Task Typology for AI Suitability
This is a framework for identifying AI suitability of different task types. It separates tasks into information addition, information equivalence, and information reduction, and outlines the potential for effective AI usage within each category. Some tasks will share characteristics of all three. When this occurs, it can be helpful to break down the task even further into component activities before applying the framework.
Information Addition
Information addition tasks often occur at the beginning of a project. These tasks rely on the cognitive work of deciding which information gets added; the primary challenge is identifying and selecting relevant information. They include literature review and user research amongst other kinds of discovery work. An internet-enabled AI such as Perplexity can help with this by identifying potential sources, but deciding what to look for and deciding which ones are relevant remains a human responsibility. The determination of which sources are appropriate for a given task often depends on the "why" motivating the discovery work, and humans can identify this "why" much better than an AI can, even with expanding context windows (A context window is how much information a given AI can “remember”, or process, in one session.)
Linguistically, we might consider this the "pragmatic" purpose of the work. A pragmatic meaning is the "meaning that arises through the USE of language" (Linguistics Society), and is heavily context dependent. For example, a research paper will generally have a different intended function than a pitch deck does, and will be written in a different style and tailored to its specific audience. This context influences the meaning and form of the communication produced, and (thus far) humans are much better at grasping this pragmatic context than AIs are. An AI can only know a document's purpose - its pragmatic meaning - if you tell the AI what that purpose is. A human can recognize these core differences in function and determine constraints on what type of information is relevant, and then these constraints can be used to effectively prompt an AI to find instances of information to potentially incorporate.
Some questions that can help identify Information Addition tasks:
- Does introducing novel content correlate highly with fulfilling task’s purpose?
- Does this task expand my breadth of knowledge on this topic or project?
- Does this task include introducing new questions?
- Would this task produce unfavorable outcomes if I couldn’t use the internet for it?
Information Equivalent
Information equivalent activities rely on translating existing information into a different format without altering the content or purpose. This is primarily a semantic difference, as the core meaning of the information doesn't change. This translation is often the core labor of a project, digesting information from one audience and preparing it for consumption by another. This can include visual formatting changes, but expands well beyond formatting. Translating existing information encompasses developing additional instances of a given theme, updating content for different audiences, and re-organizing existing thinking for greater clarity or conciseness.
Using an AI to generate code is a great example of this kind of translation. When writing code, a programmer translates the conceptual decision of what needs to be done into specific instructions for the computer to execute. Telling an AI to "write code" might not get you very far, but telling an AI to "write Python code that counts from n to n+m and then prints the sum of each number counted" is likely to give you something more workable. The addition of "what to do" stems from the human's thinking, while the translation of that concept into specific Python code is something the AI does well. Since translations ideally maintain the density and quantity of information across formats (and are, hence, “equal”), these types of tasks are categorized as "information equalization". Of the three task types, these are best suited for leveraging AI.
Some questions that can help identify Information Equalization tasks:
- Am I re-shaping existing content for a different audience or context?
- Do I already feel confident in my breadth of knowledge on this topic?
- Could I, in theory, do this task without an internet connection?
Information Reduction
The third task type is "information reductive", and is usually performed towards the end of a project. These tasks involve removing noise from a given body of work. Once again, the human must decide what the initial body of work includes and what "valuable" means for any given scenario; however, once those decisions are made, an AI can be an excellent executor of them.
An example of this task type is The Beatles using AI to reduce literal noise from a 1980 cassette recording. The original recording included low-quality audio and static that made the recording, in its entirety, unworkable. By reducing the information - in this case, sound - From the recording, they were able to extract something more valuable: John Lennon's voice. When instructed properly, AIs can do noise reduction quite well, giving the final product greater clarity of purpose and legibility. This reduction was part of the editing process for this post.
Some questions that can help identify Information Reduction tasks:
- Does this benefit from being made more concise?
- Are there elements present that don’t contribute to the purpose of this task?
- If I had to sub-divide this information, are there obvious criteria by which to do so?
Conclusion
This task typology - addition, equivalence, and reduction - offers a framework for identifying what tasks might benefit from using AI and which are likely more efficient when done by hand (as of the time of writing). Additive tasks generally benefit the most from human involvement, while equivalence tasks are most ripe for delegation to an AI. If a human provides specific criteria, an AI can be a great assistant for information reduction tasks, though it is likely to falter without specific instructions. As with all typologies, there will of course be some tasks that don’t fit neatly into any of the categories. Some tasks share characteristics of all three. These categories aren’t meant as hard-and-fast “rules”, but instead as one way of thinking about the suitability of using AI for different activities. Slow, step-by-step task analysis is more likely to lead to useful results from using this framework than broad summarizations are. If you'd like to learn more about how to identify these task types in your work, schedule a consultation.
For Further Reading
Design & UX
What is Divergent Thinking?. (2016, June 5). The Interaction Design Foundation - IxDF. https://www.interaction-design.org/literature/topics/divergent-thinking
How to improve your UX designs with Task Analysis. (n.d.). The Interaction Design Foundation. https://www.interaction-design.org/literature/article/task-analysis-a-ux-designer-s-best-friend
Linguistics
Meaning (Semantics and Pragmatics) (2024, March 6). Linguistic Society of America. https://www.linguisticsociety.org/resource/meaning-semantics-and-pragmatics
Pragmatics: An Overview (2017). LSA Linguistic Institute. https://www.linguisticsociety.org/sites/default/files/e-learning/Overview-revised.pdf
AI Links
CLAUD Programming Examples. (Accessed 2024, March 5.) Anthropic.
ChatGPT-4 Editor: Slow Writing (Unpublished). (Accessed 2024, March 4.) OpenAI.
Perplexity.ai. (Accessed 2024, March 6.) Perplexity.