Using Generative AI To Do Research

These are tools where one key feature is generating new text in response to user prompts. They are transformers because they work with enormous amounts of text and employ algorithms to find patterns and relationships in the huge datasets they are trained on (mainly from internet sources). They are commonly referred to as Large Language Models (LLMs) using probability or predictive technology to determine which words should appear in a sequence in response to prompting. 

Outputs that all these tools generate include many different text types, e.g., news articles, essays, reflective statements, blogs, poetry and more.

In the Points to Keep in Mind section of Types of Tools, we explain that some of these tools are grounded which means that they may incorporate real-time web search and/or an external knowledge base or data source, with the result that relevance and factuality is stronger. We recommend checking out tool descriptions to determine when and if this applies, but it is now common for many of them to offer this in one way or another. Even where a tool is grounded, we recommend verifying sources and critically evaluating the tool's output. 

Examples of text-generating tools include ChatGPT (OpenAI), Google (Gemini), Co-Pilot (Microsoft), Claude (Anthropic). Features will vary for different versions and/or for free versus pro models. 

These tools generate images from textual descriptions and/or from other images. They work with huge datasets of images and generate outputs based on analysis of those images, including their captions and text descriptions. 

Examples include DALL-E, Midjourney, Stable Diffusion, and Adobe Firefly. Note that the big players like ChatGPT (OpenAI) and Co-Pilot (Microsoft) now incorporate image-generation features too, but often this is a pro feature.

Outputs may take the form of many different types of graphics or media, including cartoons, photos, anime, oil paintings, illustrations and more.

This includes tools like AIVA and Soundful. They are designed for audio generation tasks. They are often broken down in to different sub-types including those that facilitate voice generation/cloning, those that generate music, and finally general audio generation types. Prompts might take the form of text, musical notes, melodies or chord progressions, or audio samples or excerpts. They train on audio data and through this process they can generate new audio that has the style and characteristics of the data they were trained on. 

Outputs may take the form of synthesized voices, sound effects, background soundscapes, and music in the form of songs, melodies, compositions, etc. This can even include new genres of music that mimic existing styles. 

This includes tools like Runway, Pika, Invideo, and Deforum that generate video content. Prompts can be given in various forms, including text, images, or videos. They are trained on large datasets of video clips, including movies, TV shows, or other video content. Using this data, they are trained to understand and replicate visual patterns and structures in the input video data, as well as temporal and spatial features. Just as with audio tools, these tools can now often assist with video editing and are starting to appear in tools like Adobe Premiere or DaVinci Resolve. 

Outputs can take many forms, including generating simple animations or special effects to complex video scenes or entire movies. 

There are quite a range of tools in this category. They work with scholarly research outputs like articles to search and analyze them, extract key information, generate literature reviews, maps and summaries, or provide features common with citation tools. In a nutshell, they are often billed as helping to automate research workflows. Examples include Elicit, Scite, ResearchRabbit, and Semantic Scholar. 

These tools are much more fully explored in the Research Discovery & Workflow Tools section of this guide. Some of these tools are not 100% free and have usage limits.