One generalist against twenty specialists

The most useful fact on this page is a dated one. OpenAI's Usage Policies, effective 29 October 2025, prohibit the "provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional." The same update restricts automated high-stakes decisions in financial, credit, insurance, legal and medical contexts without human review.

It is worth being precise about what that does and does not mean, because the internet has largely got it wrong. ChatGPT was not banned from discussing law or medicine, and in practice it still answers a great deal when asked directly. What changed is the policy position on tailored advice - guidance applied to your specific circumstances - and the practical effect most people notice is a model that hedges earlier and more often on exactly the questions they came with.

If you have found yourself getting "you may want to consult a professional" where you used to get an answer, that is the shape of it.

This is a structural difference, not a quality one. A single model serving every subject at once has to be safe across all of them simultaneously. It cannot know whether the person asking about a security deposit is a tenant, a landlord, a law student or a journalist, so it answers in a way that is defensible for all four. That produces text which is technically correct and practically useless.

A copilot configured for one field starts from a narrower assumption and can therefore be concrete. Ask both about a withheld deposit and the difference is not eloquence - it is that one of them names the statute, states the deadline in your state, and offers to draft the demand letter.

None of this makes the specialist smarter. It makes it narrower, which is the entire point.

The second difference has nothing to do with model quality. ChatGPT lives in a tab. Whatever you want help with has to be copied into it, and the context - what page you were on, what the rest of the document said - has to be re-explained.

Copilotly's browser extension runs on the page the problem is already on. Select a clause in a lease, a line in a lab result, a paragraph in an offer letter, and hand it to the relevant specialist without leaving. For the kind of work this product is for, that is often the larger practical difference.

ChatGPT Plus is $20/month, with Pro tiers at $100 and $200. Copilotly Pro starts at $4.99/week for every copilot in every domain, and the free plan needs no card.

The honest comparison is not one subscription price against another - if you want one general assistant, ChatGPT is both cheaper and better at being one. The comparison that matters is $29 against what the questions cost otherwise: a legal consultation at $350/hour, a financial advisor at $250, a tax preparer at $400 a filing.

The hedging is not a limitation of the model. It is the correct behaviour for a product that has to be safe across every subject at once, for every user, in every jurisdiction, with no idea which of those it is talking to.

Two things produce it. The first is policy: OpenAI's usage policies restrict the provision of tailored advice requiring a licence without appropriate professional involvement, and comparable terms exist at the other major providers. That is a reasonable position for a general platform and it means the shape of the answer you get on a legal, medical or financial question is set partly by liability rather than entirely by capability.

The second is calibration. A model tuned to be helpful across everything cannot know whether the person asking about a notice period is a tenant, a landlord, a law student or a journalist - so it answers in a way that is defensible for all four. The result is accurate, general, and unusable for the specific decision in front of you.

What that costs is not a worse answer. It is a wasted turn. You ask about your situation, receive the general rule plus a recommendation to consult a professional, and are exactly where you started - except now you have spent the time and formed the impression that AI is not much use for this. That impression is the real damage, and it is why so many people stop asking after two attempts.

A copilot configured for one field can ask instead of hedge. Which state, when did the notice arrive, what does the code in the corner say. Those are the questions a practitioner asks first, and asking them is what turns an under-specified question into an answerable one.

The obvious objection is that you can get the same result by starting with "act as an employment lawyer in California". For a single question, largely true. The reason it stops working is mechanical rather than a matter of prompt quality.

An instruction you type is one message in a growing context. As the conversation fills with your situation - the dates, the documents, the follow-ups - that first instruction becomes a smaller and smaller share of what the model is attending to. By message fifteen it is competing with several thousand tokens of your problem, and the generalist calibration reasserts itself.

That matters because message fifteen is where the useful part of the conversation lives. The first few turns are establishing the situation. The turns where it becomes specific enough to act on are exactly the ones where a typed prompt has lost most of its influence, which is why people describe long ChatGPT conversations as "drifting".

A configuration is applied at every turn rather than sitting in the context competing for attention. It also carries things a prompt cannot reach at all: which model is used, what the temperature is, how follow-ups are structured. You cannot prompt your way into a different model.

And there is the ordinary practical point. Nobody actually retypes a carefully written system prompt every time. The version people reconstruct from memory is worse than the one they wrote once, and much worse than one written by someone who structured it around how the field actually works. How it works sets out the full mechanics, including where the specialisation stops helping.

Often, and it is worth being specific rather than gracious about it.

For code, ChatGPT is excellent and Copilotly is not trying to compete - if you are debugging, refactoring or scaffolding, use it or a dedicated coding tool. For open-ended creative work, brainstorming and writing at length, the general model's breadth is the advantage rather than the limitation.

For anything needing live web results, it has browsing and Copilotly does not. For image generation, data analysis on an uploaded file, or voice, it has capabilities that simply do not exist here. For learning a subject broadly rather than resolving a specific situation, a generalist is the right shape.

The claim worth making is narrow. On a professional question with a jurisdiction, a deadline and a consequence attached, a copilot configured for that field asks better questions, produces something you can send, and states its own limits precisely rather than gesturing at them. That is a real difference and it is not a claim about general intelligence.

Most people who use both end up using them for different things, which is the sensible outcome. If your questions are mostly the first kind, ChatGPT alone is the right answer and you should not pay for two subscriptions.