Debunking AI Myths: 10 Common Misconceptions About Large Language Models?

Large language models (LLMs) are one of the most hyped innovations in tech. In fact, McKinsey estimates that generative AI could add between $2.6-$4.4 trillion annually to the global economy and increase the impact of all artificial intelligence (AI) by 15 to 40%.

However, while the market for generative AI continues to grow, there are still a lot of misconceptions and myths circulating about how language models work.?From thinking that LLMs are sentient to thinking they can generate content with high accuracy and no bias, there are a number of misconceptions users need to be aware of.

Demystifying Language Models: 10 AI Myths Debunked

1. LLMs Can Think

One of the most common misconceptions about LLMs is that they can think independently. In reality, language models can make inferences from a dataset and can create a summary or predict text, but they don’t understand natural language the way a human would.

They process the user’s input and use patterns they’ve learned from training data to determine how to respond. Likewise, they don’t understand emotions, sarcasm, and colloquialisms. This means that modern LLMs are a long way off from artificial general intelligence (AGI).

2. Language Models Create Content

While LLMs can be used in content creation, they don’t independently innovate and create original content. Instead, they take patterns from written or visual content they’ve observed in their training data and use it to predict and generate content that’s based on its training data.

The use of training data to generate responses is a controversial practice. For example, three artists waged a class action lawsuit against Stability AI, DeviantArt, and Midjourney, arguing that “stolen works power these AI products” because they’re trained on copyrighted images scraped from the internet.

3. All Inputs Are Confidential

Another significant misconception about LLMs is that the data entered into the input is completely confidential. This isn’t necessarily true. At the start of this year, Samsung banned ChatGPT in the workplace after an employee leaked confidential data into the solution due to concerns the information shared was being stored in an external server.

Organizations looking to use generative AI must thus highlight what information employees can and can’t share with language models, otherwise, they run the risk of falling foul of data protection regulations such as the General Data Protection Regulation (GDPR).

4. Generative AI Is 100% Accurate

Many users make the mistake of believing that the information that tools like ChatGPT and Bard generate is 100% accurate, or at least generally accurate. Unfortunately, language models are susceptible to hallucination, meaning they can fake facts and information and state them “confidently” as if they were correct.

As a result, users need to make sure that they double-check facts and logical explanations so they don’t end up being misled by misinformation and nonsense outputs.

5. LLMS Are Impartial and Unbiased

Given that LLMs are developed by human beings and mimic human language, it’s important to remember that biases are embedded into these systems, particularly if there are errors in the underlying training data. This means users can’t afford to consider them impartial and unbiased sources.

Machine bias can show up in language models in the form of inaccuracies or incorrect information or, more overtly, in the form of hateful or offensive content. The extent to which these biases show up depends on the data these models are trained on.

6. Generative AI Is Effective in All Languages

Although generative AI solutions can be used to translate information from one language to another, their effectiveness in doing so depends on the popularity of the language being used.

LLMs can generate convincing responses in popular European languages like English and Spanish but fall flat at creating responses in languages that are less commonly used.

7. LLMs Report Information from the Internet

Language models like GPT4 and GPT 3.5 don’t access the Internet in real-time but process their training data (some of which is scraped from the Internet).

For providers like Google, OpenAI, and Microsoft, the nature of this training data is largely kept in a blackbox, meaning that users have no insight into what information LLMs are using to generate outputs. This means users can’t afford to assume that the information is up-to-date or accurate.

8. LLMs Are Designed to Replace Human Employees

While AI has the potential to automate millions of jobs, LLMs in their current form can’t replace the intelligence, creativity, and ingenuity of human employees. Generative AI is a tool that’s designed to work alongside knowledge workers rather than to place them.

Combining the expertise of employees alongside the scalability and processing capabilities of LLMs can be referred to as augmented intelligence.

9. LLMs Can’t Produce Malicious Content

Some users might believe that the content moderation guardrails of providers like OpenAI prevent other individuals from using them to create offensive or malicious content, but this isn’t the case.

With jailbreaks and some inventive prompts, cybercriminals can trick LLMs into generating malicious code and phishing emails that they can use out in the wild to steal private information.

10. LLMs Can Learn New Information Continuously

Unlike human beings, LLMs aren’t learning new information all the time but are using deep learning techniques to identify new patterns within their training data. Better understanding of these patterns enables them to take more detailed inferences from a dataset.

As such, organizations would need to retrain an LLM if they want it to learn new data that isn’t part of the original training data.

Language Models: Best with Supervision

LLMs have the potential to be a force multiplier for knowledge workers, but it’s important to be realistic about your expectations for the technology if you want to get the best results.

Being vigilant for hallucinations, biases, and inaccuracies will help to avoid the chance of being misled and will enable users to increase their chances of extracting concrete insights to enhance their decision-making.