Limitations and capabilities of Generative AI in 2023

Use language models and GenAI tools correctly

It’s been a year since the launch of ChatGPT, and the interest in learning how to effectively integrate GenerativeAI (Gen-AI) based tools with business operations and processes keeps growing.

Deploying these technologies requires one to ponder the potential implications of adopting them but it is critical to firstly understand what we can realistically do, what we can’t and what we should not do with them.

We will quickly discuss the actual limitations to later list the real capabilities that can be used to any professional’s advantage. But before we dive in…

Let’s lay out the difference between GenAI and the mainstream generative tool: ChatGPT.

· GenAI is a broad AI field including methods and models to generate audio, visuals, text etc. These models’ underlying principle is to learn from previous datasets to create novel outputs based on the patterns they were trained on. GenAI is widely used in design tasks and content creation.

· ChatGPT is a specific application under the umbrella of GenAI crafted for conversational purposes. In essence, it is a language model trained on large amounts of text, enabling it to generate human-like responses, dialogue, and fake, yet realistic conversations. Its the mainstream conversational tool and it is a widely used tool for customer service, chat bots, etc.

It’s critical to know that GenAI’s scope is broader, extending into realms well beyond conversational interfaces.

The 5 key limitations of GenAI that anyone should be aware of

1. Incapable of learning independently: Generative models are trained on a fixed set of data and do not learn new information on their own. They do rely on external inputs for additional context but do not modify their underlying knowledge base. You may be able to provide them with domain specific knowledge and industry context, but these technologies will require specific training to be adapted to each business environment and company needs.
2. Unable to replace human traits: Like creativity, emotional intelligence or proactive learning. These tools are completely unable to sense human thoughts, feelings, or to grasp fresh ideas. At the moment, they are not even close to fully replacing a human being in a task where human traits are needed. There are configuration options for tools like ChatGPT, so that the output is more creative (0 to 1 “temperature” in system prompts) but this tends to lead to a higher degree of inaccuracy and the system sometimes comes up randomly with totally fake information.
3. Difficulty to cite sources. These models do not store information in a way that allows citation. Meaning, it will not be easy to substantiate your work when some of it comes from tools like ChatGPT. This requires professionals to do extra work in documenting where key data is gathered or unverified information may be a problem down the road. There are some articles (here and there), guiding users on how to get these technologies to cite sources but it does not always work and the tool may provide fake citations or just miss key sources of the main information.
4. Lack of certainty: GenAI cannot reassess or be certain about it responses, even if it seems confident. They operate based on wording probabilities, leading grammatically correct responses but potentially inaccurate outputs. Professionals should never take the validity of information for granted.
5. Hard-to-identify artificial content: As generative technologies advance, distinguishing between human and AI-fueled content becomes harder and harder. Tools designed to detect AI-generated content are not close to bulletproof while the line between artificial and genuine continues to blur. This may seem like a benefit to many professionals but in reality this is what leaves room for fake news, wrong insights and similar problems in society.

Capabilities of Generative AI that intrinsically relate to the previous limitations

1. Summarising large volumes of information into concise summaries. Meeting transcripts, and extracting key points from lengthy documents are easy tasks for LLMs, for instance. However, remember to verify the accuracy of these summaries and the data in it.
2. Vast pool of knowledge: GenAI models may seem (or act as) a knowledge repository and a “wiki-tool” on almost all topics. But remember: GenAI is prone to providing overly confident yet incomplete, biased, and inaccurate responses.
3. GenAI can aid in coding tasks by offering suggestions and helping debug errors. While it is not flawless, it can help in software development, data analysis, and similar tasks. Actually, I have seen it in real life with capable data scientists myself. Again, remember to ask your real human colleague to review your work for you too.
4. Synthetic imagination: Although GenAI does not feel, sense or imagine, there are tools like Midjourney that allow to produce images from natural language descriptions called “prompts”, which are just well-written human instructions.
5. Seamless integrations with a wide variety of software solutions: Possible connection and integration with software and technology processes to save time and generate efficiencies. GenAI can be incorporated into business processes to address the previous 4 points but remember that humans should always review the outcome of any processes where GenAI is integrated.

Conclusion

GenAI shows remarkable capabilities and is transforming how many of us approach common daily and business tasks. However, it is crucial to:

• Understand that it can assist but not lead human tasks independently.
• Be conscious of its lack of trustworthy behaviour.
• Integrate responsibly and with a critical eye. Streamlining processes where the output is not 100% reliable can amplify errors.
• Be aware of the impossibility of tracing back to the original source of information.

… and my own personal recommendations:

• For business users of GenAI, it would be beneficial to stay updated on advancements in the field and be aware of new technology and methods developed to address its limitations. Platforms like Medium can help with this, but there are also plenty of EdTech courses and online tools available.
• Adopting a collaborative approach, where GenAI and humans work in tandem, can help in maximising the benefits of this technology while minimising potential risks.