Guardrails, SEBI and the Human in the Loop — Where AI Ends and Advice Begins

Start with what actually distinguishes the two, because it is not the content and it is not the tone. You can write exactly the same paragraph twice — the same levels, the same reasoning, the same caveats — and have it be research one time and advice the other. What changes is the audience and the intent. Analysis you produce and use for your own decisions is your own business. The same analysis directed at another person, so that they act on it, is a recommendation to that person.

This surprises people because they assume the test is about certainty or about wording. Softening a call does not change its nature. "I think this looks good above 450" delivered to a group of two hundred people who will act on it is not made into education by the word "think". Conversely, a very specific, very confident note written in your own trading journal for your own use is not advice, because there is no other person on the receiving end.

A useful way to feel where the line sits is to ask what the reader is meant to do with the words. If the answer is "understand a method, and then decide for themselves whether and how it applies", that is education. If the answer is "act on this specific thing, on this specific security, now", that is a recommendation, whatever label sits at the top of the page. The word "educational" in a header does not convert one into the other; the function of the text does.

And the point of drawing this line carefully is not paranoia about compliance. It is that the two activities have genuinely different obligations attached to them, and the obligations exist because somebody else's money is at stake. That is the subject of the next section.

In India, activities around securities are supervised by SEBI, the Securities and Exchange Board of India. Two registration categories are directly relevant to the subject of this article: Research Analyst, and Investment Adviser. Broadly, one is concerned with producing and publishing research and recommendations on securities, and the other with advising people about investments, including in relation to their personal circumstances. Both are registration regimes, which means a person or entity carrying on that activity is expected to be registered and to operate under the obligations that come with it.

That is intentionally the level of detail this article gives you, and it is worth explaining why. The exact scope of each category, what triggers a registration requirement, the qualification and certification conditions, the disclosure obligations, the record-keeping expectations and the supervisory arrangements are all set out in SEBI's own regulations and circulars — and they are amended from time to time. A blog post that states them as fixed facts is likely to be wrong within a year, and being confidently wrong about the rules in an article about following the rules is the worst possible failure. Read the source: sebi.gov.in.

What is stable enough to say is why the regime exists. Registration attaches obligations to a named, identifiable person: disclosure of conflicts and of positions held in the securities discussed, defined record-keeping, a certification requirement, an accountable entity behind the words and a route for a grievance to go somewhere. All of that exists because a recommendation about someone else's money is not the same act as an opinion about a film. When it goes wrong, there needs to be a person on the other end.

Registration is not, and never has been, an endorsement. SEBI is explicit that registration and certification do not guarantee performance or assure returns. It is a floor — it means someone accountable is standing behind the words, operating under a framework, with disclosures on the record. This site is written by a SEBI Registered Research Analyst, registration number INH000015297, trade name INVESTOLOGY, and the right response to that sentence is to verify it on sebi.gov.in rather than to trust it because it is printed here.

Here is where most people meet this issue in real life, and almost nobody meets it deliberately. You ask a model something about a stock. The answer is well structured and it comes with a level, a target and a stop. It looks useful. You paste it into a WhatsApp family group, or a Telegram group of a few hundred traders, maybe with a light caveat on top — "not advice, just what the AI said". Two hundred people read a specific instruction about a specific security.

The caveat is doing far less work than you think. Whether an activity amounts to issuing recommendations is not settled by a disclaimer typed above it, and a general-audience message telling people what to buy or sell at what price is exactly the shape of thing the regime is concerned with. Adding "AI generated" changes nothing about the position either. There is no version of this in which the model is the one making the recommendation — a model has no registration, no capacity to hold one, and nothing to answer with.

It gets sharper as scale and money enter. A one-off remark to a friend is a different situation from a standing channel that publishes calls on a schedule, and both are different again from a paid group where subscribers are paying for exactly those calls. Charging for it does not create a permission; if anything it makes what the activity is harder to characterise as anything else. If you are anywhere near this territory, the question of whether you require registration is one to put to a qualified professional before you post, not after.

The distinction to hold on to is simple. Sharing a method is not the same as sharing a call. "Here is how I identify a breakout retest, and here is what invalidates it" teaches something the reader must still apply themselves. "Buy this at 450, target 500, stop 430" tells them what to do. The first is the thing this entire site is built to do. The second is a regulated act, and running it through a model first does not change what it is.

This is the load-bearing idea of the article, and it is worth stating without hedging. When a claim goes out under your name, you made that claim. Not the model. It does not matter that the sentence was generated, that you accepted a suggestion, or that the number came from a summary rather than from the filing. Publication is an act performed by a person, and the person is the one who published.

Compare it with the tools you already use without thinking. A calculator produces a figure and you put it in a note; the figure is yours, and "the calculator said so" is not a defence anybody would offer. A spreadsheet applies a formula you wrote; the output is your work. AI feels different only because it produces prose rather than digits, and prose carries an impression of authorship that a number does not. The impression is false. It is still a tool, and its output is still your work the moment you pass it on.

The practical consequence is that "the AI got it wrong" is not available to you. Not as an explanation to a reader, not as a defence in any process, and not as a way of thinking about your own errors. If an AI-produced figure was wrong and you published it, you published a wrong figure. The verification step you skipped is where the responsibility landed, and that step was always yours.

This is precisely why the audit-trail habit runs through the whole of this module rather than sitting in one article. Every claim traceable to a primary source, every figure checked against the document it came from, and a record of what you verified and when. That habit is not bureaucratic caution — it is the only way a person can honestly stand behind text that a tool helped produce.

The phrase gets used loosely enough to mean nothing. It is often taken to mean a person glanced at the output before it went out, which is a description of proximity rather than of a control. A human who reads without checking is not in the loop; they are next to it. The distinction is the difference between a real safeguard and a story you tell yourself about one.

A human genuinely in the loop does four specific things. They verify — every factual claim is checked against a source the human opened themselves. They judge — deciding what the material actually means, which is an act of interpretation the tool cannot perform because it has no stake in the outcome. They decide — choosing what to publish, what to hold back and what to say differently, which is a decision about consequences. And they take responsibility — their name is attached, and the accountability that follows is real.

None of those four can be automated, and each is a separate job. Verification without judgement produces accurate but useless material. Judgement without verification produces confident nonsense. Deciding without responsibility is how a publication ends up with nobody able to explain why a claim was made. The four have to travel together, and they have to travel with a person.

It is also worth naming the way this control quietly dies, because it almost never fails loudly. Review fatigue: after the fortieth output that turned out fine, the forty-first gets a glance rather than a read, and that is the one with the wrong company name in it. The defences are the ones this module keeps returning to — sample and check specific claims rather than reading for general plausibility, keep the volume low enough that real checking is possible, and remember that the outputs which read most smoothly are the ones that get checked least.

Guardrails are more reliable than intentions, because intentions have to be re-summoned every time and a guardrail only has to be built once. The most valuable one is a physical separation between private and published work. Keep AI-assisted material in a clearly marked working area, and treat the move from that area into anything public as a deliberate crossing with a checklist attached — never a copy-paste that happens because the text was already written.

The second is a standing instruction to the model itself. Every prompt you use for market work can carry a preamble that forbids recommendations, forbids invented figures, forbids claims of assured outcomes and requires an honest "not stated in the source" wherever the source is silent. This does not make the output compliant — nothing makes output compliant except a human checking it — but it removes a category of text you would otherwise have to catch and delete later, and it keeps the tool doing analysis rather than drifting into instructions.

The third is a rule about what you never ask for at all. Do not ask a model for a buy or sell call, even for yourself. Not because the request is forbidden, but because it produces exactly the artefact that is dangerous to have lying around: a well-formatted, confident-sounding instruction that is one paste away from an audience. Ask instead for the things a model is genuinely suited to — summarising a filing, extracting a fact, listing what a document does not say, structuring an argument you then have to defend yourself.

The fourth is a correction habit, and it matters more than any of the preventive ones because errors will get through eventually. Decide now what you do when you find a published mistake: correct it promptly, visibly, at the original location, and say what was wrong rather than editing quietly. A correction issued fast costs a little credibility. A wrong claim left standing costs considerably more, and it is the one people remember.

The division of labour that runs through this entire module is the same one that keeps you on the right side of everything discussed here, and that is not a coincidence. The tool reads, compresses, extracts and structures. The human verifies, judges, decides and takes responsibility. That split is not a temporary limitation waiting for a better model to dissolve it. It is the arrangement that makes AI-assisted market work defensible at all, and a more capable model does not move the line by a millimetre — because the line was never about capability.

For a retail participant researching their own trades, almost none of this is restrictive in practice. You can use these tools for filings, transcripts, screening logic, journal review and a hundred tedious tasks that were pure clerical labour before. The constraint only becomes live at one specific moment: when output stops being for you and starts being for somebody else. That is the moment to slow down, and it is worth knowing in advance that it will not feel like a moment at all — it will feel like pressing send.

If you are considering anything that involves telling other people what to buy or sell, whether free or paid, whether AI-assisted or entirely your own work, treat the registration question as the first question rather than the last one. Read what SEBI publishes at sebi.gov.in, and speak to a qualified professional about your specific situation. It is a much cheaper conversation before you start than after.

And apply all of it here as well. This site is written by a SEBI Registered Research Analyst — registration number INH000015297, trade name INVESTOLOGY — and that is a reason to check the registration on the regulator's own site, not a reason to skip the check. Registration means someone accountable is standing behind the words. It has never meant, and can never mean, that the words are right.