Will AI replace humans?
A look at Models for AI/human interaction
One of the hot topics in the ecosystem has been the question “Will AI replace human jobs”? A lot of the discussion on the topic centers around the capabilities of the AI (ie can AI write marketing copy, can AI write emails etc), but I think the more important question is on HOW AI works. When I think about how AI works in this context, I don’t mean the guts of the LLM or diffusion model, but how in practice a process is applied the uses AI to create adequate or exceptional results (depending on what is needed). At Gen AI Partners we look at 4 models of human/AI interaction when designing systems. I’ll describe those models and then revisit where AI will augment humans and where it may replace human tasks.
Methods of working with AI - from mostly human to mostly AI
Four Models for Working with AI
Model 1: Augmented Human
Augmented human interaction is where humans produce the bulk of the work and AI is used to improve, review or carry out repetitive work based on the human starting point. This is probably the most familiar type of AI interaction as we’ve almost all been using things like spelling and grammar checks and auto-complete for years. The level of “augmentation” will increase significantly in the coming year with both MSFT and Google rolling out augmentation tools in office and Google Workspace.
Google Workspace is rolling out templated augmentation across experiences
Model 2: AI following human process
AI following a human process is where humans provide a prescriptive process or framework that guides multiple AI components to work together. This is one of the more excited and emergent interaction patterns and is often implemented through scripts, automation flows or AI agents. The process can be designed to mirror humans or can merely be one that is created by humans to provide more guardrails on how AI works within more complex processes. This can be an important pattern for enabling objective based AI commands. Less obviously, this can be an important step in enabling creativity in AI systems. Consider the example below asking ChatGPT to write an ad.
Direct Approach:
The direct approach yields content highly associated with Cheetos and Spring
A simple human defined process results in something that is similar but unlocks directed creativity:
Write an ad brief (LLM)
Create a character to star in the ad(LLM)
Design the character (Diffusion Model)
Write the Ad (LLM)
An AI image of the campaign character using an AI prompt
Marketing content generated around the fictional character and spring campaign. The human process yielded a different type of result. In both cases I’d expect a human revision in the real world today.
Model 3: Human-in-the-loop
Human-in-the-loop is similar to Augmented Human - but in this context AI does the bulk of the work and humans serve to prompt, review/reject and do final editing where needed. This is one of the most common patterns right now as there are often AI trust gaps and accuracy gaps in many cases. A couple of the most common workflows using this pattern are Github Copilot and Midjourney (via Discord)
In the Midjourney flow a human would prompt, review 4 potential options, then have editing options to upscale, add after effects or create variations
The initial mid-journey response gives 4 variations (at a lower resolution) for a human to choose by default
The upscaled higher resolution version is potentially usable but also provides options for further human interaction
Model 4: AI Automated
In this interaction pattern, process or workflow is fully automated by AI and a final version is presented to end consumer without any opportunity for intervention. The setup can be either human or machine generated.
It’s actually a bit of a stretch at the moment to come up with examples of this in action, currently this is limited primarily to parts of a workflow like Web Search (which is usually a part of something bigger) and used to augment an existing experience where the review summarizations are additive to the existing search result.
Fully automated review summaries compliment more traditional search content in Google’ experimental AI search
SO… will AI replace humans?
It’s complicated, but I think we can make reasonable predictions by looking at two dimensions:
The completeness in which AI serves the full workflow
The requirement for accuracy of the use case
There’s probably a second division within accuracy requirement that subdivides based on the risk of an error (how bad for you is it if AI screws up?) but to keep things simple we can look at it like this:
If this grid holds true, basic workflows with low accuracy requirements like writing a meeting summary or synthesizing feedback into bullets will be the first areas that will be replaced. In many cases, replacing these tasks will reduce a part of someones job instead of replace a complete role.
In this case and in the top-left and bottom right quadrants, it get’s more complicated on whether jobs are lost based on efficiency gains in roles. If five marketing professionals can do the work that ten could do previously, surely five jobs are lost right? History has shown that it’s not usually that simple, we’ve been through many technological advancements that reduce work for businesses, and in most cases the result is a shift in roles for workers. This chart from FEE article does a good job showing a macro perspective on how workforce changes with technology.
Shifting labor force roles due to technology is NOT a new phenomenon
The key takeaway for workers who fear AI replacements is that it’s likely overstated, but it may be useful to think about how your role will evolve or if there will be new roles available that are more appealing.
I recently attended the Brxnd.ai conference in NY and one other exciting thing to consider (raised by Mark D’Arcy) is whether or not the quality of work in your domain will be improved, many of these examples are considered as binary, but in reality we don’t just look at software code or marketing copy as done/not done, but there are many levels of performance in the output and one of the exciting areas of AI augmentation is the thought that AI may boost the overall level of performance of humans.
