Can ChatGPT pick stocks? What AI is actually good for in investing, and what it is not

Type “which Indian stocks should I buy for 2026” into any chatbot and you will get a list. It will contain the same six large companies everyone already owns, a paragraph about diversification, and a disclaimer. Type the same question tomorrow and the list will change slightly, for no reason. This is the first thing to understand about using these tools for money: they are not looking anything up. They are producing the most plausible-sounding text.

What a language model is doing

A large language model is a system trained on an enormous amount of written text to predict what words come next. That makes it remarkably good at explaining, summarising, translating and rearranging language. It does not make it a database. Unless the tool is explicitly connected to live data, it has no idea what the Nifty closed at today, it does not know your portfolio, and its “knowledge” of a company’s numbers is a blurry memory of documents that were old when it was trained.

Worse, when it does not know, it does not say so. It fills the gap with something that sounds right. Ask for the price to earnings ratio of a mid-sized company and you will get a precise-looking number, often with two decimal places, that you should treat as a guess until you have seen it on the page of an annual report. Ask for sources and it may invent a plausible-looking SEBI circular or a research paper that does not exist. This habit has a polite name, hallucination, and it is not a bug that is about to be fixed; it is a property of how the thing works.

The study everyone quotes

In 2023 two researchers at the University of Florida published a paper asking whether ChatGPT could forecast share price movements. They fed it news headlines, asked whether each was good or bad for the company, and found that the model’s scores predicted the next day’s returns in their sample better than older sentiment tools. The paper is real and interesting, and it is also a backtest: a test on past data by people who knew the answers, before transaction costs, before every hedge fund did the same thing and traded the edge away. Nothing in it suggests that a retail investor asking a chatbot for picks will beat the market. It suggests that headlines carry information, which markets already knew.

Where it genuinely helps

The useful work is reading, not predicting.

Give the model a document and ask it questions about that document. An annual report, a results presentation, a SEBI consultation paper, an RBI circular: upload it and ask what the auditor’s opinion was, which risk factors are new this year, what the related party note says. The companion piece in this section, on reading an annual report in twenty minutes, is a complete workflow for this, including the verification step that makes it safe.

Ask it to explain anything in this course at the level you need. “Explain free-float market capitalisation as if I am fifteen” produces a good answer, because the explanation exists in a thousand textbooks and the model has read them all.

Use it as an adversary. Paste your reasons for buying a company and ask for the five strongest arguments that you are wrong. It will not know the company’s latest quarter, but it knows every way a thesis like yours has failed before, and a list of those is worth more than a list of picks.

Ask it to write a small script that screens the exchange’s data for the ratios in the P/E, P/B and ROE chapter, or to turn a messy table you pasted from a PDF into a clean one. Ask it to draft the questions you want to send to a company’s investor relations team before the annual general meeting. These are clerical tasks, and clerical tasks are where the hours go.

Where it does not help

Anything that depends on the future. A model has no more information about next quarter’s profit than you do, and less about this one. Anything that depends on current data, unless the tool is visibly browsing and citing the page. Arithmetic across a long document, which it will get right most of the time and wrong at the moment it matters. Your own circumstances: it does not know your income, your tax bracket, your existing holdings or what you need the money for, which is the whole of what advice is.

The rules apply to the machine too

Under Indian law only a person or firm registered with SEBI as an investment adviser may give personalised investment advice, and only a registered research analyst may publish recommendations. A chatbot is neither, and the company behind it is careful to say so in the disclaimer you skipped. The same rules sit behind SEBI’s actions since 2024 against unregistered “finfluencers” and the restrictions on regulated firms associating with them. SEBI has also started consulting on how brokers and fund houses may use artificial intelligence, and the thread running through its proposals is that the regulated firm, not the software, is responsible for the outcome. For an individual the implication is simple. If a tool gives you a tip, nobody is accountable for it, and nobody will be when it goes wrong.

Rules for using it well

Give it the document, never just the question. Ask for page numbers with every number and check the ones that would change your decision. Ask it what it does not know; a model asked directly will usually admit the gaps it would otherwise paper over. Keep your account details, passwords and personal statements out of it; public filings are public, your demat statement is not. And treat whatever comes back as the first draft from a bright, tireless junior who has never once been held responsible for being wrong.

Used that way, AI shortens the time it takes to understand a company from a weekend to an evening. It does nothing at all to the uncertainty about what the company earns in five years, and that uncertainty, not the reading, is where investment returns come from.