Can AI Pick Stocks? What ChatGPT Gets Wrong About Indian Markets

A hundred million weekly users, and some of them want a tip#

In February 2026, ahead of the AI Impact Summit in Delhi, Sam Altman wrote in the Times of India that India had crossed 100 million weekly ChatGPT users, the first user figure OpenAI had put out for the country. That makes India its second-largest market. Nobody publishes a breakdown of what those hundred million are typing, but anyone who has sat in a WhatsApp group of retail traders knows that "which stock should I buy tomorrow" is in there somewhere.

Six months later, on 19 August 2026, SEBI chairman Tuhin Kanta Pandey told the FICCI Capital Markets Conference in Mumbai that the regulator would "shortly be issuing guidelines for responsible use of AI/ML in our markets", with kill-switch and humans-in-the-loop controls. He was speaking to a market of 14.9 crore unique investors.

Between those two facts sits a question worth answering carefully. Not "is AI smart", which is unanswerable, but something narrower. Given what the published research shows, and given how Indian markets actually work, what can a chatbot do for you and where does it fall over?

What a language model is doing when it "analyses" a stock#

A large language model is a statistical text predictor. It was trained to guess the next fragment of text, called a token, given everything before it. Trained on enough text, that turns into something that reads like reasoning. It is not a database and it does not look things up unless you connect it to something that can.

Three properties matter here.

The first is the training cutoff. Everything the model learned came from text gathered up to a certain date. Ask about anything after that and it either says it does not know, or it guesses in a confident voice. That guessing has a name: hallucination. The model produces fluent text that is simply untrue, and it has no internal signal telling it which is which.

The second is retrieval, usually called RAG, or retrieval-augmented generation. Give the model a live search tool or a document to read and it works from current material. That helps, but it does not solve the problem, as the benchmark below shows.

The third is that stock prices are close to a fair game in the short run. If a piece of public information reliably predicted tomorrow's price, traders would act on it until it stopped predicting. Any edge that is easy to copy is an edge that dies. Keep that in mind when reading the research, because the research itself says so.

The evidence, read honestly#

The published work on AI stock market prediction says neither "it beats the market" nor "it is useless". It says something more specific.

Read the first and second rows together and you get the honest summary. A good model is reliable at judging the tone of news, which is a language problem. It is unreliable at pulling a number out of a filing correctly, which is a precision problem. Investing needs both.

The third row is the one worth pausing on. That paper was reported around the world as proof that AI beats analysts, and its authors pulled it themselves when the numbers would not replicate. Nobody has alleged bad faith, and the withdrawal notice says the review is ongoing. But if you have seen "GPT-4 outperforms human analysts" quoted at you, that is where it came from, and it is currently not standing.

Five things it gets wrong about India specifically#

The research above is mostly American. The failures compound when the subject is Indian.

Stale law is the biggest one. The Income-tax Act, 1961 stopped being the operative statute when the Income-tax Act, 2025 came into force on 1 April 2026. A model working from older text will cheerfully quote section 112A for capital gains or section 80C for deductions, neither of which is the operative provision any more. Long-term gains on securities transaction tax paid listed equity now sit in section 198, taxed at 12.5% above a Rs 1.25 lakh exemption, with general long-term gains in section 197. The Rs 1 lakh at 10% that dominated a decade of internet writing went in July 2024, and it is still all over the training data.

Second, contract mechanics change by circular. SEBI's circular of 26 May 2025 required every equity derivative on an exchange to expire on a Tuesday or a Thursday. NSE took Tuesday, BSE took Thursday, effective 1 September 2025. Thursday-expiry Nifty options are a fact of the pre-2025 internet and no longer a fact of the market.

Third, prices go stale and corporate actions are invisible. A model quoting a share price from memory is quoting history. Worse, it may quote a pre-split or pre-bonus price as though it were current, because nothing in the text told it an adjustment had happened.

Fourth, home bias, and it is not your home. The PLOS ONE study measured what happens when you ask a chatbot for a portfolio: over 93% American equity. An Indian investor asking a general-purpose assistant for "a diversified portfolio" is likely to be handed someone else's country.

Fifth, fabricated specifics. Ask for the promoter pledge percentage in a mid-cap's latest shareholding pattern and you may get a plausible number that does not appear in any filing. The FinanceBench result is the warning: even with the document in front of it, the model got most questions wrong or declined.

What the regulators have actually said#

SEBI published its consultation paper on guidelines for responsible usage of AI/ML on 20 June 2025, with comments open to 11 July. It proposes five areas of obligation for market participants whose AI touches clients: model governance, investor protection and disclosure, a testing framework, fairness and bias, and data privacy. Outsourcing the model does not outsource the liability. Firms stay responsible, and a senior officer must own the system's whole life cycle. The final circular had not been issued as this was written.

The Reserve Bank got there first for lenders. Its FREE-AI committee reported on 13 August 2025 with 26 recommendations across six pillars and seven guiding principles it calls sutras. The committee's survey of regulated entities is the part investors should read. Around 21% were using or building AI systems. Only a third had board-level oversight. Just 10% had bias-mitigation protocols in place, 18% kept audit logs, and 14% monitored model performance in real time. Institutions with compliance departments are still working this out.

For anything that places orders, SEBI's circular of 4 February 2025 on retail participation in algorithmic trading sets the rules: broker APIs locked to vendor-specific keys and whitelisted IPs, OAuth two-factor authentication, algo providers empanelled with the exchange, exchange approval before an algo is offered, and every algo order tagged with an exchange-issued unique identifier. The timeline was later extended. Wiring a chatbot to your trading account through an unapproved route is not a grey area.

And the boring rule that covers all of this: under regulation 3 of the SEBI (Investment Advisers) Regulations, 2013, nobody may act as an investment adviser without registration. A chatbot is not registered. It carries no fiduciary duty, no suitability assessment and no grievance redress. That matters more than it sounds, because the supply of the real thing is thin. A SEBI board memo from October 2024 recorded 927 investment advisers and 1,380 research analysts as on 31 August 2024, against more than 12 crore investors. The vacuum is why people ask the machine.

Where it earns its keep#

None of this makes the tool worthless. It shifts what you use it for.

A language model is good at language. Paste in a 200-page annual report and ask what the auditor's qualifications say, or what changed in the related-party transactions note. Ask it to translate "deferred tax asset" or "contingent liability" into plain words. Ask it what questions you should be asking about a company before you buy. Ask it to explain why an option lost value on a day the stock did not move. These are comprehension tasks, and it does them well, especially when you give it the document rather than trusting its memory.

What it cannot fix is the thing that actually costs Indian investors money. SEBI's study published on 20 August 2026 found that 87.7% of individual traders in equity derivatives lost money in FY26, with 92% of the aggregate losses coming from options. Those losses are not caused by a shortage of analysis. They come from position sizing, leverage and the urge to trade. A chatbot that answers instantly and never sounds unsure is, if anything, likely to make that worse. That last sentence is interpretation, not something the study measured.

Key takeaways#

1. The headline claim that GPT-4 beats human analysts rests on a paper its own authors withdrew in February 2025 after a co-author could not replicate the data.
2. Language models read news sentiment well. Lopez-Lira and Tang found roughly 90% hit rates on the initial reaction to headlines, while warning that the edge shrinks as more people use the same tool.
3. On precision tasks they are weak. FinanceBench found GPT-4-Turbo with retrieval got 81% of open-book financial questions wrong or refused them.
4. Indian answers go stale fastest: the Income-tax Act, 2025 replaced the 1961 Act on 1 April 2026, and NSE expiries moved to Tuesday from 1 September 2025.
5. SEBI's AI guidelines are pending, its algo trading framework is already in force, and no chatbot is a registered investment adviser.

Frequently asked questions#

Can AI predict the stock market? Not in any dependable way. The published evidence shows short-lived predictability from news sentiment, which decays as adoption spreads, and poor accuracy on precise financial facts.

Is it legal to use ChatGPT for stock research in India? Yes. Researching for yourself is not regulated. Selling AI-generated recommendations to others without SEBI registration is, and connecting a bot to your broker's API to place orders falls under the algo trading framework.

Why does ChatGPT get Indian tax rules wrong? Most of its training text predates the Income-tax Act, 2025, which took effect on 1 April 2026, and the 2024 change to capital gains rates. Older rules dominate the internet, so the model repeats them.

Are AI-managed funds available in India? Some schemes and platforms use machine learning in their processes, and SEBI has required mutual funds to report their AI and ML systems quarterly since May 2019. There is no Indian equivalent of a fully AI-picked equity fund with a long public record.

Will SEBI ban AI in the markets? Nothing in the consultation paper or the chairman's remarks suggests a ban. The direction is disclosure, accountability, human oversight and a kill-switch.

Does connecting ChatGPT to live market data fix the problem? It fixes staleness, not precision. FinanceBench tested retrieval-equipped models with the source documents supplied, and most answers were still wrong or refused.

What is the single safest way to use it? Give it the document and ask it to explain, not to decide. Then verify any number it quotes against the filing, the exchange or the regulator's own site.

Glossary#

Large language model. A system trained to predict the next fragment of text. It is optimised for fluency, and accuracy comes along only to the extent that fluent text in its training data happened to be true.

Token. The unit of text a model works in, roughly a word piece. Models predict tokens, not meanings.

Training cutoff. The date after which the model saw no data. Anything later is unknown to it unless supplied at the time of asking.

Hallucination. Confident, fluent output that is factually wrong, produced without any signal to the model that it is wrong.

RAG, retrieval-augmented generation. Feeding the model live search results or documents so it answers from supplied material rather than memory.

Backtest. Running a strategy over historical data. Easy to make look good by accident, which is why out-of-sample testing matters.

Sharpe ratio. Return above the risk-free rate divided by volatility. A measure of return per unit of risk taken.

Kill-switch. A control that halts an automated system immediately. SEBI's chairman named it as a requirement in the coming AI guidelines.

References#

1. TechRepublic, India hits 100M weekly ChatGPT users, becoming OpenAI's second-largest market, 16 February 2026, reporting Sam Altman's column in the Times of India
2. News Arena India, SEBI to soon issue AI/ML guidelines for capital markets, remarks by chairman Tuhin Kanta Pandey at the 23rd FICCI Capital Markets Conference, 19 August 2026
3. Securities and Exchange Board of India, Consultation paper on guidelines for responsible usage of AI/ML in Indian securities markets, 20 June 2025
4. Alejandro Lopez-Lira and Yuehua Tang, Can ChatGPT forecast stock price movements? Return predictability and large language models, arXiv:2304.07619
5. Patronus AI, FinanceBench: a new benchmark for financial question answering, arXiv:2311.11944, November 2023
6. Alex G. Kim, Maximilian Muhn and Valeri Nikolaev, Financial statement analysis with large language models, arXiv:2407.17866, withdrawn 20 February 2025
7. Philipp Winder, Christian Hildebrand and Jochen Hartmann, Biased echoes: large language models reinforce investment biases and increase portfolio risks of private investors, PLOS ONE 20(6), 27 June 2025
8. International Monetary Fund, Artificial intelligence can make markets more efficient and more volatile, summarising Chapter 3 of the Global Financial Stability Report, October 2024
9. Reserve Bank of India, Committee on a Framework for Responsible and Ethical Enablement of Artificial Intelligence, report of 13 August 2025, as summarised by Dvara Research
10. Securities and Exchange Board of India, Safer participation of retail investors in algorithmic trading, SEBI/HO/MIRSD/MIRSD-PoD/P/CIR/2025/0000013, 4 February 2025, and extension of timeline, 30 September 2025
11. Securities and Exchange Board of India, Final settlement day (expiry day) for equity derivatives contracts, SEBI/HO/MRD/MRD-TPD-1/P/CIR/2025/76, 26 May 2025
12. Securities and Exchange Board of India, SEBI (Investment Advisers) Regulations, 2013, last amended 25 November 2025
13. Securities and Exchange Board of India, Review of regulatory framework for Investment Advisers and Research Analysts, board memorandum, October 2024
14. Securities and Exchange Board of India, Study: profitability of individual traders in the equity derivatives segment (FY25 to FY26), 20 August 2026, as reported by Moneylife
15. Income Tax Department, Income-tax Act, 2025 comes into force from 1 April 2026, section 197 and section 198
16. Securities and Exchange Board of India, Reporting for AI and ML applications and systems offered and used by mutual funds, SEBI/HO/IMD/DF5/CIR/P/2019/63, 9 May 2019

This article is journalism, not investment advice. Regulatory positions quoted are as at 28 September 2026 and SEBI's final AI/ML guidelines had not been issued at that date. Verify rules with SEBI, the RBI and the Income Tax Department, and consult a SEBI-registered adviser before acting.