AI Assistants: What They Actually Do, and What Gets Claimed

You have been hearing about this for a while now, from people who say it changes everything and people who say it is worthless. Neither is a useful starting point.

Here is a practical account of what these tools do well, where they fail, and how to tell the difference without any technical knowledge.

What they actually are

A system that has processed an enormous quantity of text and produces plausible continuations of whatever you give it.

That description sounds dismissive and is not meant to be — it explains both the surprising capability and the specific failures. The system is producing what a good answer would look like. Usually that coincides with a correct answer. Sometimes it does not, and nothing in the output distinguishes the two cases.

This is the single most useful thing to understand. The tool is equally confident when it is right and when it is wrong. There is no hesitation, no uncertainty in the tone, no signal.

What they are genuinely good at

Rewriting something you have written. Making a letter shorter, more formal, less angry. You supply the substance and it handles the shape — the task it is best suited to.

Explaining something at the level you ask for. "Explain what an annuity is as if I know nothing about finance." Genuinely useful, and worth verifying anything that matters.

Drafting a starting point. A complaint letter, a note to a landlord, a difficult message. Easier to fix a draft than to face a blank page.

Summarising a long document you already have. Paste in the text and ask what it says. Works well because the material is in front of it.

Finding the word you cannot recall. Describe it and ask. Surprisingly good.

Answering questions about a document you provide. "Does this contract say anything about cancellation?" — with the caveat that it should be checked against the document, which is easy when the document is right there.

Where they fail

Anything requiring current information. These systems have a cutoff, and beyond it they do not know what happened. Some can search the web, and knowing which mode you are in matters.

Precise facts. Dates, figures, names, quotations, citations. It will produce something plausible and specific. It may be invented, and it will look exactly like a real answer. Legal cases and academic references that do not exist have caused people real professional trouble.

Arithmetic and anything with numbers. Improving, still unreliable. Use a calculator.

Anything about your own accounts, money or medical situation. It has no access to your circumstances and no responsibility for the outcome.

Local and specific information. Opening hours, which office to go to, current prices.

Knowing what it does not know. The most important failure. It will not say "I am not sure." It will produce an answer.

The rule that covers all of it

If it matters, verify it. If it does not matter, do not bother.

A rewritten letter does not need verification — you can read it. A summary of a document you have does not need much, because you can check it. A claim about a law, a figure, a medication or a date needs checking against a real source, every time.

Practically: use it for shaping what you already know, and be careful using it for what you do not.

Getting better results

Say what you want it to do, specifically. "Make this shorter" works less well than "Cut this to half the length, keep the first and last paragraphs, and make it less formal."

Give it the material. Paste in the letter, the document, the details. It performs far better with something in front of it than with a general question.

Say who it is for. "Explain this to someone who has never used a computer" produces something different from the same question without that.

Push back. "That doesn't sound right, check it" and "give me the source for that" frequently produce a correction. This is worth doing routinely.

Ask for options. "Give me three versions with different tones."

The privacy question

Do not paste in anything you would not want stored. Bank details, Social Security numbers, medical records, other people's private information.

These conversations are typically retained and may be reviewed. Most services have a setting controlling whether your conversations are used to improve the system — it is worth finding and deciding deliberately rather than by default.

The part worth being alert to

The same technology is being used against people, and it is worth understanding as one subject rather than two.

AI-generated phishing has eliminated the grammar mistakes that once helped people spot fakes. The advice to look for bad spelling no longer works. See fake emails.

Voice cloning has made family emergency calls far more convincing, which is why the defence is now a family word rather than recognising a voice. See the family emergency call.

Real-time deepfake software runs on a standard laptop, so a video call is no longer proof of who you are speaking to.

Understanding what these tools can do is itself the protection. People who know that a voice can be cloned in seconds are not fooled by a voice.

Where to start, if you want to

Pick something low-stakes and real. A letter you have been putting off. A document you cannot face reading. A thing you want explained.

You will get a feel for it in twenty minutes, and that feel — knowing when it is being useful and when it is being fluent — is worth more than any list of instructions.

And the one thing to carry with you: it sounds equally certain whether it is right or wrong. Everything else follows from that.

For additional public information on investor education and financial planning, see Investor.gov.