What if i could confidently give you the wrong answer?
Hallucinations in Generative AI are a real challenge. They happen when language models produce inaccurate or fabricated responses, often with absolute certainty. For businesses, this can mean misleading insights, compliance risks, and lost trust.
Why Do Hallucinations Happen?
Think of a chat model like a multitasker—listening to the radio, talking on the phone, and answering a test question simultaneously. With divided attention, accuracy suffers. Similarly, a general-purpose AI without clear instructions can “guess” its way through a response, leading to hallucinations.
Most chat models are generalists, designed for a wide range of tasks rather than specialising in one domain. Without clear instructions, they “guess”—and that’s where hallucinations creep in. They can be referred to as “Jack of all trades”.
Key causes of hallucinations include:
- Ambiguous prompts: Vague instructions leave the model guessing.
- Lack of domain context: General models lack deep expertise in specialised fields.
- Overconfidence in output: AI often presents fabricated answers with high certainty.
Selecting the right language model for your use case is a great way to ensure your chat agent produces the correct response which aligns with user expectation.
It is therefore important to set the stage for your agent and give it specific and clear instructions. This enables the agent to tailor it’s output.
Some chat agent language models and their key uses are listed below;
| Language Model | Key Use |
|---|---|
| GPT-4o | Multi-modal model for generating texts. Can process images and audio prompts too. |
| DALL-E | Text to Image generation model |
| Stability AI | To create realistic visuals from text prompts. |
| Text-embedding- ada | An embedding model used for improved search results. Embeddings models convert text into numerical representations. Can be used to ground data. |
| Llama 3.1 8B Instruct | For general, efficient multilingual chat |
| Deepseek-R1 | For complex logical reasoning, intricate math problems, coding, research, and agentic workflows |
| Core42 JAIS | An Arabic language LLM, making it the best choice for applications targeting Arabic-speaking users. |
| Mistral Large | Has a strong focus on European languages, ensuring better linguistic accuracy for multilingual applications. |
| Nixtla TimeGEN-1 | Specializes in time-series forecasting, making it ideal for financial predictions, supply chain optimization, and demand forecasting. |
If your project has regional, linguistic, or industry-specific needs, these models can offer more relevant results than general-purpose AI.
Use the Microsoft Copilot agent readily available in your Microsoft 365 suite to tailor your agent responses and reduce hallucinations.
See how to demo here.
It is important to note that when your ChatGPT hallucinates or doesn’t give the expected outputs, it’s likely the model is not optimised for the task you’re trying to perform.
Try one of the other models which are optimised for your use. Prompt engineering is key for non developers as access to underlying models is usually restricted. You can use prompts to guide the chat agent to produce desired outputs.
Ensure your prompts give clear and well defined instructions in order to generate the best outcome. If you’re on a Microsoft platform, make the most of the 365 copilot – see demo – to increase productivity.
Leave a comment