The Honest Limits Rule

There is a version of AI use that creates risk without the user realizing it.

A consultant publishes a report citing a study that the AI fabricated. A doctor sends a patient a letter with a drug interaction detail that sounded right but wasn't. A realtor quotes a zoning regulation that was repealed two years ago. In each case, the professional trusted the output without checking it — and the confident tone of the AI's response gave them no reason to suspect a problem.

This isn't a reason to avoid AI. It's a reason to use it with a clear-eyed understanding of exactly where it falls short.

There are three failure modes that are structural — meaning they apply to every AI tool, not just one brand or one model. They're not going to be "fixed in the next version." They're properties of how these systems work. Once you know them, you can build a simple workflow that protects you from all three.

The three failure modes

It makes things up confidently. The technical term is hallucination — AI generates text that sounds authoritative but is factually wrong. A statistic that doesn't exist. A case citation that was never published. A name attributed to the wrong person. The dangerous part isn't that it makes things up; it's that it makes things up with the same confident tone it uses when it's completely correct. There is no signal to warn you. The text reads identically.

The rule: any specific claim that matters — a number, a date, a regulation, a named source — requires verification from a primary source before it leaves your hands. Not because AI is usually wrong, but because you have no way to tell when it is.

It sounds like everyone else. AI has been trained on an enormous volume of text from an enormous range of sources, and it has learned the patterns of what text "usually" looks like. Its default output is the statistical average of professional communication — competent, clear, and completely indistinguishable from a thousand other documents. Unless you actively bring your voice, your opinion, and your distinctive perspective, the output will read like it was written by no one in particular.

This isn't a flaw in your prompting. It's the nature of a system trained on averages. Your voice is the differentiation, and you have to supply it.

It can't verify its own claims. If you ask AI whether something it said is accurate, it will often tell you it is — and be wrong. If you point out an error, it may "correct" it with a different error. It doesn't have access to an independent truth-checking system. It generates the most plausible-sounding continuation, and "this is correct" is a very plausible-sounding continuation to almost any claim.

This is why asking the AI to double-check its own work is not a reliable verification step. You are the verification layer.

The workflow that keeps you in control

Knowing the failure modes doesn't mean adding a complicated new process to everything you do. For most tasks — brainstorming, drafting, summarizing, rephrasing — the risk is low and the loop is fast. For tasks where specific facts matter, there's a simple five-step loop that handles all three failure modes.

The loop is direct. You write a clear brief (the work covered in "Prompting in Plain Language"). The AI produces a draft. You read it specifically looking for factual claims — numbers, dates, regulatory language, named individuals — and verify those against a primary source. You then rewrite the parts that sound generic until the voice is yours. Then you publish or send.

The feedback step at the end is what makes the loop compound over time. When a particular format or brief structure produces a draft that needs minimal editing, note it. Use it again. Each iteration gets faster.

What not to paste

Everything in this track so far has pushed in one direction: give the AI more context and you get better output. This section is the counterweight, and for doctors, lawyers, realtors, and consultants it may be the single most important rule in the track — some context should never leave your hands in the first place.

De-identify before you brief. Look again at the cardiologist brief from the prompting lesson: "I'm a cardiologist. A patient just received borderline cholesterol results." Notice what makes it work — the role and the situation. The patient's name, date of birth, and record number would add nothing to the quality of the draft; they would only add risk. The same is true for a lawyer's client, a realtor's buyer, a consultant's account. Describe the situation, strip the identity. The de-identified version produces the same letter.

Know your profession's rules and your employer's policy. If you work in healthcare, pasting patient-identifying information into a consumer AI tool can put you on the wrong side of privacy law — in the US, that's HIPAA and protected health information (PHI). Lawyers carry privilege and confidentiality duties. Consultants carry NDAs. And a growing number of employers have an explicit AI-use policy that says which tools are approved and what data may go into them. If you haven't read yours, do that before your next brief — "I didn't know" is not a defense in any of these professions.

Understand the consumer-tier training question. Many AI tools on their free or consumer plans reserve the right to use your conversations to improve their models unless you opt out in settings. Business and enterprise plans typically commit by contract not to train on your data. If your work involves anything confidential, which tier you're on matters more than which model is smartest — it's a question worth asking before the tool ever sees real work.

A simple gate that covers all three: before you paste, ask whether you'd be comfortable if that exact text appeared somewhere outside your control. If the answer is no, de-identify it first — or don't paste it at all. The brief barely suffers; your exposure disappears.

The disclosure question

One question that comes up frequently: do you have to disclose that something was AI-assisted?

The honest answer is: it depends on your profession, your context, and your relationship with the recipient.

In some fields and some contexts, disclosure is legally or ethically required — healthcare, legal, financial advice, academic work. In others, the norm is still forming. In many everyday professional contexts, AI assistance is treated no differently than using spell-check, a grammar tool, or a research database.

What is always true: if you publish something, send something, or present something under your name, you are responsible for its accuracy and its content. The AI being involved doesn't transfer that responsibility anywhere. You are the professional. The draft is yours to verify and own.

That's not a burden — it's actually what keeps you indispensable. The person who can use AI to work faster while still exercising the judgment that the AI can't replicate is more valuable than someone who can't do either.

Putting it together

The four lessons in this track form a complete picture of what it takes to use AI effectively as a non-technical professional.

First, the right mental model: AI is a capable assistant with a specific skill set, not an oracle and not a vending machine. Understanding what it actually does — and what it doesn't — lets you use it deliberately instead of being surprised by it.

Second, the brief: context plus task plus format plus an example gets you a first draft worth editing. The investment in the brief is always returned in time saved on revision.

Third, picking your tool: choose the mainstream assistant that fits how you already work, start on a real task this week instead of researching forever, and remember that switching later is cheap because the skills transfer across every tool.

Fourth, the honest limits: know the three failure modes, run the verify loop for anything factual, and add your voice and judgment before anything leaves your hands — and just as importantly, de-identify anything sensitive before it goes in. Confidential client and patient details don't belong in a consumer AI tool.

None of this requires technical skill. It requires the same professional discipline you already apply to every other tool and every other assistant in your work.

For professionals who want to go beyond the basics — building systems, automating workflows, and developing the kind of compounding AI advantage that actually changes what's possible in your practice — academy.jeremyknox.ai is built exactly for that next level.