They arrived nearly as quickly as the excitement surrounding generative AI (GenAI)—hallucinations. Hallucinations occur when large language models (LLMs) create incorrect outputs that are out of alignment with their training data and algorithms. The tendency of GenAI to hallucinate has created some interesting and high-profile real-world examples. There was the New York attorney who, after consulting with ChatGPT, presented precedent-setting cases in a personal injury lawsuit against an airline. The problem? The cases didn’t exist. In another instance, Google’s chatbot claimed that the James Webb Space Telescope (JWST) took the first image of a planet outside our solar system. In reality, the first images of an exoplanet had been taken 17 years before JWST’s exoplanet images. After NASA corrected Google, its parent company Alphabet lost $100 billion in value.
As the concern over hallucinations grows, it’s no wonder that AI’s trustworthiness has taken a hit. Last year, a survey by KPMG reported that 61% of respondents were wary of AI systems. That wariness extends beyond the systems to AI companies as well. Axios recently reported that faith in AI companies hovers around 53%. Even with these issues, businesses are adopting AI at warp speed. McKinsey’s 2024 “State of AI” survey revealed that 65% of companies are using AI for at least one business function. Of those respondents using GenAI, their companies are using it for at least two functions. It’s impossible to slow or halt the adoption of AI across industries and yet, hallucinations can lead to major concerns like flawed decision-making, missed market opportunities, and revenue loss.
So, the following questions arise: How can you minimize hallucinations and mitigate the risk of negative impact when they do occur?
HTEC’s latest white paper, When Machines Dream: Overcoming the Challenges of Fake Data and AI Hallucinations, by Sanja Bogdanovic-Dinic, Head of Data Solutions outlines steps organizations can take to minimize hallucinations in their AI systems. Here, we provide an exclusive look at the paper and insights from the author.
The paper begins by establishing the leading causes of hallucinations. Studies show that GenAI chatbots are inaccurate 3-27% of the time. The reasons fall into three main categories:
- Overfitting: This refers to AI models that do well with the data they’re originally trained to interpret but are unable to properly provide predictions for new data. Essentially the model is fixated on the original data, which can be a result of the training data’s complexity and noise or an over-extended training period.
- Low-quality, Insufficient, and Outdated Training Data: According to Bogdanovic-Dinic, “LLMs are only as good as the data on which they’ve been trained.” Outdated, biased, and incorrect data will inevitably lead to output errors.
- Prompting Mistakes: Simply stated, poorly worded or phrased prompts will lead to poor outputs. Other incorrect prompt approaches include using vague, ambiguous, or conflicting language, slang or idioms, or engaging in “adversarial attacks.” Adversarial attacks can be used by threat actors to intentionally confuse an AI model and commit a cybercrime.
How to avoid hallucination triggers
There are currently no universal guidelines for AI use, so companies often find themselves trying to minimize problems on their own. To help, HTEC experts created a list of recommendations to help reduce the frequency and impact of hallucinations:
- Carefully Consider Your GenAI Use Cases: Certain industries (retail or entertainment) may be more risk tolerant than others (healthcare or finance). It’s still important, however, for companies to assess the potential impact of hallucinations before implementing GenAI anywhere in their organizations. Before using GenAI, develop a company-wide strategy for the technology, and if even the occasional hallucination can cause damage, GenAI is the wrong use case.
- Invest in Data Quality: Poor quality data is a major contributing factor to hallucinations. Make sure your data sets are cleansed, well-organized, and validated. (Want more information on data maturity and readiness? Read HTEC’s white paper, “Data and AI Readiness Assessment: Gen AI is Here to Stay — Is Your Data Ready?”)
- Integrate Explainable AI (XAI) into Your User Experience: Transparency in your AI systems can help maintain user trust even when problems do occur. XAI explains a model’s logic and decision-making processes in plain language. It can be helpful for compliance and debugging, while aiding end users in recognizing and navigating hallucinations.
- Use Retrieval-Augmented Generation to Improve Output Accuracy: Retrieval-augmented generation (RAG) improves AI model accuracy by consulting outside knowledge sources before generating a response. By cross-referencing its responses against outside sources, the system is less likely to produce incorrect outputs.
- Use Better Prompting Techniques: Similar to using a good data set, increasing the quality of your prompts can reduce hallucinations. Recommended approaches include using detailed language, affirmative directives, and breaking down complex questions into smaller prompts that build upon themselves.
- Enable Secondary AI Monitoring Agents: Monitoring agents are specialized software that run continuous scans to track performance. These agents are crucial for your organization’s GenAI models, and secondary LLMs. With secondary LLM monitoring, the same prompt can be put into both primary and secondary models. If the responses differ, a hallucination can be detected.
- Consider Implementing a Fault-tolerant Agent Structure: Fault tolerance refers to a system’s ability to operate even when software malfunctions. Building fault-tolerant agent groups using feedback loops, redundant pathways, and adaptive mechanisms can create more resilient systems better equipped to recognize and address hallucinations.
- Build a Strategy to Investigate and Evaluate Hallucinations: To make hallucinations as infrequent as possible, the paper recommends meticulously assessing ones that do occur. Establish a dedicated team of experts in your organization to dissect and analyze hallucinations. With each evaluation, search for patterns, biases, and system vulnerabilities that you can correct to strengthen your systems.
Are hallucinations all bad?
Bogdanovic-Dinic notes that there are moments when hallucinations can actually be helpful. GenAI visuals, for example, can introduce novel design ideas. You can also feed your system seed prompts and use its hallucinations for brainstorming.
When your company’s reputation is on the line, however, remember that reducing the negative impact of hallucinations begins with good data, better prompting, and careful use case consideration. HTEC partners with clients to build hallucination-resistant AI products from the ground up — starting with best-in-class data preparation techniques and cross-functional teams to understand each client’s GenAI needs.
“Hallucinations are an important topic because they’re a byproduct of a very exciting technology,” Bogdanovic-Dinic says. “They stand in the way of AI’s promise by affecting user trust. So, it’s important to understand what they are, what they mean, and what kind of impact they make.”
