Did You Read It Before You Sent It?

Maybe you’ve sent an embarrassing text because autocomplete finished a sentence for you. You noticed after pressing send.

That was a small version. The bigger versions are showing up now.

Someone lets an AI tool tidy up their CV, it adds a skill they do not have, and the first they hear of it is in the interview.

Canadian politician Bill Oliver read an AI response aloud during a speech, including the words, “Here’s a more natural, flowing version of that section…” The moment was recorded and spread online. Ouch.

Or, the deluge of brand-damaging ads featuring bizarre anatomy and impossible physics. Honestly, at this point there are too many to even give examples.

These are the kinds of mistakes we risk when we accept automated output without checking it.

What this is

Researchers call it automation bias. When a machine makes a suggestion, people tend to accept it without checking, even when they would have caught the mistake on their own.

The term is from the 1990s. It was a problem with plane autopilots and hospital systems long before it was a problem with chatbots. What has changed is how many people now work with a machine that suggests things all day.

It shows up in two ways. You skip a check because the machine did not flag anything, or the machine produces something wrong and you go along with it.

Why it happens

Most automation we are used to just works, so we learned to trust automated output. Some of that trust is understandable. We’re used to tools that reliably handle narrow tasks. For instance, a calculator applying math rules to give you a result.

Generative AI works differently. It produces content using learned statistical patterns. It can give us useful answers, but also convincing nonsense. Clear, polished writing and a confident tone don’t make a claim true.

When you create something yourself, you make decisions and correct mistakes as you go. When an AI creates it, that step is gone. Reviewing AI generated work carefully, hunting for what is wrong, is a different skill from creating. Most of us have never had to do it at this volume or speed.

Then there’s the volume. A request for a short answer produces pages of explanation. Every extra paragraph adds something to check. It’s easy to start skimming.

How to reduce it

Keep tasks small. Start with jobs where you already know what right looks like. Tasks like tightening a paragraph you wrote, recreating an image you already have or filling in a form with your own details. You can check these in seconds because the answer is already in your head.

Write the instructions down. If you want an email in a certain tone, under a certain length, with no exclamation marks, say so. Give the tool your standards and constraints before it starts. The clearer the brief, the smaller the pile of things you have to catch on the way out.

Plan the review before you plan the task. Decide what you are checking for, things like numbers, claims, citations. Write it as a short checklist and use it every time.

Measure how long review takes. A person reads about 250 words a minute. Careful review is slower than that. If a tool hands you 2,000 words, you are looking at ten minutes minimum, and that is before you fix anything. Track this for your own work. Once you know your numbers, you can ask the tool to tell you the review time up front, and you can decide whether the task was worth automating at all.

Learn where it struggles on your work. Keep a short record of missed instructions, invented facts, broken constraints and unfinished tasks reported as complete. Use that experience to decide what to delegate.

Resist the pressure to move faster. Count the time spent giving instructions, reviewing and correcting the result. Include any cleanup that falls to someone else. A draft produced in seconds may still leave you with more work overall.

The tools are useful but they cannot replace your judgement.