top of page

Ask AI - What Can't You Do?

In this new series from PanOptic we will be asking AI questions about AI to get it's perspective of what it is.

What Generative AI is NOT Good At

Generative AI has made remarkable strides in recent years, creating everything from lifelike images and coherent paragraphs of text to innovative music compositions. However, it's important to understand its limitations. Despite its impressive abilities, there are areas where generative AI still struggles or is inappropriate for use:

  1. Emotion and Empathy: Generative AI can simulate human-like responses, but it doesn't truly understand or feel human emotions. It can't empathize with human experiences, and therefore, it might not capture the emotional nuance required in certain contexts.

  2. Context Awareness: AI models, especially when working in isolation, might not always grasp or respect the broader context. This can lead to outputs that may be inappropriate or offensive.

  3. Creating Truly Original Ideas: While generative models can combine and rephrase existing information in new ways, they don't genuinely "think" or "innovate" in the human sense. Their creations are reflections and remixes of what they've been trained on.

  4. Ethical Judgement: Generative AI doesn't possess a moral compass. It can't inherently discern right from wrong or make ethical decisions.

  5. Accountability and Responsibility: If something goes wrong, it's tough to hold an AI model accountable. It lacks intention and consciousness. Instead, the onus falls on its creators, trainers, and users.

  6. Adapting to Real-time, Dynamic Situations: Generative models, like those based on the GPT architecture, typically don't adapt in real-time to new information unless they are specifically designed or combined with other models to do so.

  7. Recognizing and Avoiding Bias: If a dataset used to train the AI is biased, the resulting model will likely perpetuate that bias. While there are methods to mitigate these biases, it remains a challenge in the field.

  8. Complex Multistep Tasks: Tasks that require multiple steps, especially if those steps are in different domains (like cooking a meal from scratch, then explaining a philosophical concept about it), can be challenging for current generative models.

  9. Interacting with the Physical World: Generative models are software-based. They lack a direct interface with the real world, and while they can be integrated into systems that do (like robots), there's a significant difference between generating content and physically acting upon or within the world.

  10. Understanding Limitations: Generative AI doesn’t inherently know its limitations. It can't tell you when it's out of its depth or making guesses.

  11. Data Dependency: The quality and nature of the data used to train generative AI play a huge role in its outputs. Poor quality or limited data will lead to subpar or skewed results.

  12. Over-Generation: One of the challenges with generative AI is knowing when to stop. It can over-generate content, leading to verbosity or unnecessary details.

In summary, while generative AI presents numerous opportunities across various fields, it's crucial to approach its capabilities with a nuanced understanding. Recognizing what it can't do as much as what it can do ensures that we harness its potential responsibly and effectively.

The interesting take aways from this were #10. Try to wrap your head around something answering something in great detail and then saying it cannot understand what it just answered?

The accountability piece will be found solely resting on the organizational ownership and not the model itself. It is going to be extremely vital for organizations to understand how to feed data into the model and train them appropriately to get a desired outcome. Also, understanding what each model is good at and picking appropriately will be key.

Overall, what is starting to be apparent is that Generative AI should be looked at as a resource for humans to do more meaningful work more often, it's ability to comprehend massive amounts of context is AMAZING. It has limitations around creative ideas from scratch and is only as good as the data and prompting.

14 views0 comments


bottom of page