Phoenix 24
Artificial intelligence can impress, but it cannot replace the gaze of those who truly analyze.
New York, September 2025.
Artificial intelligence models are advancing at a pace that astonishes governments, markets, and citizens alike. They have proven capable of writing complex texts, producing realistic images, translating languages with accuracy, and even designing marketing strategies. Yet, a recent investigation revealed that there is a field where their apparent infallibility collapses: financial analysis. When confronted with the interpretation of economic contexts, the reading of balance sheets, and the projection of scenarios that demand more than statistical training, AI falters and exposes its limits.
A team of researchers at Bernstein Société Générale, led by Venugopal Garre, tested some of the most widely used and publicized systems today: OpenAI’s ChatGPT, Google’s Gemini, X’s Grok, along with Claude, Microsoft’s Copilot, and Meta AI. The test was simple but telling: ask them to act as financial analysts capable of projecting a company’s performance based on market information, historical data, and sector trends. The results made it clear that, while the answers were fluent, the content could not stand up to scrutiny in rigor or accuracy.
According to the report, these models often deliver responses that sound convincing and are neatly structured, but when compared with the interpretation of a trained human market analyst, they lack the sensitivity to detect risks, interpret weak signals, or question the data they receive. In many cases, the AI simply repeated textbook information or gave generic responses without anchoring them in specific contexts. In other cases, it fabricated nonexistent data or projected impossible scenarios, a problem widely known as hallucination.
Garre himself stated bluntly that AIs are good at synthesizing but not at analyzing in depth. Financial analysis requires precisely that: human judgment, intuition, and experience that cannot be reduced to statistical patterns. Detecting a subtle policy shift in a company, anticipating how an election will affect markets, or predicting the moves of a competitor is not a mechanical exercise. It demands historical memory, access to live sources, contact with actors, and a sense of reality that AI cannot yet replicate.
The study also showed that when confronted with questions about corporate balance sheets, AI models often misread numbers or misinterpreted consequences. They failed to distinguish clearly between structural and cyclical debt, or to properly weigh the impact of inflation, interest rates, or regulatory shifts. The absence of context led to projections that looked solid on paper but collapsed when confronted with reality.
This does not mean artificial intelligence has no value in the financial sector. In fact, its ability to process massive amounts of data in seconds makes it an irreplaceable ally for screening tasks, comparisons, and drafting summaries. It can save time by organizing information, flagging obvious patterns, or even producing draft reports. But its weakness appears in the decisive stage: when numbers must be interpreted and translated into strategic decisions. At that point, the human role remains central.
The phenomenon highlights a broader tension: fascination with AI sometimes leads us to forget that it is not intelligence in the full sense, but a statistical tool trained to predict the next word, the next image, the next likely fragment. Its power is astounding, but also limited. The danger arises when its imitative ability is mistaken for true comprehension.
The moral of the study is clear. The human financial analyst, despite biases and limitations, remains irreplaceable in critical contextual reading. AI can assist, but it cannot replace the judgment of someone who combines data with experience, intuition with caution, and memory with foresight. The report concludes that in the short term, rather than seeking to replace analysts with algorithms, the sensible approach is to build collaborations in which the machine brings speed and the human provides judgment.
On a deeper level, the findings serve as a warning for other industries. If AI falters in a field where data is abundant and seemingly structured, what happens in even more uncertain arenas such as politics, ethics, or diplomacy? The illusion that a trained model can answer everything without error must give way to a more sober vision: AI is powerful, but it is not infallible.
What this experiment revealed is not AI’s failure, but a reminder that there are still tasks requiring the human touch. While algorithms calculate, the analyst interprets; while AI projects, the human anticipates; while models predict, experience identifies the improbable. In that gap lies the enduring difference between noise and understanding.
La verdad no se grita: se decodifica en silencio.
Truth is not shouted: it is decoded in silence.