subjective state is precisely the link most prone to problems when parties handle matters themselves. At the audit or anti-smuggling investigation stage, the party usually says that he was unsure at the time and therefore asked artificial intelligence. To the party, this sentence is an explanation; to the investigating authority, it may be interpreted as: the actor had already realised that a doubt existed here and still declared without verifying with customs or a professional. The fact of "having once consulted artificial intelligence", in a case file lacking professional organisation, can easily turn from a favourable circumstance into an unfavourable one. In which direction the same fact is explained depends on whether anyone fixed and organised it before the procedure began. This point will be discussed further below.

II. Outside the Customs Law Perspective: Response Deviations of Artificial Intelligence in Other Legal Professional Fields and Their Causes

Errors by artificial intelligence in professional fields are not occasional technical malfunctions. Zhao Yinhao, Chief Judge of the Appeal Review Division of the Beijing No. 2 Intermediate People's Court, pointed out in an analysis in Legal Daily the most fundamental layer: "The essence of large language models is probabilistic prediction rather than logical reasoning. When training data is insufficient, information is ambiguous, or user instructions are too leading, in pursuit of textual coherence and task completion, they tend to create content on their own." [4]

Judge Chen Ying of the Civil Adjudication Division of the Shanghai No. 2 Intermediate People's Court added from the perspective of how questions are posed: "Parties often ask questions based on their own understanding, providing incomplete or even biased information, so the answers given by AI tend to favour the questioner." This observation is repeatedly confirmed in practice. Questioners usually have already formed some expectation when asking, and the case descriptions they provide often reflect that expectation. What artificial intelligence responds to is, rather than an objective legal question, the answer the questioner hopes to have confirmed. [5]

The third link deserves more attention. The Liberation Daily article "When AI 'Stirs Up' the Courtroom" records [6] that at the end of October 2025, Chen Ying, head of the labour dispute adjudication team of the Civil Division of the Shanghai No. 2 Intermediate People's Court, received appeal materials. The party was a 65-year-old man, and of the four items of material attached to the appeal, two were marked by him personally on the envelopes as coming from the search results of a certain artificial intelligence software. This was a repetitive lawsuit; the dispute had already gone through first instance, second instance and appeal review, and none of the claims had been supported. The old man sued again, the first instance ruled to dismiss, and he then filed an appeal with the assistance of artificial intelligence. During the offline reception, the old man repeatedly began with "according to what AI said", with a firm attitude. Chen Ying found that the analysis given by the artificial intelligence was quite specific, but the old man had selected only the part favourable to his own claims, and had not adopted most of the analysis inconsistent with his claims.

The problem revealed by this detail is deeper than the errors of artificial intelligence itself. When using artificial intelligence, users often unconsciously perform a filtering: retaining conclusions that match their expectations and ignoring content to the contrary. And artificial intelligence will not give any warning about this.

The superposition of these three factors makes errors by artificial intelligence in professional fields such as law, to a considerable degree, inevitable. In the same analysis mentioned above, Judge Zhao Yinhao of the Beijing No. 2 Intermediate People's Court also noted that the legal knowledge system is vast, updates rapidly, and contains a large number of unpublished documents or internal norms, which objectively tends to create "information vacuum zones." This judgment holds true in many fields. Take tax cases as an example: the enforcement standards of tax authorities in various localities are not fully public, and the industry risk parameters held by inspection departments are even less disclosed to the outside. AI training data does not include such content; what it can rely on is mainly public provisions and scattered materials retrievable on the internet. Between judgments formed on this basis and the professional argumentation that should be achieved, there is a gap difficult to bridge.

It should also be pointed out that the form of the model's output itself can also aggravate misuse. A response that is structurally complete, clearly organized, and confident in tone will make people mistakenly believe that it contains a complete argument. In reality, the reasoning process between the provisions and the conclusion may not exist, the cited sources may not be genuine, and the version applied may long have been invalid. In the aforementioned Beijing No. 2 Intermediate People's Court case, not a single one of the adjudicative bases submitted by the party could be matched to current law, yet the party had not the slightest doubt about the appearance of these contents. It can be seen that the key to the problem is not whether the model will make errors, but that when it makes errors there are no external signs whatsoever.

For enterprises, the real risk lies in directly embedding this set of outputs into business processes. For checking a tax rate, finding a legal provision, or understanding the general steps of a procedure, artificial intelligence is far more efficient than manual retrieval, and whether to use it is not really a question. The problem arises at the step of treating it as a final conclusion. A classification opinion is written into a customs declaration without review, a contract clause is signed and sealed without examination, a legal opinion is issued externally without argumentation—at this point, a probabilistic deviation of the model has already turned into the enterprise's declaration conduct, contractual obligations, and external commitments. In the aforementioned Kunming case, Luo may not have been a rash person; he merely handed a question requiring professional judgment to a tool that could only provide probabilistic answers, without setting any intermediate link between the two. That link should originally have been the work of a lawyer.

What is harder to detect than fabrication is another situation: the provision given by the AI model itself genuinely exists, the article number is accurate, the original text of the legal provision is even word-for-word identical, and the sentencing range is cited without error, but this provision does not fit the case to be resolved. Judging this mismatch requires grasping two things at the same time: first, what the applicable premise of that norm is; second, whether the facts of the case fall within that premise. And these two things are precisely where retrieval can give the first half but cannot give the second half. What the user sees is that the provision is genuine, the citation is standard, and the expression is professional; what the user does not see is that it has been placed at the wrong level on the norm's ladder of application.

The reason this form is dangerous is that it cannot be identified by common sense. A fabricated legal provision, once retrieved, will be found not to exist, and even a layperson can become suspicious. But when a genuine legal provision is placed in the wrong position, retrieval can verify only that "this provision exists," not whether "this provision should be used here." To discover this error, the judge must have an overall grasp of the normative system of the field, know whether provisions are in a parallel or exclusive relationship, and know that once a certain premise does not hold, the entire chain of reasoning becomes invalid. This kind of judgment does not come from retrieval; it can only come from long-term experience in handling similar issues. Precisely for this reason, such errors are often not discovered at the time of use, but only upon peer review, or even only after external publication.

III. Corroboration of Similar Situations in Litigation and Non-litigation Fields

Beyond customs, situations in which parties use artificial intelligence on their own and suffer adverse consequences have appeared multiple times in civil litigation, and in many cases have been disclosed by courts under real names. What these cases have in common is that the parties were all convinced they had obtained a reliable answer, while the error landed in the most direct way on the adjudicative documents.

In an inheritance dispute case heard by the Beijing No. 2 Intermediate People's Court in September 2025 [7], the party submitted to the court adjudicative bases obtained through artificial intelligence searches. The presiding judge, Tu Yu, upon comparison, found that not a single one was an accurate current legal provision.

In a decoration and renovation contract dispute case disclosed by the Shanghai Baoshan District People's Court in April 2026 [8], the plaintiff, Tan, in order to prove that the underfloor heating product involved had insufficient power, asked an artificial intelligence software, "What area is this model of boiler suitable for," and received the reply that "the heating area is suitable for residential housing of 80 to 120 square meters," and submitted that reply to the court as evidence. After verification, the court found that the materials cited in the reply mainly came from online marketing information and product introductions, and lacked authoritative bases such as industry norms and national standards. To verify the stability of the conclusion, the court in court used two other generative artificial intelligence tools to ask the same question, and obtained conclusions that were not entirely consistent or even broader. On this basis, the court held that the objectivity, authority, and probative force of this AI-generated content were all insufficient.

The common feature of the above cases is that their false content was all seen through on the spot by the presiding judges. The number of situations that were not seen through, or that were seen through but not publicly reported, is certainly not small. Si Yanli, Deputy Director of the Research Office of the Supreme People's Court, stated at the press conference on the Opinions on the Lawful Trial of Artificial Intelligence-Related Dispute Cases on September 7, 2026 [9]: "The people's courts have discovered multiple false cases generated by artificial intelligence, and the parties and agents had not conducted comprehensive verification before submitting them to court." Article 19 of the Opinions further clarifies that if litigation participants submit litigation documents, case retrieval reports, and other materials generated by artificial intelligence, they shall verify the authenticity and accuracy of the relevant laws, judicial interpretations, cases, and other content before submitting them to court, and shall explain the use of artificial intelligence assistance at the time of submission, and bear legal responsibility for the authenticity and accuracy of the relevant content in accordance with law. Artificial intelligence is only an auxiliary tool, and litigation participants may not exempt themselves from their own legal responsibility on the grounds of "AI hallucination" or technological neutrality.

In litigation there is still a judge to guard the gate, and the cost is mostly criticism and education or non-adoption. In non-litigation projects, however, this layer of gatekeeping does not exist. Deliverables such as contracts, legal opinions, and due diligence reports directly enter the transaction once issued, with no third party to verify them on behalf of the party. Defects in the validity of clauses, omissions of rights and obligations, and unenforceability of dispute resolution mechanisms often only come to light when performance is obstructed or a dispute arises. In October 2026, the government of Western Australia commissioned an international consulting firm to issue a feasibility report on the merger of three universities, paying AUD 166,000. After the report was published, it was found that some literature citations had been generated out of thin air, and one citation pointing to a World Health Organization report had metadata markers of generative artificial intelligence remaining in its URL. Two cost totals in the appendix also did not match the sum of the sub-items, with discrepancies of AUD 260 million and AUD 311 million respectively. The state government subsequently wrote to the firm demanding accelerated refund of the fees already paid. This was not the first time this firm had encountered such a problem in Australia; in 2025 it partially refunded the federal government due to similar errors in another report [10].

IV. The Necessity of Professional Lawyer Involvement and Its Specific Manifestation in Different Scenarios

At this point, the question is no longer whether artificial intelligence can be used, but whether the party's own retrieval can replace a lawyer's professional judgment. The author's view is that it cannot. This judgment has its own specific reasons in the three scenarios of litigation, non-litigation, and customs procedures.

In litigation, the involvement of a lawyer means that someone verifies item by item before submission. The reason the aforementioned cases were seen through precisely shows that the courts are undertaking this verification. But the consequence of court verification is non-adoption, criticism and education, or even fines. What the party truly needs is someone to tell him before submission whether this material can be used. Litigation is characterized by concentrated issues, clear time limits, and the presence of the opposing party; once a fabricated adjudicative basis or a "piece of evidence" of unknown origin is submitted, it is difficult to withdraw, and it directly affects the judge's assessment of the party's credibility. The cost of such errors often exceeds the material itself. Yet court verification can only target materials already submitted, and cannot replace preparation on the party's side. What should and should not be submitted within the time limit for evidence, how the claims should be organized, and what defenses the opposing party may raise—all these questions require judgment to be completed before the hearing. Artificial intelligence can generate a document, but it does not know which judge, which case file, or which changed adjudicative rule this document will face.

In non-litigation projects, the involvement of a lawyer means that someone bears responsibility for the documents issued. Unlike litigation, non-litigation has no judge adjudicating in the middle, nor an opposing party cross-examining in court. When it is discovered only after the contract is signed that a clause cannot be enforced, or only after the legal opinion is issued that the premise does not hold, the transaction has often already been completed and the loss has already occurred. The characteristic of non-litigation is that once a deliverable is issued it creates reliance, and the cost of correction is higher than that of redoing it. In this link there is neither a judge nor a second chance. The focus of non-litigation work therefore lies not in writing, but in upfront verification and structural design. Whether a transaction arrangement touches export controls, constitutes a concentration of undertakings, or triggers the negative list for foreign investment access—these judgments need to be completed before the plan takes shape; once the structure is built and then adjusted, it often entails renegotiation of commercial terms. Artificial intelligence can provide checklist-style prompts, but the parts beyond the checklist are precisely where professional judgment comes into play.

In customs procedures, the involvement of a lawyer means that someone is able to complete the assessment and response within an extremely short window. In a customs administrative penalty case that the author handled, the party was ordered to pay the equivalent value of smuggled goods of more than two million yuan for under-declaring the price. The lawyer handling the case began reviewing the case file immediately upon accepting the engagement, sorted out the key facts, concluded that the penalty suffered from erroneous application of laws and regulations and insufficient reasonableness, and advised the party to apply for a hearing immediately. After two hearings, the customs authority ultimately revoked the penalty decision within two months. The characteristic of customs procedures lies in the rigidity of time limits. The most economical way to resolve a classification dispute is to fix the code in advance through an advance ruling before declaration; one ruling is valid for three years and applies across the entire customs territory. Once goods are questioned, complete technical materials and a statement of circumstances must be submitted within seven working days, and the goods are not released in the meantime. At the administrative penalty stage, a decision on whether to apply for a hearing must be made within five days. The existence of these time limits means that by the time the party receives the relevant documents, the time left is no longer sufficient to start from scratch. Take the questioning procedure for dual-use item export control as an example: when customs has doubts about the nature of the goods, it may require the enterprise to submit materials within seven working days, and customs completes its determination within fifteen working days; where identification by the Ministry of Commerce is needed, the matter is further referred, and during the identification period the goods are not released and may not be released against a guarantee. From the enterprise's receipt of the notice to the completion of material submission, the time available for technical argumentation and legal analysis is extremely limited; any enterprise that has not established complete technical files and compliance processes in advance will find it almost impossible to respond calmly when the time comes.

What the three scenarios have in common is that what the lawyer provides is not knowledge, but judgment, responsibility and timing. At the knowledge level, artificial intelligence can already cover a considerable portion, and covers it more comprehensively than an individual's search. But judgment requires applying norms to specific facts and choosing among multiple available paths, which requires understanding the full picture of the case and knowing how similar disputes are usually handled in practice. Responsibility requires a subject that can be held to account, one that can be pursued after an error and is willing to bear the consequences of its conclusions. Timing requires someone to remember what should be done at the right time, rather than starting preparations only after the deadline has expired. None of the three can be replaced by search.

This is also a point the author repeatedly emphasizes in daily work: the effectiveness of risk handling in legal affairs depends to a large extent on the timing of intervention. A classification argument before a declaration may cost far less than two hearings afterwards. A review before a transaction document is issued may matter more than several rounds of debate after a dispute arises. When a party comes to a lawyer after a dispute has erupted, the options at that point can often only be adjusted within the existing framework to a limited extent. But when intervention occurs before matters take shape, many disputes simply never arise.

V. Conclusion

Returning to the penalty decision at the beginning of this article. The person who carried out the inquiry was Luo, yet those who bore the administrative penalty liability were the two companies that lent their legal person names; the former was not listed as a penalty target, and the latter were unaware of the illegal export. This allocation of liability is not a peculiarity of this case, but the inevitable result of customs administrative penalties being based on the declaration act as the basis for attribution. But it reveals a reality after artificial intelligence intervenes in professional affairs: the person who uses artificial intelligence, the person who makes decisions based on artificial intelligence, and the person who ultimately bears the legal consequences are not necessarily the same subject. In this process, no party can claim that the consequences of inaccurate declaration should be borne by the artificial intelligence software. Artificial intelligence can quickly list several tariff codes that a certain commodity may correspond to, can explain the penalty range corresponding to inaccurate declaration, and can generate a statement of circumstances with smooth sentences; it completes these tasks faster than a person. But it cannot determine what the goods involved should be evaluated as in law, cannot conduct argumentation with customs at a hearing, and cannot complete the organization of technical materials and the citation of legal bases within seven working days. The formation of professional judgment requires facts as the foundation, norms as the basis, and procedure as the path; this process cannot be replaced by artificial intelligence, and still requires the professionalism and experience of lawyers to accomplish.

Appendix

Questions and answers

What happened in the customs administrative penalty case involving AI commodity classification?

In November 2025, You entrusted Luo to declare three lathes for export to Myanmar. Luo used AI software, which said the goods were not export-controlled and gave code 8461909000. Customs later found the correct code was 8458110010, requiring a dual-use export licence. Two companies that lent their names were fined 10,500 yuan each.

Why is commodity classification a legal judgment rather than a simple search task?

Classification requires applying the General Rules for the Interpretation of the Harmonized System, such as Rule 3 and Rule 6, to the specific goods. The same goods may fall under different headings depending on technical parameters and use. AI can list possible codes but cannot determine what the goods should legally be evaluated as.

What are the legal consequences of an incorrect commodity classification under Chinese customs law?

Consequences range from recovery of duties plus late payment surcharges, to administrative penalties under Article 15 of the Implementing Regulations of the Customs Administrative Penalty (fine of 30% to twice the duties evaded), and potentially the crime of smuggling under Article 151 of the Criminal Law if intent to evade customs supervision is found.

How do Chinese courts treat AI-generated materials submitted by parties in litigation?

Courts have found multiple false cases generated by AI. Under Article 19 of the Opinions on the Lawful Trial of Artificial Intelligence-Related Dispute Cases, parties must verify authenticity and accuracy before submission, explain AI assistance, and bear legal responsibility. Courts may refuse such materials and criticize or fine parties.

Why can't AI replace lawyers in customs procedures and other legal matters?

Lawyers provide judgment, responsibility and timing. Judgment requires applying norms to specific facts and choosing among paths; responsibility requires an accountable subject; timing requires meeting strict deadlines such as seven working days for technical materials or five days for hearing applications. AI can retrieve knowledge but cannot perform these roles.

Laws and cases cited

Laws

  • Customs Law
  • Criminal Law

Regulations

  • Implementing Regulations of the Customs Administrative Penalty
  • Import and Export Tariff

Judicial interpretations

  • Opinions on the Lawful Trial of Artificial Intelligence-Related Dispute Cases

Other

  • Wanding Guan Ji Wei Zi [2026] No. 17

References

  1. http://kunming.customs.gov.cn/kunming_customs/2026-07/21/article_2026072121465513217.html
  2. http://www.customs.gov.cn/customs/2008-06/04/article_2025121220574118704.html
  3. http://www.shfzb.com.cn/shfzb/html/2026-08/18/content_150505_3751191.htm
  4. http://www.ce.cn/xwzx/gnsz/gdxw/202601/t20260108_2689678.shtml
  5. Ibid.
  6. https://www.jfdaily.com/staticsg/res/html/web/newsDetail.html?id=1064857
  7. http://www.ce.cn/xwzx/gnsz/gdxw/202601/t20260108_2689678.shtml
  8. https://m.thepaper.cn/newsDetail_forward_32928379
  9. https://www.court.gov.cn/zixun/xiangqing/514921.html
  10. https://zot.news/article/wa-government-seeks-reimbursement-deloitte-flawed-university-b5vxd