Generative AI and Tesler’s Law

6 min readJan 12, 2024

We talk a lot about the new innovations enabled through generative AI, but we haven’t talked as much about how we expect users to interact with AI applications. Because of the more unpredictable nature of generative AI, user interfaces will shift from deterministic to probabilistic user interfaces.

In mathematics, a deterministic mathematical model is meant to return a single solution, an outcome of an experiment following specific inputs. In a contrast, a probabilistic model gives a distribution of a possible outcomes, describing all outcomes and giving some measure of how likely each is to occur.

Depending on the prompt that you provide to the chatbot, you’ll get a certain outcome. The form of this outcome then depends on the modality of the system, e.g. voice, text or imagery. The probabilistic aspect of AI raises the question about how to best design an AI interface and best manage its inherent complexity.

To what extent is the user expected to handle this complexity?

How will design patterns evolve when data is the product?

These questions relate to the law of conservation of complexity, also known as ‘Tesler’s Law’. Larry Tesler was a scientist at Xerox Parc who realised that any system houses a certain amount of complexity and that this complexity can’t be reduced. Tesler famously invented the ‘copy and paste’ functionality which is so engrained in all software today. By doing ‘Control c’ (cut) and ‘Control v’ (paste) the complexity of manually duplicating text is shifted from the user to the system.

According to Tesler’s law, any inherent complexity that can’t be removed or hidden should be dealt with — either in product development or user interaction. It raises the question about the complexities that we need to consider when designing applications based on generative AI:

→ Data is the output (e.g. in the form of text, audio or an image)

→ The output is probabilistic and depends largely on the quality of user input

→ The output is harder to interpret for the user as the logic behind the data decision-making

It means that we need to focus on the complexity involved in getting the user to provide the right input so that the AI can deliver the right output (even when the user can’t fully understand how AI arrived at this output). ‘Bad data in = bad data out’ applies to the prompt that users give to the chatbot. If the user gives a prompt that is vague or too broad, they’ll get an unhelpful answer in return. Over time, users will learn how to provide the right level of input so that they get the right output.

I envisage the design of interfaces for applications or features based on generative AI to make it as easy as possible for users to generate the right prompt to get the right generate the desired output from an AI. This goes beyond simply giving users easy access to a chatbot, such as through Microsoft’s new Copilot key.

To help users get to the desired output, interfaces will need to provide them with specific guardrails or set categories to choose from. Avatar creation app Lensa AI provides different styles in which a user’s image can be generated. If users aren’t happy with the results generated from applying one style, they can go back and iterate the avatar by selecting an alternative style.

It’s the same for products like Snapchat+ which lets users create images based on a text prompt and send them to friends. Snapchat+ users can also use AI to fill in the background of their images.

I see tools like Copy.ai providing its users with helpful questions to provide with quality output. Questions like “what are you looking to create?” and “what are the main points you want to cover?” are designed to shift complexity from the user to the AI.

You can then refine the returned content through “rewrite with AI” type functionality. Another emerging generative AI trend are the platforms that let people design their dream house or generate a new look for an existing room. On platforms like Interior AI you can upload an image of the existing interior of your kitchen or bathroom, apply a different ‘mode’ and ‘style’ to get a range of images that you can then refine further.

My prediction for 2024 is that we’ll see a clear distinction between application that are fully AI based and products with AI features. The former group will consist mostly of professional service products like Harvey, Robin AI and H&R Block AI Tax Assist, as I believe that the world of professional service (and its processes) is ripe for innovation through AI ‘copilots’. The interfaces of such products will need to double down on providing users with the right level of guidance for their prompts.

For example, Robin AI — an AI copilot for drafting, reviewing and querying legal contracts — offers guidance on how users can improve the quality of the answers that they get from Robin AI’s Contract Copilot:

Explain in plain english [insert clause number]
What are my [right / obligations] in [clause / contract]
Does this [clause / contract] contain [insert terms]
What happens if [insert event]
Can I [insert scenario]
What does [insert clause number] mean
Can you summarise [insert terms]
Draft an email to [insert request and audience]

Finally, search is another area that will change drastically due to generative AI being implemented by the likes of Chat GpT, Bing Copilot and Perplexity. The challenge is to make it as simple and fast for users to get the right search results (and refine their search prompts if required).

There’s a real opportunity to improve the quality and relevance of search results, and companies like Perplexity AI have made it their mission to provide information that is free from advertising influence and feels highly relevant to the individual user:

Perplexity’s interface allows users to ask follow up questions. It also addresses the lack of transparency inherent in a lot of AI output by highlighting the sources as well as the steps the copilot took to arrive at the answer in question.

Main learning point: How will product designers and developers apply Tesler’s law to generative AI? Who will carry the burden of the complexity and probabilistic outcomes of generative AI — the product developer or the user? I’m looking forward to seeing the design of AI based products and features evolve in such a way that it becomes easy for the user to get the right outputs.

Related links for further learning:

Generative AI and Tesler’s Law

Written by MAA1