Countering Hallucinations Through Prompt 'Design': Practical Methods for Installing Safety Mechanisms

Hello, I am Poke-go, and I research prompts for writing high-quality articles using generative AI.

As the time spent using AI increases,

"Are the numbers generated by the AI really correct...?"
"When I left research to the AI, it included plausible-sounding lies."

Don't these anxieties keep growing?

The phenomenon where AI generates plausible but factually incorrect information is called hallucination.

This cannot be completely eliminated, no matter how high-performance a model you use.

But in reality, the root cause of AI lying is surprisingly simple.

AI lacks a standard for determining when it 'doesn't know'.

Without a standard for judgment, it has no choice but to piece together the information it has to produce a 'plausible answer'.

That is precisely why providing that standard for judgment in the prompt in advance is the starting point for countermeasures.

I personally add a safety mechanism to my article creation prompts that says, 'Do not write information for which you cannot provide a source.'

Just by doing that, the errors found during fact-checking clearly decrease.

In this article, I will specifically introduce how to incorporate such safety mechanisms into your prompts.

You can start by 'just adding one line,' so please feel free to give it a try.

What is the 'root cause' of AI hallucinations?

Many people who use AI have likely had the experience of receiving information that is factually incorrect.

To begin with, why does AI make mistakes with such confidence?

First, we will determine the direction of our countermeasures by understanding the root cause.

AI lacks the criteria to answer 'I don't know'

When you ask an AI a question, it will always return some kind of answer.

It rarely says, 'I don't know.'

This is because the criteria for determining 'I don't know' are not built into the AI's mechanism.

AI language models operate on a mechanism that predicts the most natural word to come next based on vast amounts of text data.

In other words, what is important to an AI is not 'whether it is a fact,' but 'whether it is natural as a sentence.'

Therefore, even if the information is ambiguous, it will string together 'probabilistically likely expressions' to generate a plausible-sounding sentence.

AI does not lie out of malice; it simply does not have 'rules for not lying.'

If you set judgment criteria in your prompt, the AI can become honest.

Conversely, this means that if you explicitly provide the criteria for judgment in the 'prompt,' the AI will be able to honestly pause.

The same can be said for human teams.

When I was running a web media outlet, I commissioned articles from over 100 writers.

Depending on the writer, some manuscripts would come back containing uncertain information...

So, I established a rule: 'Do not write about anything you haven't experienced.'

The same applies to AI; simply writing 'If you are not certain, please answer "I don't know"' in the prompt changes its behavior significantly.

In one study, data was reported showing that adding such instructions 'reduced the hallucination rate from 53% to 23%.'

'Do not lie' instead of 'providing criteria for judgment'.

This approach is the essence of hallucination countermeasures in prompts.

Related article:

Three Safety Mechanisms for 'Uncertainty Protocols' to Incorporate into Prompts

So, what specific instructions should you incorporate into your prompts?

Let me introduce how to do that.

I call these instructions 'Uncertainty Protocols.'

The name might sound grand, but what you actually do is simple.

How to behave when you are not certain—teach this to the AI in advance.

It is just that simple of a mechanism.

Reservation instruction: "If you are not certain, answer that you do not know."

This is the simplest and most effective safety mechanism.

By simply giving the AI permission to say "I don't know" when it doesn't know something, the frequency of it forcing a fabrication of information decreases.

This is because, without instructions, the default behavior of an AI is to "answer something."

The background behind why reservation instructions work is related to the AI's learning process.

Since AI learns a massive number of patterns for "returning an answer to a question," the behavior of "not answering" is difficult for it to select as an option.

That is precisely why you need to explicitly provide the option of "not answering" in the prompt.

In an actual prompt, you add a sentence like the following.

確信度が低い情報については、推測せず「この点については確認が必要です」と明記してください。

It is just this much, but it can significantly suppress patterns of "answering without evidence."

Confidence labels: Rules for distinguishing between facts and guesses

The second safety mechanism is a rule that requires the AI to label its answers with a 'confidence level'.

In fact, without labels, it is impossible to distinguish which parts of the information output by the AI are facts and which are guesses.

This is a very dangerous situation for the reader.

This is because the more plausible a text sounds, the more it appears as if 'everything is a fact'.

The confidence labeling mechanism instructs the AI to 'indicate how confident you are about each part of your answer'.

For example, you can write this in your prompt.

各情報に対して、以下の3段階で確信度を付記してください。
- 【確実】十分な根拠がある事実
- 【推測】根拠はあるが確定ではない情報
- 【未確認】裏取りが必要な情報

Just having this label makes it much easier to judge 'what to believe' when receiving output from the AI.

Source Requirement: 'Do not write information for which you cannot provide evidence'

The third safety mechanism is to instruct the AI 'not to write information for which you cannot provide evidence'.

Hallucinations are likely to get mixed in because the AI outputs information while keeping the boundary between 'knowledge' and 'guesswork' ambiguous.

Therefore, by establishing a rule to 'output information as a set with its evidence,' information without evidence is naturally excluded.

You can instruct the AI in the prompt as follows.

情報を記載する際は、必ずその根拠や出典を併記してください。
根拠を示せない情報は記載しないでください。

By adding this single sentence to your prompt, the AI will be able to honestly report, 'I could not verify this point from public information'.

Requesting citations is not a 'mechanism for finding lies,' but a 'mechanism for making the AI pause before it lies.'

By combining this with hold instructions or confidence labels, it should function as an even more powerful safety mechanism.

Implementation examples of uncertainty protocols for different purposes

Now that you understand the mechanisms of the three safety devices, let's look at how to use them differently in actual work.

The key is to adjust the 'tightness' of the safety devices depending on the application.

Business research and report creation: Setting examples for strict mode

For research and reports used in work, the accuracy of information is the top priority.

Errors in figures or descriptions that differ from facts can lead directly to the risk of making incorrect decisions.

In such situations, a 'strict mode' that combines all three safety devices at maximum strength is recommended.

Below is an example of a prompt for business reports.

あなたは正確性を最優先にする調査アシスタントです。以下のルールを厳守してください。
- 確信がない情報は「確認が必要です」と明記し、推測で埋めないこと
- 各情報に【確実】【推測】【未確認】のラベルを付記すること
- 根拠を示せない数値やデータは記載しないこと

Please use this template as a base and customize it to fit your own work.

Learning and creation: Points for adjusting to flexible mode

For brainstorming for learning purposes or generating ideas for creative work, there are times when you want the AI to think freely.

If you apply strict mode as-is here, the AI might become inhibited and stop producing even useful ideas.

Therefore, in such situations, let's switch to a 'flexible mode' where you only loosely set hold instructions and confidence labels.

自由にアイデアを出してください。
ただし、事実として述べる情報と、あなたの推測・提案は明確に区別して表示してください。

The point is to remove the requirement for sources while maintaining only the 'distinction between fact and speculation'.

With just this, you can manage the risk of hallucinations without compromising creativity.

Please try to be conscious of using it differently depending on your purpose.

Note: Prompts alone cannot completely prevent it

Up to this point, I have conveyed the effects of the uncertainty protocol.

However, to be honest, it is not possible to prevent all hallucinations with prompts alone.

Finally, I will also share the limitations of the protocol and the mindset to compensate for them.

Knowing the limitations of the uncertainty protocol

If you use the uncertainty protocol, the quality of the AI's output will likely be significantly improved.

However, you also need to be aware that there are some structural limitations.

First, if there are errors in the AI's training data itself, it cannot be prevented by prompts.

This is because the AI does not have a means to verify whether 'what it knows is correct'.

Also, as conversations grow longer, there is a tendency for the effectiveness of the instructions set at the beginning to fade.

When using it in practice, you may encounter situations where safety mechanisms that worked with high precision in the first turn gradually lose their effectiveness as the interaction continues.

In other words, the uncertainty protocol is not a 'panacea,' and it is appropriate to position it as the 'first line of defense' for managing risk.

It is precisely because you use it with an understanding of its limitations that the protocol can demonstrate its true power.

Increasing reliability with the idea of 'layering' safety mechanisms

In addition to prompt safety mechanisms, layering safety mechanisms from different angles further reduces the risk of hallucinations.

For example, the following approaches are practical:

• Self-fact-checking:
  Having a human verify the AI's output at the end. The simplest and most reliable safety mechanism

• Comparison with multiple AIs:
  Asking the same question to another AI and checking the consistency of the answers

• Utilizing RAG (Retrieval-Augmented Generation):
  A mechanism that guides the AI to provide evidence-based answers by having it reference a reliable database

I personally use a combination of these when writing actual articles.

The important thing is not to try to prevent everything perfectly with just one, but to adopt the mindset of 'increasing defensive strength by layering multiple safety mechanisms.'

Start with the uncertainty protocol introduced in this article, and combine it with other measures as needed.

Summary: Give AI a 'standard for judgment'

In this article, for those feeling anxious about AI hallucinations, I have discussed:

I have covered the above while incorporating the practices of the author, who continues to research prompts for AI article creation.

The essence of countering hallucinations is not to doubt AI output, but to proactively incorporate a 'mechanism that prevents lying' into the prompt.

If you provide criteria for judgment, the AI will be able to honestly pause and say, 'I don't know'.

Even if it is difficult to prevent them perfectly, by layering safety mechanisms, you should be able to steadily build a relationship of trust with AI.

Start by adding the sentence 'If you don't know, answer that you don't know' to your usual prompts.

Thank you for reading until the end 😇

✅ Click here for Poke-go's content