New Constructs Warns AI-Generated Investment Research Could Fuel Market Misinformation

As the use of generative AI accelerates across investment research to summarize companies, screen stocks, and generate investment ideas, these tools are increasingly presenting fabricated information, incorrect earnings figures, and misinterpreted disclosures in persuasive, authoritative language, making inaccuracies difficult to detect before real money is put at risk. 

David Trainer, CEO of New Constructs, a data-driven investment research firm, discusses how research platforms grounded in forensic accounting, proprietary financial models, and a purpose-built ontology can provide more reliable insights. 

1. What are the biggest blind spots emerging as generative AI models are increasingly being implemented in investment research? 

The biggest blind spot is complex and nuanced financial disclosures. Current AI models are very good at reading what companies put in the body of a filing, and very bad at finding what’s buried in the footnotes, which is often where the real economics live. Off-balance-sheet purchase commitments, unusual income and expenses, and residual value guarantees rarely appear as clean line items; they’re disclosed in dense footnote language specifically because that’s harder to parse. Standard AI models summarizing an earnings report are likely to reproduce management’s framing rather than the underlying economic reality. That’s the mistake that has let major companies push trillions of dollars of AI-related debt off their balance sheets without most investors noticing.  

A second blind spot is hallucination: these models are optimized to always produce an answer, not to flag when the data is incomplete or contradictory, so investors get a number without ever learning how uncertain that number really is.  

And a third is groupthink: when many firms lean on similarly trained models pulling from overlapping, unverified data, an error stops being one analyst’s mistake and becomes a shared blind spot sitting in a large number of unrelated portfolios at once. 

2. In what ways is AI-generated erroneous financial analysis becoming the next major source of market misinformation? 

Traditional misinformation spreads because people repeat a claim in print or on TV. While it’s always been true that a lie can travel around the world while the truth is still putting on its boots, AI-generated misinformation spreads potentially much faster than traditional misinformation because it is becoming the primary source of information for so many people, who mistake the fluency and confidence of the AI for trustworthiness.  

When a chatbot misreads an accounting adjustment or invents a figure that isn’t in the filing, the output looks identical, in tone and polish, to research that’s actually correct. I dubbed the term “the 99% rule” to raise awareness of the risk inherent in any data that is not 100% accurate. Even if the model is 99% correct, an investor can’t rely on any of an AI-generated analysis because if investors don’t know which 1% of AI analysis is wrong, then there’s a chance the input on which they rely to build their thesis is wrong.  

Once an error gets copied into a spreadsheet, a research note, or a trading algorithm, it’s no longer contained to one investor’s mistake. It gets redistributed and compounded across every other tool and portfolio drawing on the same or similar sources, which is exactly the dynamic that turns an isolated hallucination into a market-wide misinformation event. Some examples of possible errors: 

• Immediate Scale of Unverified Content: Generative AI allows low-quality or hallucinated financial research to be produced and distributed instantly across social platforms, trading forums, and financial blogs, flooding the market with unverified assertions. 

• The Illusion of Authority: Advanced natural language capabilities give AI outputs a polished, professional tone. Retail and institutional investors alike are easily misled into assuming that a well-formatted report represents rigorous quantitative analysis when it is merely statistical text generation. 

• Systematic Trading Risk: Automated sentiment scrapers and quantitative algorithms ingest AI-generated articles and headlines. If those articles contain hallucinated earnings beats, incorrect guidance figures, or false debt numbers, quantitative models can trigger unwarranted price volatility or momentum shifts. 

• Compounding Accounting Distortions: When AI tools analyze companies using unadjusted accounting figures, they misrepresent fundamental valuation multiples (such as P/E or EV/EBITDA). As these distorted numbers proliferate through automated screening tools, capital is misallocated into economically unprofitable companies. 

Note: The 99% rule and the forensic-accounting/footnote framing are David Trainer’s own analytical framework, previously published by New Constructs, rather than third-party data requiring separate citation. 

3. How should investors rethink their confidence in AI-generated stock recommendations as autonomous AI agents become more capable? 

Capability and trustworthiness are not the same thing, and conflating them is the mistake investors need to understand. A more capable agent that can execute multi-step research or even place trades on its own doesn’t automatically become more accurate or more helpful; it just becomes able to act on its errors faster and with less human review in the loop. FINRA has flagged exactly this in its recent guidance on AI agents, warning that “agents may act beyond the user’s actual or intended scope and authority,” which is a real risk if you’re not 100% confident in the data driving the agent. I think many investors intuitively understand this risk: HSBC’s 2026 survey of nearly 10,000 affluent and high-net-worth investors found that 73% already use AI to explore ideas, but only 12% said AI was the most influential factor in their last investment decision, and roughly half said their ideal approach combines AI with a human adviser. Investors need to be more careful with agents that are more autonomous, not less. The more autonomous the system, the more essential it becomes to know exactly what data it uses, what assumptions drive the model, and where exactly does human oversight live in the final decision-making process. Here are a few specific suggestions: 

• Demand Source-Level Auditability: Investors must move from passive acceptance to a “trust-but-verify” model. Every AI recommendation should be accompanied by direct links to the primary source documents (e.g., exact page numbers and footnotes in SEC filings) and step-by-step mathematical breakdowns. 

• Prioritize Input Quality Over Output Fluency: Confidence should be based on the integrity of the underlying dataset, not how convincingly the agent writes. An agent operating on unvetted web data will generate polished errors, whereas an agent constrained to audited, forensic data can produce actionable insights. 

• Shift Focus to Expectations Investing (Reverse DCF): Rather than relying on AI to make speculative direction calls or price predictions, investors should leverage AI agents to run Reverse Discounted Cash Flow (DCF) models. This approach quantifies the exact operational performance (revenue growth and margins) implied by current stock prices, letting investors judge if expectations are realistic. 

• Require Human-in-the-Loop Validation: Autonomous agents should assist human judgment rather than replace it. Investors should maintain high confidence only in systems where human forensic accounting experts review and validate data inputs before models issue final recommendations. 

4. How can financial firms ensure that AI improves investment decision-making without compromising on transparency and accountability? 

There should be no compromise on transparency and accountability. They are essential to any AI system on which important decisions are based. 

Firms need to treat explainability as a deployment requirement, not a feature to add later. Research from Sage found that 71% of finance leaders would reject an AI tool that was 99% accurate if it couldn’t explain its answers, and 54% would pay a premium for a system that could. 

Financial institutions can safely harness AI for investment decision-making by embedding strict architectural safeguards: 

• Traceable “Marked-Up” Filings: AI interfaces must link every data point directly back to marked-up primary SEC filings. Portfolio managers should be able to click any line item or adjustment and immediately view the original text and footnote context within the 10-K or 10-Q. 

• Fully Auditable Valuation Models: AI outputs must not sit inside a closed system. Platforms should export complete financial models (e.g., into Excel or via API) with transparent calculations, allowing analysts to inspect every formula and tweak underlying assumptions. 

• Domain-Restricted Retrieval (RAG on Validated Data): Instead of allowing AI to search the open web, firms should enforce Retrieval-Augmented Generation (RAG) restricted to proprietary, standardized databases that have undergone rigorous accounting normalization. 

• Fiduciary Human Oversight: Establish institutional protocols where human accounting analysts monitor anomalous disclosure trends, audit edge cases, and validate model inputs to maintain compliance with fiduciary standards. 

5. How does New Constructs’ Robo-Analyst technology transform complex financial disclosures into actionable investment insights without relying on unverifiable information? 

We built the Robo-Analyst to solve these exact problems detailed. It doesn’t summarize a filing; it’s trained to read through it the way a forensic accountant would, pulling data out of footnotes, the MD&A, and other disclosures that legacy datasets simply miss. The Robo-Analyst’s accuracy, transparency and value-added insights have been validated and vetted by: 

• The Journal of Financial Economics, “Core Earnings: New Data & Evidence,” found that a majority of the difference between our data and legacy providers like Compustat comes from real data collection oversights, cases where the legacy dataset simply never captured information relevant to a company’s core earnings. The authors also hand-checked our footnotes adjustments for 350 companies and found our data was 100% correct in every instance. 

• Ernst & Young’s “Getting ROIC Right” explains why our models and the metrics they produce are more reliable than any other research firm. Specifically, on return on invested capital (ROIC), one of Warren Buffett’s favorite metrics, the firm notes that an accurate calculation “requires more diligence than often occurs in some of the common, off-the-shelf ROIC calculations.” Ernst & Young credits our machine learning approach for applying accounting adjustments buried deep in footnotes across thousands of companies as a differentiating factor in our ability to produce superior research.  

• Harvard Business School study, featured in Bloomberg, found that portfolios built on our Robo-Analyst’s buy recommendations earned statistically and economically significant abnormal returns, concluding that Robo-Analysts are “a valuable, alternative information intermediary to traditional sell-side analysts.”  

Every output from our Robo-Analyst is auditable back to the specific filing and adjustment behind it, which is the difference between an insight you can act on and one you have to take on faith. 

6. In what ways does FinSights AI, developed by New Constructs with Google Cloud, differ from general-purpose AI tools when it comes to analyzing companies and generating investment insights? 

Google built FinSights for the purpose of demonstrating the power of Google Cloud’s AI technology when it’s endowed with proven, superior, accurate, and auditable data and investment research. From the press release: 

“AI can be a direct engine for advances in intelligent investing when it’s powered by high quality data,” said Brad Little, Vice President, Global Head of Professional Services at Google Cloud.”  

The power of this combination manifests immediately in the kinds of insights FinSights can reliably deliver.  

If investors ask a general-purpose model “which companies are most likely to miss EPS estimates” or “what are the best stocks in the tech sector,” they will get plausible-sounding platitudes about the largest and most popular companies. FinSights answers those questions based on data and research that’s proven to beat the stock market. In other words, FinSights has a live track to back its answers to those questions. Google Cloud’s decision to build FinSights on our data and technology was a validation of exactly that distinction: an AI agent is only as reliable as what it’s grounded in, and most of the tools on the market today are grounded in the same broad, unverifiable internet content.  

AI based on unreliable data is exactly the risk that investors and regulators need to address and ensure is not embedded in their research process. 

Summary: Developed by New Constructs on Google Cloud, FinSights AI addresses the primary shortcomings of generic consumer chatbots: