CNIPA’s Two-Part Strategy: Better Patents and Smarter Examination

The China National Intellectual Property Administration, commonly known as CNIPA, used its June 27, 2025, press conference to highlight two closely connected priorities. The agency described an intensified campaign against abnormal patent applications and unlawful patent agency practices. It also explained how artificial intelligence would play a larger supporting role in patent examination, particularly in prior art searches, technical analysis, document classification, and case distribution.

These initiatives may look unrelated at first. One is about fraud, forged records, questionable agencies, and applications created mainly to collect subsidies or satisfy performance targets. The other is about machine learning, semantic retrieval, image recognition, and large AI models. In practice, however, both programs serve the same objective: improving patent quality.

A patent system works only when granted rights represent genuine inventions that have survived a credible examination. If applications are fabricated, copied, filed under false identities, or submitted in industrial quantities with little technical merit, even the cleverest examination software becomes an expensive garbage sorter. CNIPA is therefore attacking the quality problem from both endswho files the application and how the application is reviewed.

Why Patent Fraud Became a Major Regulatory Target

Understanding “Abnormal” Patent Applications

CNIPA generally uses the term “abnormal patent applications” for filings that appear inconsistent with legitimate innovation activity. Examples may include applications fabricated without real research, repeated filings with substantially similar content, applications using false applicant information, or mass submissions designed to obtain subsidies, professional credentials, tax benefits, awards, or other non-patent advantages.

Not every rejected or weak application is fraudulent. Inventors can misunderstand the prior art, draft claims too broadly, or simply overestimate how revolutionary their new bottle opener may be. Abnormal filing behavior becomes more serious when applications are deliberately manufactured, identities are misused, official documents are forged, or an agency promises a guaranteed grant before an examiner has even opened the file.

The 35 Institutions Placed Under Centralized Rectification

At the June 2025 press conference, CNIPA said it had placed 35 institutions associated with large numbers of abnormal patent applications under centralized rectification. The agency indicated that serious sanctions, including business suspension and license revocation, were being considered. Regional authorities were also investigating 65 additional institutions, with warnings, fines, and other penalties expected depending on the facts of each case.

CNIPA also reported investigations involving 12 key institutions suspected of conduct such as fraudulently using company information, forging seals or documents, and fabricating patents. Suspected criminal conduct involving forged official seals was being referred to public security authorities rather than handled only as an administrative paperwork problem. That distinction matters: an agency license violation is one thing; an organized scheme involving stolen identities and forged government documents is a much less charming creature.

The enforcement campaign extended to online advertising. Regulators identified 89 examples of improper marketing, including “guaranteed authorization” claims and language that allegedly impersonated official organizations. Ten online platform companies were instructed to remove or correct the problematic promotions. CNIPA also reported action involving agencies that failed to meet licensing conditions, institutions that neglected annual reporting duties, and businesses suspected of obtaining licenses through concealed or false information.

The Continuing “Blue Sky” Campaign

The enforcement effort forms part of CNIPA’s long-running “Blue Sky” campaign to regulate the patent and trademark agency industry. In 2025, the campaign emphasized coordinated enforcement, credit-based supervision, public disclosure, cross-regional information sharing, platform accountability, and the use of AI-powered monitoring tools to identify risks earlier. The stated goal was to move supervision from passive reaction toward active prevention.

This approach reflects a broader policy change. Chinese authorities have increasingly discouraged evaluation systems that reward raw patent numbers without considering commercial value, technical quality, or legal stability. Quantity can be useful, but when every employee needs three patents before lunch to qualify for a promotion, the incentive system may produce more paperwork than innovation.

How CNIPA Is Expanding Artificial Intelligence in Patent Examination

AI Tools Were Already Operating Before the Announcement

CNIPA did not suddenly discover artificial intelligence during the 2025 press conference. The administration launched an intelligent patent examination and retrieval system in 2023 and tested it across several examination scenarios. These included semantic searches based on an invention’s underlying concept, image searches using portions of industrial designs, and the automatic clustering and allocation of application files.

CNIPA has also described using machine translation, natural language processing, and image recognition to distribute relevant prior art, locate visually similar designs, classify documents, and identify formal deficiencies in application materials. By April 2025, the administration reported that the average examination period for invention patents had fallen to 15.5 months, while the accuracy rate for completed invention patent examinations had reached 95.2%.

Those figures should not be interpreted as proof that AI alone produced the improvements. Staffing, procedural reform, specialized examination programs, better databases, and workload management also influence examination time and quality. AI is one tool in a crowded toolbox, not a magic robot examiner wearing a tiny judicial wig.

Improving Prior Art Searches

Prior art searching is one of the most promising applications for AI. Patent databases contain millions of documents written in different languages, using inconsistent terminology and often describing familiar concepts with heroic levels of verbal creativity. A conventional keyword search can miss a relevant document simply because an earlier inventor used different vocabulary.

Semantic retrieval systems can analyze meaning rather than relying solely on exact words. An examiner reviewing a patent for battery temperature management, for example, may receive technically similar documents that refer to thermal regulation, heat dissipation, cell conditioning, or energy-storage safety without using the applicant’s preferred phrase.

CNIPA said future upgrades would focus on improving the accuracy of searches for comparison documents and helping examiners understand inventive concepts and complete technical analysis more quickly. Better search results may improve examination consistency, reduce the chance that important prior art is overlooked, and produce patents that are more reliable when challenged.

AI Will Assist Examiners, Not Replace Them

CNIPA placed an important limit on the new technology: AI-generated results cannot be used directly as examination opinions. Human examiners remain responsible for reaching conclusions under China’s Patent Law, its implementing regulations, and the Patent Examination Guidelines.

That human-in-the-loop rule is essential. A search model can rank documents, identify linguistic similarities, or flag suspicious patterns. It cannot independently guarantee that a claim lacks novelty, involves an inventive step, contains adequate disclosure, or complies with every legal requirement. Patent examination requires legal interpretation and technical judgment, often in cases where reasonable experts disagree.

The model may find the haystack faster. The examiner must still determine whether the shiny object inside it is actually a needle.

AI Is Changing Patent Offices Around the World

CNIPA’s modernization is part of a larger international trend. The United States Patent and Trademark Office has integrated AI-supported similarity searching into its examiner tools and launched programs testing automated prior art searches before substantive examination. The USPTO has consistently described these systems as tools that augment rather than replace examiner searches and judgment.

The European Patent Office has likewise expanded AI-supported search technology, including its ANSERA platform and related tools used to search patent documents and non-patent literature. These developments suggest that AI-assisted examination is becoming normal infrastructure among major patent offices rather than an experimental side project maintained by three engineers and a nervous legal department.

CNIPA, the USPTO, the EPO, the Japan Patent Office, and Korea’s intellectual property office cooperate through the IP5 framework. Together, the IP5 offices handle roughly 90% of global patent applications, making their technical and procedural choices influential far beyond their national borders.

What the Changes Mean for AI Patent Applicants

Detailed Disclosure Is Becoming More Important

Companies filing AI-related patents in China should not assume that an enthusiastic regulatory attitude toward artificial intelligence means easy approval. CNIPA’s guidance divides AI-related inventions into categories such as improvements to models, AI-based technical applications, AI-assisted inventions, and AI-generated inventions. It also emphasizes technical effects, sufficient disclosure, and a genuine relationship between algorithmic and technical features.

An application that merely announces “we used AI to improve everything” is unlikely to impress an examiner. Applicants may need to explain model architecture, processing steps, input and output relationships, important parameters, training procedures, data characteristics, and how the claimed system solves a technical problem.

Natural Persons Must Still Be Named as Inventors

CNIPA’s AI patent guidance maintains that an inventor must be a natural person who made an inventive contribution. An AI system cannot be listed as the inventor because software does not possess the legal capacity associated with patent ownership and inventor rights.

Businesses using generative AI in research should therefore document human contributions carefully. Records should show who identified the problem, selected inputs, designed experiments, evaluated results, modified the model, and recognized the useful technical solution. Waiting until filing day to reconstruct inventorship through a group chat titled “Who actually invented this?” is not a recommended governance system.

Foreign Companies Must Vet Chinese Patent Agencies

The enforcement campaign creates a practical due-diligence obligation for foreign applicants. Companies should verify that a service provider holds the required patent agency license, confirm the identities and qualifications of the professionals handling the work, review engagement terms, and avoid providers promising guaranteed grants or suspiciously inexpensive filing packages.

Price matters, but a bargain application based on copied technical material, false information, or mass-produced claims can create prosecution problems, reputational harm, and future validity risks. Saving money during drafting is less exciting when the resulting patent collapses during litigation.

Potential Benefits and Risks of AI-Assisted Examination

Possible Benefits

• Faster prior art discovery: Semantic and image searches can surface relevant references that ordinary keyword searches may miss.
• More consistent classification: Automated document clustering may route applications to examiners with appropriate technical expertise.
• Earlier quality control: AI systems can flag missing information, suspicious filing patterns, duplicated content, or unusual agency behavior.
• Reduced repetitive work: Examiners can spend less time sorting documents and more time applying legal and technical judgment.
• Improved multilingual searching: Machine translation may help examiners find relevant foreign-language prior art.

Risks That Require Careful Management

AI search results depend on training data, indexing quality, model design, and the information supplied in each query. A system may overvalue linguistic similarity, underrepresent older records, struggle with formulas or chemical structures, or return confidently ranked documents that are legally irrelevant.

Automation bias is another concern. Examiners may place too much weight on highly ranked AI results, while applicants may find it difficult to understand why certain references were selected. Patent offices will need strong audit trails, security controls, performance testing, and procedures for challenging decisions influenced by unreliable automated output.

Confidentiality also matters. Unpublished applications may contain commercially sensitive technical details. Patent offices and agencies must ensure that AI tools do not expose confidential information, use it for unauthorized training, or transmit it to uncontrolled third-party systems.

The Crackdown Continued Into 2026

Developments after the June 2025 announcement indicate that CNIPA’s enforcement program was not a one-day press conference performance. CNIPA reported that, during 2025, authorities imposed 256 administrative penalties for serious agency violations, canceled registrations involving 180 patent agencies and 1,208 branches, and removed thousands of patent attorney practice registrations.

In April 2026, CNIPA, the Ministry of Public Security, and the State Administration for Market Regulation launched a nationwide “Year of Rectification and Standardization” for the intellectual property agency industry. The plan targeted forged applicant data, fabricated patents, unauthorized agencies, rented qualifications, improper solicitations, suspicious payments, and black or gray commercial chains connected to patent fraud.

CNIPA later publicized a risk list containing 14,460 companies whose names referred to intellectual property or patent agency services even though they had not obtained a patent agency practice license. The notice did not declare that every listed company was unlawfully performing patent agency work, but it warned innovators to verify qualifications before purchasing regulated services.

These later measures reinforce the underlying policy: China wants fewer applications created merely to inflate statistics and more patents tied to genuine research, industrial value, and enforceable technical contributions.

Practical Experiences and Lessons for Patent Teams

Experience One: Incentives Shape Filing Behavior

Organizations managing large patent portfolios repeatedly discover that internal incentives determine what reaches the patent office. When bonuses, promotions, academic evaluations, or departmental rankings depend mainly on application counts, employees naturally optimize for volume. A team may divide one modest development into numerous narrow filings, submit speculative concepts before testing them, or treat patents as certificates rather than legal rights.

The practical lesson is to measure quality before quantity. Internal review committees can assess technical readiness, commercial relevance, available prior art, claim scope, inventorship, and alignment with product strategy. A smaller portfolio of defensible patents usually serves a business better than a warehouse full of decorative certificates.

Experience Two: AI Search Often Exposes Drafting Weaknesses Early

AI-assisted prior art tools can be uncomfortable in a useful way. A development team may believe its invention is unique because nobody in the room has seen it before. A semantic search may locate similar systems from another industry, translated documents using unfamiliar terminology, or older patents that solve the same technical problem through slightly different components.

Finding this material before filing gives the applicant options. Engineers can refine the invention, counsel can focus the claims on the genuine improvement, and management can decide whether patent protection is worth the expense. Discovering the same references in an examiner’s rejectionafter translation fees, filing fees, and several rounds of attorney workis the more traditional and considerably less festive route.

Experience Three: Agency Verification Cannot Be Treated as Clerical Work

International companies sometimes select local providers through procurement departments focused heavily on cost and turnaround time. CNIPA’s enforcement activity demonstrates why professional qualification, supervision, confidentiality, and drafting quality deserve equal attention.

A sensible onboarding process should verify the agency’s official status, identify the responsible patent attorneys, examine sample work, confirm conflict-checking procedures, and establish rules for using subcontractors or AI drafting platforms. Companies should also require written approval before an agency changes inventors, applicants, claim scope, or technical content.

Experience Four: Human Review Must Remain Visible

AI tools are increasingly useful for searching, translation, classification, summarization, and document comparison. They can reduce repetitive labor, but they can also invent citations, misunderstand technical relationships, or produce language that appears polished while quietly changing the invention.

Effective teams assign accountable human reviewers at every important stage. Engineers verify technical accuracy. Patent counsel verifies legal strategy and claim support. Inventors confirm their contributions. Filing specialists verify names, addresses, priority data, and formal documents. The goal is not to reject AI; it is to prevent automation from becoming a very fast method of producing avoidable mistakes.

Experience Five: Strong Records Become More Valuable as Screening Improves

As patent offices deploy better pattern detection, organizations should expect closer scrutiny of duplicated language, unusual filing volumes, shared addresses, questionable signatures, and inconsistent inventor information. Applicants with real research activity should maintain dated invention disclosures, laboratory records, technical drawings, meeting notes, source-code histories, and assignment documents.

Good documentation supports inventorship, proves ownership, helps counsel draft stronger applications, and distinguishes authentic innovation from mass-produced filings. In an AI-assisted enforcement environment, a clean evidence trail is not bureaucratic clutter. It is the patent portfolio’s seat belt.

Experience Six: Coordinate Global Filing Strategies

A patent application prepared for China may also support filings in the United States, Europe, Japan, and other markets. However, each jurisdiction applies its own rules concerning disclosure, eligibility, inventorship, data-related inventions, and foreign filing authorization.

Global teams should therefore plan claims and supporting descriptions before the first filing. An AI invention may need detailed technical interactions for CNIPA, eligibility-focused explanations for the USPTO, and a clearly demonstrated technical effect for the EPO. Drafting one generic application and hoping it charms every patent office equally is a strategy best described as optimistic.

Conclusion

CNIPA’s campaign against patent fraud and its expansion of artificial intelligence are parts of a single modernization project. Enforcement is intended to reduce fabricated applications, unlicensed services, deceptive advertising, and filing practices driven by artificial incentives. AI-assisted examination is intended to help human examiners search broader collections, understand technical concepts faster, and manage a massive volume of patent information.

The success of the program will depend on balance. Aggressive enforcement must distinguish fraud from honest drafting mistakes. AI systems must improve searches without becoming unchallengeable black boxes. Examiners must remain accountable for legal conclusions, while applicants must provide accurate information, meaningful technical disclosure, and reliable evidence of human invention.

For businesses, the message is straightforward: select qualified patent professionals, document real inventive activity, conduct stronger prior art searches, and treat AI as an assistant rather than an oracle. The future patent system may be faster and more automated, but it will not be more forgiving of weak applications wearing a fashionable algorithm as a hat.