# How do you use AI without letting it think for you?

Date: 2022-08-20 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/use-ai-without-letting-think/

**Answer:** Use an AI writing tool as a sparring partner, not an oracle: form your own view first, then ask the model to challenge it, and check at least one of its claims against a source you trust. That order keeps you doing the thinking while still getting the speed and range a model like GPT-3 offers.

A language model such as GPT-3 writes fluent, confident text on almost any subject in seconds. That is exactly what makes it easy to lean on too hard. If the first thing you see on a question is a polished paragraph, it quietly becomes your starting point, and your own thinking shrinks to editing someone else's argument.

Researchers have a name for this pull: **automation bias**, the tendency to favor what an automated system suggests and to discount other information, even correct information. It was studied long before text generators, in cockpits, hospitals and with ordinary spell checkers. A familiar case: the spell checker proposes a wrong word and you accept it without opening a dictionary. A closely related problem, automation complacency, is paying less attention to a system's output because you assume it is reliable.

## Why fluent text deserves extra care

Fluency is not accuracy. In 2021, Stephanie Lin, Jacob Hilton and Owain Evans built TruthfulQA, a set of 817 questions on topics such as health, law and finance, written so that some people would answer them wrongly because of a common misconception. The best model they tested gave truthful answers to 58% of the questions; people managed 94%. GPT-3, for example, repeated the popular claim that cracking your knuckles often may give you arthritis.

The researchers also found that, on these questions, the biggest models tended to be the *least* truthful. Their explanation is that a model trained to imitate text from the web learns to repeat popular falsehoods along with everything else. A bigger model is better at sounding like the internet, mistakes included. So the more natural the answer sounds, the less you can tell from its tone alone whether it is right.

## Three habits, one worked example

Imagine you run a small bakery and are wondering whether to close on Mondays. Here is how to use GPT-3 in the Playground without letting it make the call.

**Habit 1: Write your own view first.** Before opening the tool, spend five minutes writing what you think and why. For example: "Mondays bring in the least money and I'm exhausted. Closing would cost a little revenue but give me a real day off. I'm worried regulars will go elsewhere." This is your anchor. Whatever the model says, you now have something to compare it to.

**Habit 2: Ask the model to argue against you.** Instead of asking "Should I close on Mondays?", paste in your view and ask for the strongest counterarguments.

```prompt
I run a small neighborhood bakery. Here is my current thinking:
"Mondays bring in the least money and I'm exhausted. Closing would cost a little revenue but give me a real day off. I'm worried regulars will go elsewhere."

List the four strongest arguments AGAINST closing on Mondays. For each one, add a question I could answer with my own sales records or customers to test it.
```

This framing keeps the decision yours. The model supplies angles you might have missed, and each angle comes with a way to check it against your own facts instead of its say-so.

**Habit 3: Check at least one claim.** Suppose the response includes a confident general statement, for instance that customers who find a shop closed tend not to come back. Pick the claim that would most change your mind and look for evidence: your own records, a trade association, a fellow shop owner. If you can't verify it, treat it as a hypothesis, not a fact.

The table shows how the roles divide:

| Step | Who does it | What you get |
|---|---|---|
| Form a first view | You | Your reasons, in your words |
| Find objections | GPT-3 | A wider set of angles |
| Test the objections | You | Evidence from your own situation |
| Decide | You | A choice you can explain |

> **Try it:** Choose a decision you are weighing this week. Write your view in three sentences, then use the prompt above with your own details. Pick the one counterargument that surprised you most and spend ten minutes checking it with a source outside the model.

## Which habits keep your own judgment sharp?

Research on automation bias offers a few useful clues. One study found that holding people accountable for the accuracy of their decisions reduced their bias. You can apply that to yourself: before you use a model's output, ask whether you could defend each point if someone challenged you. If the honest answer is "the computer said so", you are not ready to rely on it.

Another finding is that training people with a system that makes deliberate, visible errors works better than simply warning them that errors can happen. The everyday version is to catch the model being wrong now and then. Ask it questions you already know the answer to, from your own field or town, and notice how calmly it can state something false. Those moments recalibrate your trust better than any rule.

A few more habits help over time. Keep a short note of where you changed your mind after using the tool and why, so you can tell persuasion from evidence. Draft important pieces yourself first and use the model to critique them rather than the reverse. And keep the parts of a task that carry your judgment, such as the final recommendation or the choice of facts, in your own hands. Studies of automation bias also show that experts are not immune, and experience alone does not cure it.

## Sources

- Automation bias (revision of 13 February 2022): https://en.wikipedia.org/w/index.php?title=Automation_bias&oldid=1071666121
- TruthfulQA: Measuring How Models Mimic Human Falsehoods: https://arxiv.org/abs/2109.07958v1