Introduction to generative AI

ChatGPT and similar services are based on what is known as large language models. These models can be characterized as systems developed with algorithms, trained on extensive data sets, capable of solving tasks without relying on pre-programmed rules for executing solutions. Various applications like ChatGPT can then use these models to generate text, based on the so-called prompts (roughly instructions) that we write.

The models do not contain a database of information; instead, the answers are generated based on probability. The language models are very good at mimicking human language, and the responses can therefore often seem correct, but they still make mistakes too often to be used for creating fact-based content.

It is also important not to use ChatGPT in its basic version (i.e., the free version without internet access) for seeking information. In the paid version of ChatGPT, there are options to activate general internet access, and there are other services that have this in their basic versions (such as Google Bard or Bing Chat/Copilot).

As a result of the development around generative AI, there is also a need for new types of competencies. One that is often mentioned is "prompt engineering", that is, the ability to formulate instructions that generate the desired result.

It might be easy to think that competencies in formulating prompts is similar to formulating searches in Google or databases, but the difference is quite significant. Where you can benefit from focusing on keywords in searches, it is more important to be as detailed and comprehensive as possible when formulating prompts. To get good results, you may need to provide clear instructions that limit and guide the generation, and it is common for prompts to consist of longer text passages. Here, it can be helpful to be inspired by others, and it is relatively easy to find examples through simple searches on the internet or in groups on Facebook. OpenAI has also published a guide to prompt engineering. 

The service most commonly mentioned in discussions about education at the time of writing is ChatGPT, but there are several other services that use different language models and have different purposes. Therefore, it is relevant to explore various services to see what opportunities they offer and what limitations they have. Other services and tools that may be relevant from specific subject areas and that individual teachers may need to know about include Perplexity, Elicit, jenni.ai, and GitHub Copilot.

It is also good to be aware that there are services for generating images, sound, and video. In particular, image generators have developed very quickly in terms of quality, which can pose challenges in different subject areas. The most common image generators as of this writing are Midjourney, DALL-E, and Stable Diffusion.

Given that the ability to generate content is not exclusive to specific AI-based services, it is beneficial to evaluate if comparable features are integrated into alternative systems or tools. An example of this is Copilot in Microsoft365 where GPT4 is integrated and can be used to generate texts, images, layouts, etc.

A good benchmark to look out for is whether there is an offer to generate text, image, video, or sound without you having to produce it yourself. If so, generative AI is most likely integrated in the tool or service.