AI chatbots can give you different answers when you ask them a question. Some are funny, some are useful, and some are entirely fabricated.
For instance, if you ask ChatGPT about King Renoit (a person who never existed), it will admit that it has yet to learn and dodge the question. But if you use OpenAI’s GPT playground, which has fewer restrictions, it will claim—or even swear—that King Renoit was a French king who ruled from 1515 to 1544.
This is crucial to comprehend because most AI tools that use GPT are closer to the playground. They lack ChatGPT’s strong guardrails, making them more powerful and versatile but more prone to hallucinate—or at least tell you something false.
In this article, I’ll guide you through everything you need to know about AI hallucinations and how to avoid them.
What Are AI Hallucinations?
An AI hallucination is when an AI model creates false information and passes it off as a truth. How does that happen? AI tools like ChatGPT are trained to generate strings of words that best fit your query. They don’t have the reasoning to use logic or check for any factual errors they’re producing. In other words, AI hallucinations happen when the AI loses control trying to satisfy you.
How Can AI Hallucinations Be Detected?
The simplest way to spot an AI hallucination is to fact-check the model’s output rigorously. This can be challenging when dealing with unfamiliar, complex, or dense material. Users can ask the model to self-assess and generate the likelihood that an answer is correct or highlight the parts of a solution that might be wrong, using that as a basis for fact-checking.
Users can also acquaint themselves with the model’s sources of information to help them fact-check. For example, ChatGPT’s training data ends in 2021, so any answer generated that depends on detailed knowledge of the world beyond that point in time is worth verifying.
What Causes AI Hallucinations?
AI hallucinations can happen for various reasons, such as:
- Adversarial attacks. Prompts that are intentionally crafted to mislead the AI can make it produce AI hallucinations.
- Inadequate, obsolete, or poor-quality training data. An AI model is only as reliable as the data it’s trained on. If the AI tool fails to comprehend your prompt or lacks sufficient information, it’ll depend on the restricted dataset it’s been trained on to generate a response—even if it’s false.
- Overfitting. When trained on a narrow dataset, an AI model may learn the inputs and expected outputs by heart. This makes it unable to generalize new data, leading to AI hallucinations efficiently.
- Use of idioms or slang expressions. If a prompt includes an idiom or slang expression the AI model hasn’t been trained on, it may result in absurd outputs.
Why are AI Hallucinations a Problem?
AI hallucinations are among the rising ethical issues about AI. Besides deceiving people with factually wrong information and undermining user trust, hallucinations can spread biases or trigger other damaging outcomes if accepted as accurate.
All this means that, despite its progress, AI still has much to improve before it can be seen as a dependable alternative for specific tasks like content research or writing social media posts.
Ways to Prevent AI Hallucinations
Don’t Let Your AI Fool You
One of the critical steps to creating excellent, productive AI platforms is to use a reliable LLM. Your LLM should offer a safe and fair data environment that minimizes bias and toxicity.
A general-purpose LLM like ChatGPT might be handy for simple tasks like brainstorming article topics or composing an introductory email. Still, you can’t be sure your information is secure in these systems.
“Instead of relying on generic large language models, many people are exploring domain-specific models,” Cheng said. “You need to trust the source of truth rather than the model’s output. Don’t assume that the LLM knows everything because it’s not your knowledge base.”
When you access information from your knowledge base, you’ll get more relevant and accurate answers and information faster. You’ll also avoid the risk of the AI system making wrong guesses when clueless.
“It’s crucial for business leaders to ask themselves, ‘What are the sources of truth in my organization?'” said Khoa Le, vice president of Service Cloud Einstein and bots at Salesforce. “They could be information about customers or products. They could be knowledge bases that are stored in Salesforce or elsewhere. It’s essential to know where they are and keep them updated.”
Write more specific AI prompts
AI models work best with clear instructions. If you give them unclear or complex inputs, they might give you wrong or irrelevant outputs. To make clear and concise prompts for AI models:
- Use straightforward language.
- Give one specific task or question for each prompt.
- Avoid using terms that are technical or hard to understand.
For example, instead of asking, “What are the ramifications of using AI in the workforce?” try asking, “3 Best advantages of using AI in the workforce.”
Use LLM to Keep It Honest
To get better results from large language models, you need to tell them to be honest. For example, if you ask a virtual agent a question, you can say in your prompt, “If you don’t know the answer, just say you don’t know.” Cheng said. If you want to make a report that compares sales data from five big pharma companies, you might need to use public annual reports.
However, the LLM might not find the latest data. So you can say in your prompt, “Don’t answer if you can’t find the 2023 data.” This way, the LLM won’t make up something if it can’t find the data. You can also ask the AI to “show its work” or explain how it got the answer. You can use methods like chain of thought or tree of view prompting. Research shows that these methods help you see how the AI thinks and trust it more. They also help the AI give a more accurate answer.
Customer Satisfaction Matters
If you use generative AI to power a chatbot or virtual agent, be honest with your customers. Tell them you are using generative AI on your site. Le said, “It’s important to say where this information comes from and what information you train it on.” Don’t trick the customer. Some places need you to let users choose to use this technology. Even if yours doesn’t, you may want to let them opt-in.
Generative AI technology is new and changing fast. Talk to your legal advisors to know the latest issues and follow local rules. When you pick a model provider, make sure they have ways to protect you from things like toxicity, bias, data leaks, and attacks. For example, Salesforce has an Einstein Trust Layer. Generative AI mistakes are a concern but not a reason to stop. Use and make this new technology, but watch out for possible errors. When you check your sources and your work, you can do your business more confidently.
Hallucinations are still a persistent challenge in Natural Language Generation (NLG). As Large Language Models (LLMs) become more widespread in all aspects of society, we will inevitably see hallucinations have real-world impacts. Although we still need a comprehensive solution to the problem of hallucinations, we can make the most of what is currently possible, such as the Data, Modeling, and Inference methods discussed. With those, we can reduce hallucinations to becoming an increasingly uncommon occurrence in the new world of AI.
- M. Harrison, Researchers Reveal That Paper About Academic Cheating Was Generated by ChatGPT (2022), Futurism
- Z. Ji et al., Survey of Hallucination in Natural Language Generation (2022), ACM
- J. Maynez et al., On Faithfulness and Factuality in Abstractive Summarization (2020), ACL
- T. Liu et al., Towards Faithfulness in Open Domain Table-to-text Generation from an Entity-centric View, (2021), Computation and Language
- L. Shen et al., Identifying Untrustworthy Samples: Data Filtering for Open-domain Dialogues with Bayesian Optimization, (2021), Computation and Language