Incidents of analysts using AI to write ostentatious research reports still occur. Recently, a real estate industry weekly report signed by two analysts mistakenly labeled the statement on real estate at the 2024 State Council Executive Meeting as the latest policy on June 7, 2026. Many industry insiders told reporters that this error mode of "time misalignment + complete content transfer" is in line with the typical characteristics of the "illusion" of large AI models.
The research report mentioned that the State Council executive meeting on June 7, 2026 will deploy real estate-related work, which is actually information from two years ago.
This research report reads, "The State Council executive meeting on June 7, 2026 deployed real estate-related work, requiring the implementation of existing policies, reserve new measures to destock and stabilize the market, steadily revitalize existing real estate and land, accelerate the construction of a new model of real estate development, and improve the 'market + security' housing system." However, through public information tracing, this is actually an old document from 2024. The real estate issue was mentioned at the State Council executive meeting held on June 7, 2024.
Judging from common sense, for such a high-level State Council executive meeting, as a real estate industry analyst who follows policy, it is impossible for a fundamental cognitive error to be made about the time and content of the meeting. The State Council did hold a relevant meeting on June 5 this year, but the content and time did not correspond.
Coincidentally, reporters also found almost identical erroneous statements on some self-media platforms, except that the time was wrongly written as "June 6, 2026." Obviously, analysts or underlying large models encountered serious "information pollution" when capturing public information on the Internet, and AI directly "traveled" old news from two years ago to today without strict review.
Some self-media also quoted similar erroneous information.
Investment research is the field with the deepest implementation and fastest results of AI in the securities industry, because the core work of investment research is information processing and knowledge production, which happens to be the capability of large models. Using AI to assist in writing research reports has become a common practice in the industry. Many brokerage research institutes have also established AI investment research groups and regularly publish AI-generated research reports. At present, AI mainly plays an auxiliary role in research report writing, such as polishing text, quoting public information, etc.
Although it is a auxiliary role, information pollution on the Internet is entering the field of professional research through AI tools, which deserves vigilance. Just in 2025, a brokerage company circulated an erroneous PPT caused by information pollution, which stated that "80% of retail investors will lose money this year." In the end, it was found that everything from the data to the data source was wrong.
As a key information bridge connecting investors and the capital market, securities firm research reports play a vital role in guiding investment directions, improving market transparency, and safeguarding the interests of investors. Nowadays, as the securities industry deeply embraces AI technology, the challenges of process compliance and trust mechanisms cannot be ignored.
Why do hallucinations occur?
The type of research report that exposed the problem this time is a weekly report regularly released by securities companies. Such normalized and high-frequency research products are precisely the areas where AI hallucinations are most likely to occur.
It's not hard to understand why. Weekly newspapers and daily newspapers are “low-cost” research outputs and generally have a fixed template framework. As long as you apply a template, import data, and cite public information, you can quickly write a document. The introduction of AI tools has further shortened the production cycle.
Convenience often comes with carelessness. In the templated production process, analysts often do not go back to the original sources one by one to review the policy content automatically captured and generated by AI. In particular, some policy statements that appear to be “correct” are more likely to be let go during the review process.
On the other hand, the popularity of AI tools in the research departments of securities firms has created a superimposed effect with changes in personnel structure.
A large number of young analysts and interns have been exposed to and accustomed to using AI writing tools earlier. For them, using large models to organize data and generate first drafts is standard in their daily work. However, some personnel lack awareness of the importance of source verification and rely too much on AI output results.
When there are already a large number of erroneous statements about the "2026 State Council Real Estate Conference" on the Internet, AI may output this pollution information as "facts" during the training and reasoning process and enter it into a formal research report.
This is not the first time that AI has exposed the problem of “illusion” in the securities field. The Financial Associated Press previously published in "Will the "Model Illusion" of Robo-Advisors Mislead Investors?" In the report "Three Major Pain Points Survey", we have conducted follow-up observations on AI factual deviations in the investment advisory business.
How to solve hallucinations?
A principle that has formed a consensus in the industry is that the output of investment research AI must always bear the label of "human review".
Any AI-generated content that has not been reviewed by analysts shall not appear in official research reports. This is not an efficiency issue, but a compliance issue - if something goes wrong, it is the analyst who signs off, not the AI.
However, in actual operation, how to define the standards of "review" and how to balance efficiency and quality are still issues that each securities firm needs to face.
As of now, the China Securities Regulatory Commission has not issued new regulatory regulations specifically for AI-generated research reports. The "Administrative Measures for Information Disclosure of Listed Companies" revised in March 2025 added regulatory provisions for "outsourcing" behavior, but did not explicitly mention AI.
AI compliance in investment research scenarios currently mainly relies on industry self-regulatory guidelines and securities firms’ internal risk control systems. This "grey area" is both a challenge and an opportunity - the lack of clear rules means room for innovation, but it also means that the boundaries of responsibility need to be grasped by oneself.
It is worth mentioning that since 2025, there has not been a large-scale distortion of research reports due to AI writing in the industry, which shows the overall importance that securities firms’ research businesses attach to compliance.
But once a problem occurs, the impact cannot be underestimated. A research report with factual errors will not only arouse public opinion, but more importantly, damage the trust of buy-side institutions and investors in sell-side research.
“Is the research report everyone sees necessarily correct?” Once doubts like this take root in the minds of investors, it will damage the credibility of the entire sell-side research industry.
What’s interesting is that AI is both creating problems and solving them. According to the reporter's research, the intelligent manual operation platforms introduced by many securities firms are bringing systematic innovation to the seller's research and compliance management system. The design of standard workflow and the support of artificial intelligence technology have achieved the dual upgrade of standardizing the writing process of securities research reports and unifying the review standards.
Using more mature AI to constrain and standardize the content generated by AI may be an inevitable choice on this technical path. A brokerage analyst told reporters that no matter how technology evolves, one thing will not change. The capital market’s requirements for information authenticity are higher than all efficiency improvements.