https://english.news.cn/20260612/36f9bb93ae6f45b397c9c5d6c29acf04/c.html

During an elevator procurement evaluation in eastern China, an AI system analyzed a tender and generated a comprehensive report in just 15 minutes, a task that would typically require five human experts three and a half hours to complete. The human review panel examined the AI’s findings and unanimously approved its conclusions.

This scene in Hefei, capital city of Anhui Province, marked the first real-world deployment of a large language model tailored for public bidding, a domain long plagued by graft risks, paperwork overload and hard-to-detect collusion.

China is fast-tracking the application of its new intelligent computing power across industries, and this technological leap is also expected to transform how the society is governed. Developed by the local public resource trading center in partnership with AI tech firm iFLYTEK in 2024, the model was trained on 290 documents of laws and regulations, as well as 82.9 terabytes of data including past tenders and bidding documents. Since the start of 2025, the system has audited about 36,000 tender documents, flagging 3,264 questionable clauses, including hidden discriminatory or exclusionary terms that are often difficult for humans to detect. Among 657 AI-monitored project cases reviewed later by human experts, unanimous expert adoption rates stand at 91.17 percent.

In one case, the AI detected a non-compliant qualification requirement in an outdoor wall repair tender. The model correctly noted that such restrictions are generally prohibited for standard construction projects, proving its ability to spot hidden irregularities.

One of the model’s most prized features is its ability to detect bid-rigging and collusion, a chronic headache for regulators worldwide. Traditional detection relies on superficial signals such as IP addresses or machine codes, which sophisticated cheaters can easily bypass. The AI, by contrast, uses semantic analysis, logical reasoning and multimodal analysis to compare thousands of documents for unusual similarities in phrasing, paragraph structure, embedded images and even identical typos.

Unlike general-purpose AI, the tendering model is designed to minimize “hallucinations,” which is the tendency of large language models to generate plausible-sounding but false information. It does so by anchoring its output to specific regulations.

In early 2026, eight departments including the NDRC jointly issued a directive to expand AI use in tendering nationwide, aiming for full coverage of key scenarios such as document scrutiny, assisted evaluation and collusion detection in selected provinces and cities by the end of 2026, and nationwide by the end of 2027.