Recently there has been a surge of interest in generative AI models such as GPT and Stable Diffusion.

An interesting quality of these AI models is that, rather than making a single prediction for a given input, they can generate many different sample outputs. The human operator can then scrutinize the sensibility of the output and can choose to accept the model's output or edit the prediction in some way, e.g. by adjusting the model inputs.

These new models make it seem as if we have transitioned away from the rigid supervised learning framework of machine learning, where models would make an exact single prediction. In the new framework the model output is seen more as a suggestion where a human subsequently performs some kind of quality assessment.

To describe this new framework, I am proposing the term Suggestive AI or Suggestive Models. Because we are now using models that make, possibly multiple suggestions for a given input. An academic audience might say this new paradigm is just a switch to generative models, and they would be right in that many of the recent large AI models are generative. However, in my opinion, the term generative models does not quite capture the way people are using these technologies, so Suggestive models seems more descriptive.

I expect that Suggestive AI will be much easier to integrate into current business processes. Most decisions in business involve some degree of subjectivity, due to e.g. uncertainty or unobserved inputs, as such these are hard to automate with models that can only make a single rigid prediction. Using suggestive models we can represent the full range of possible decisions. However, in this case, a human is still involved in an inspection role, rather than being responsible for crunching through all the decision making logic themselves.

This of course opens up many other questions. What other applications will be made possible with suggestive AI? What tools and processes will be needed to better develop and utilize Suggestive AI? What are the security and societal concerns if Suggestive AI becomes widespread? Will we need new tools to better enable the interaction between humans and AI models?

The Failed Promise of AI Automation

Contrary to what many headlines reported at the time, we have not seen widespread adoption of AI and machine learning in most industries. Only in select industries, and for specific tasks that are very repetitive and homogeneous has AI readily been used for automation.

So was AI simply overhyped? Why did the adoption of AI fail, despite the impressive results coming out of research labs?

My hypothesis is that most AI models were using prediction targets that were too rigid, and that Suggestive models have fixed this problem.

For example, in supervised learning a training dataset consisting of example inputs and outputs is used to calibrate a machine learning model. The model is supposed to generate single exact outputs that are the same as in the training output for the corresponding training input. The hope is that unseen data points will be sufficiently similar to those in the training dataset, that the model will still work well on new unseen data points.

This kind of supervised learning works well when we have very complete data, e.g. we know all the relevant inputs and we have wide coverage of the possible input and output examples. In real applications this is almost never the case. Thus when using this framework, the machine learning model is pretty much guaranteed to be wrong some of the time. Since the model generates only a single output, there is not much that can be done if it is wrong. Additionally we do not know a-priori when or how the model will fail, making it hard to develop contingencies.

Contextual Suggestions + Human-in-the-Loop is Better than Point Predictions

There are two main factors of Suggestive AI that fix the problems with old AI. The first is the move from a single point prediction to many contextual suggestions. The second is the involvement of human operators. Additionally, both these changes are necessary, one or the other is insufficient to fix the previous problems with AI.

For example, suppose we decided that we are just going to add human operators to an old machine learning model that makes point predictions. The problem here is that humans are pretty good at determining when a particular decision is good or bad, but, due to limited attention, they may not be as good at absorbing many different data points to generate good decisions. As such adding humans to the previous process might help prevent mistakes, but it would not tell you how to fix things when they went wrong.

Similarly, we could try just switching and old AI model to a suggestive model that makes many possible suggestions. But often we can only make one decision, so out of the possible decisions we would need a way to pick one. Sometimes this problem has been solved by adding an additional AI model that acts as an assessor of the suggestions, for example the actor-critic approach in reinforcement learning. However an assessor model still relies on having good data, and so can also fail in real world scenarios, thus again requiring human intervention.

The Path Forward for Suggestive AI

The recent interest in Suggestive AI shows that it is a rapidly developing area. Suggestive AI is also quite different from previous machine learning approaches which raises some interesting questions.

Potential questions revolve around how the typical model development process will change? AI and ML practicioners have developed good processes around model development, e.g. training data collection, tools for creating labeled data, training algorithms, training and testing splits to assess model quality, hyperparameter tuning, model logging and debugging technologies etc. How will all this change with the move to Suggestive AI?

E.g. What does training data look like for a Suggestive Model? How much data is sufficient? Do we still need labels, or could we just create a big pile of data and let the model+human operator figure it out? Will we need different tools to collect and curate data for Suggestive models?

Currently Suggestive models are trained using similar methods to previous AI models. Could there be better ways of training Suggestive models?

How do we debug Suggestive model training and how do we asses their quality? Since we no longer have point predictions it becomes more difficult to create a single quality metric. Will we need tools that make it easier for humans to inspect the model manually?

Finally are the additional tools needed to improve the interaction between the human operator and the model? What kind of user interface is most suitable here? Will there be differences between running this on mobile phones, vs desktop computers or other hardware? Additional monitoring tools will probably also be needed.

Finally another important questions is in what other areas Suggestive models can be used to improve current processes. We have seen lots of use for creative subjects like art and writing where suggestions can be very free-form and there is lots of data online. What other areas where decisions are subjective could be improved, Law perhaps? Hiring? Generating business presentations? Customer support? Those all seem areas that could benefit from Suggestive AI.