AI hallucinations – those convincing but incorrect responses – can be a major challenge when working with AI writing tools.
Now that we’ve had AI tools in healthcare for a couple of years, we’re starting to learn how they can help doctors in various healthcare settings, from patient communication to medical documentation. We now know that the quality of AI-generated responses largely depends on how you structure your AI prompts.
By creating clear, specific and well-thought-out prompts, you can extract more accurate, reliable and actionable information.
Here are some useful strategies to identify and prevent AI hallucinations.
What are AI hallucinations?
AI hallucinations occur when your AI writing tool generates plausible-sounding but factually incorrect information, ranging from subtle inaccuracies to completely made-up details.
Spotting and preventing these errors is crucial if you’re using tools like ChatGPT in healthcare. It relies on a combination of external validation and critical thinking.
Just as you would verify the accuracy of a research summary found in a health blog, feature article or social media post by consulting the original source, the same principle applies here – if you don’t know where the information or claim originated, it’s best not to use it.
When it comes to health claims and medical terms, always trace them back to their original source, and trust your instincts – if something doesn’t sound or feel right, it probably isn’t
How to identify AI hallucinations
AI-generated text can mislead due to its confidence in delivering incorrect information. The responses are very convincing! But that doesn’t mean they are true.
Here are three easy ways to catch those errors.
1. Implement step-by-step verification
One of the most effective ways to catch AI hallucinations is by breaking down the verification process. Prompt your AI writing tool to explain its logic or cite sources. This makes it easier to detect gaps in reasoning or unsupported assertions. Create prompts that:
- Request detailed explanations for each claim
- Ask for specific sources and references
It’s also your responsibility to:
- Look for logical inconsistencies in the reasoning
- Cross-reference important information with reliable sources
Example prompts:
- “Explain how you arrived at this conclusion.”
- “Include your references for each point.”
2. Use confidence indicators
Asking the AI to indicate its certainty about various points can reveal areas where the response might lack reliability. You can also look for qualifying language that might indicate uncertainty or areas where the generated response may express doubt.
Example prompts:
- “Highlight the points that you’re most/least certain about.”
- “Rate your confidence in [part of response]”
3. Break down complex questions
Complex queries often lead to confused responses and increase the likelihood of errors. Split complex questions into smaller, focused prompts to help the AI tool process them more accurately.
This helps you to focus on specific aspects one at a time, build up to more complex topics gradually and verify each component separately.
Example prompts:
Instead of asking, “What are the causes and treatments of diabetes?” break up your question into smaller questions:
- “What are the primary causes of diabetes?”
- “What are common treatments for diabetes in adults?”
- “What are the least common treatments for diabetes in adults?”
Improving accuracy through prompting
Clear, well-structured prompts will always lead to better responses. By creating clear, specific, and well-thought-out prompts, you can extract accurate, reliable and actionable information.
Here are practical tips for crafting better prompts.
1. Set clear constraints and parameters
Specify what the AI should include or exclude, and ask it to note uncertainties.
- Example prompt: “Answer this question based only on high-confidence information. Flag any parts where the answer might be uncertain.”
2. Request supporting examples
Ask the AI to include examples that back up its key points. This makes it easier to assess the validity of its claims.
- Example prompt: “Provide three real-world examples that illustrate the benefits of exercise for heart health.”
3. Use comparative prompts
Encourage the AI to analyse concepts or weigh options, anchoring its response in established knowledge.
- Example prompt: “Compare the benefits and risks of using telemedicine versus in-person consultations for chronic disease management.”
Practical use case: Avoiding AI hallucinations
Suppose you’re researching the impact of AI in healthcare. Here’s an example of three different types of prompts with different specificity levels.
| Vague | More specific | Very specific |
| “How is AI improving healthcare?” | “List three specific applications of AI in healthcare. Explain the data or studies supporting each.” | “Describe three applications of AI in healthcare, focusing on their impact on improving patient outcomes. For each application, cite one peer-reviewed study or credible report published after 2020 that demonstrates its effectiveness. Explain the specific problem the AI addresses, how it works, and any limitations or challenges associated with its implementation.” |
In the example above, the very specific prompt provides even more detailed instructions about the scope, focus, and desired structure of the response.
This prompt sets clear expectations for:
- Number of examples (three).
- Focus (impact on patient outcomes).
- Evidence (peer-reviewed studies or credible reports).
- Context (problem addressed, how it works, limitations).
- Timeframe (after 2020).
By adding these layers of specificity, the prompt minimises ambiguity and encourages a response grounded in relevant, recent and actionable information.
You’ll still need to verify the information yourself, but these types of refined prompts reduce the likelihood of vague or inaccurate responses while yielding actionable insights.
The final word on AI hallucinations
Preventing AI hallucinations requires a systematic approach combining well-crafted prompts, regular verification and critical thinking. If something feels off, it’s definitely worth double-checking the information.
By implementing the strategies outlined in this post, you can significantly improve the accuracy and reliability of your AI interactions.
