AI Can Hallucinate Too: What Are the Dangers, and How Can We Protect Ourselves?

Artificial Intelligence (AI) progresses every day, attracting an increasing number of followers aware of its potential. However, it is not infallible and every user must maintain a critical mindset when using it to avoid falling victim to an “AI hallucination”. What exactly is an AI hallucination? How do we identify them, and most importantly, how do we protect ourselves from them? Let’s explore this together.

Yes, AI also hallucinates!

Just like their creators, AI systems can sometimes experience hallucinations. This is a phenomenon where an AI generates a response that seems plausible, but is in fact partially or entirely false or unrelated to the given context. In other words, the AI “imagines” information that is not true or relevant. This can occur in various contexts, from text generation to image recognition.

An AI could, therefore, whip up literary works that never existed, just to back up its argument. Or it could respond to you with details about the life and history of people who never existed.

It could also contradict itself, as seen in our test on GPT-4:

Here, for example, we can see that despite asking for “the number of victories of the New Jersey Devils in 2014”, the AI's response is that it “unfortunately does not have data after 2021”. Since it doesn't have data after 2021, it therefore can't provide us with an answer for 2014. Doesn’t make sense, does it?

This is yet another hallucination. But why are these happening?

Why do AI hallucinations occur?

In the context of large-scale language models (LLM), like GPT-3 or GPT-4, hallucination can be attributed to a lack of truth coming from external sources. LLMs are indeed based on vast amounts of textual data, which allow them to predict the next word in a sentence based on the context provided by the previous words. However, they do not have access to external sources of truth to verify their predictions, which can lead them to learn and propagate inaccuracies present in the training data.

Of course, this last parameter tends to become less true today with AI models like Google Bard, or Bing, constantly connected to the Internet. Thus, while hallucinations are becoming less and less frequent, it is crucial to remember that they still can exist and that they can even have disastrous consequences.

AI Hallucinations can be disastrous, causing major challenges for tomorrow

While being wrong about an ice hockey team’s number of wins or creating a historical fact out of thin air might not have major repercussions, AI hallucinations can still turn out to be disastrous.

Imagine an AI spouting off hallucinated medical information during a medical analysis. This could easily throw doctors for a loop and lead to the patient's death. The same is true in connected cars where an AI could hallucinate the presence or absence of a physical element, causing it to react inappropriately.

To give a concrete example, lawyer Steven Schwartz fell victim to AI hallucinations in May 2023 during a lawsuit against the Colombian Airline company Avianca. To defend his client, who suffered a knee injury after being struck by a metal cart, Steven had to dig up similar cases in order to win the lawsuit.

Steven used ChatGPT to conduct his research. The AI found six cases similar to his, which the artificial intelligence assured were true and consultable in the "LexisNexis" and "Westlaw" databases, used by lawyers.

Alas, once the pleading was done, no one could find the judgments, nor the quoted and summarized excerpts, seeing that they did not actually exist. 

The AI had simply hallucinated them to help Steven win his case.

Such a mistake is unforgivable, and Steven's reputation has been annihilated. He is now awaiting trial for serious misconduct.

From this example alone, it’s important to note that we should always keep a sharp mind when using an AI and always double-check its response by Doing Your Own Research (DYOR)!

Fortunately, while there are no other miracle recipes to protect you from these  hallucinations, there are still good practices to reduce the risk of getting them when using AI.

How to Protect Yourself from AI Hallucinations

This list is not exhaustive, but when used properly, it can significantly reduce the risk of encountering hallucinations. However, as we just mentioned, it is imperative that you always verify the truthfulness of the provided responses:

• Improved training data: One of the main causes of AI hallucinations is the use of poor quality or biased training data. By using more accurate and diverse training data, we can reduce the likelihood of the AI hallucinating. This is indeed what we allow at HUMAN Protocol, curated data, leading to fewer hallucinations.
• Simulation of adverse scenarios: Another strategy is to test the AI in adverse scenarios to see how it reacts. This can help identify the AI's vulnerabilities to hallucinations.
• Transparency and explainability: It is important that AI systems are transparent and explainable. This means that we should be able to understand how and why the AI makes certain decisions. This can help identify when and why the AI hallucinates.
• Incorporating human reviewers: Using human reviewers to validate the outputs of the AI system can be an effective strategy to prevent hallucinations. Human reviewers can identify hallucinations that AI might not be able to recognize. Again, HUMAN Protocol is a major ally in these tasks and is already performing millions of operations in this sense.
• Domain-specific refinement: Help the model better understand the context and conventions of a particular domain for more accurate and reliable outputs.
• Adversarial training: Train the model to recognize and avoid generating hallucinated content.
• Multimodal models: Use models that combine language models with other modalities such as images or structured data to provide additional context and help anchor the model's outputs in reality.
• Reinforcement learning with human feedback (RLHF): Use human feedback to reward the model when it is correct and get it back on track when it deviates. At HUMAN Protocol, we are aware of the importance of RLHF and have enabled the launch of Ask Athena, a market research application that uses the HUMAN protocol to ask humans for their opinions, for example in the context of A/B testing or multiple-choice question responses.
• Limit possible results: When giving instructions to the AI, limit possible results by specifying the type of response you want.
• Include relevant data and sources unique to you: Anchoring your prompts with relevant information or existing data that you have gives the AI additional context.
• Create a data model for the AI to follow: Provide example data in a prompt to guide the behavior of an AI model.
• Assign a specific role to the AI and tell it not to lie: Assign a specific role to the AI to stop any hallucination.
• Tell the AI what you want – and what you don't want: Anticipate an AI's response to get rid of superfluous responses.
• Demonstrate the inconsistency of an AI's response: If an AI gives you an inconsistent response, don't hesitate to demonstrate it logically to force it to give you a better response.

To test out these precautions, we used the last point during our conversation on “the number of victories of the New Jersey Devils", and here’s the response we got:

The response is indeed correct: 35 victories! This is no longer a hallucination, but a correct answer.

Final Thoughts

By getting a solid handle on what AI hallucination is, how to spot it, and how to guard against it, we're able to leverage AI technologies in a safer and more efficient way. However, it must be recognized that the hallucination of artificial intelligence is a complex challenge that requires constant vigilance and ongoing research. At HUMAN Protocol, we actively contribute to this fight against hallucination by facilitating the verification of data accuracy and collecting feedback from millions of human users who are transparently compensated for their efforts.

Although AI is powerful, it is not infallible. It is our responsibility to remain vigilant and collaborate to improve it.‍

‍
_{Legal Disclaimer}

_{The HUMAN Protocol Foundation makes no representation, warranty, or undertaking, express or implied, as to the accuracy, reliability, completeness, or reasonableness of the information contained here. Any assumptions, opinions, and estimations expressed constitute the HUMAN Protocol Foundation’s judgment as of the time of publishing and are subject to change without notice. Any projection contained within the information presented here is based on a number of assumptions, and there can be no guarantee that any projected outcomes will be achieved.}

‍