Upon reading "Power to the People: The Role of Humans in Interactive Machine Learning." by Amershi et. al. I stumbled across the following sentence:

“Consequently, interactive machine learning can facilitate the democratization of applied machine learning, empowering end users to create machine-learning-based systems for their own needs and purposes.”

Is democratization of AI even a good idea? And should interactive machine learning contribute to democratization? I am concerned about what would happen if every human on this planet could simply build whatever AI application they wanted. AI might be weaponized and result in harmful effect on our society. With the interactive approach that Amershi et al. are proposing the middleman, the expert who changes design parameters based on user feedback is skipped. When end users become teachers to AI systems, they need to be educated about the inner workings of the system to allow for a shared mental model and good user interaction.

Regarding the question of established conventions or unquestioned assumptions in the application of AI, I have two thoughts. Firstly, I am wondering if AI must always come in the shape of a software program. Does it always have to feed on digital data, or can we create an AI system that is for example based on nature like tree or a mycelium? Secondly, it might be important to question the type of tasks that we use AI for. We use it to make Ad targeting better, recommend us music and movies, translate between languages, or enable autonomous cars. Most of the above-mentioned applications are being developed by large corporates with the aim to maximize profit. But who asks us humans if we even what an AI to do that? Nobody asks the end user what we want to use AI for. We simply let companies decide for us. But where there is no profit, there is no product and no AI. What if instead of applying these powerful algorithms to maximize profit, we dedicate them to creating social good? Can we use AI to counteract climate change, foster social cohesion or enhance mental health?

New tasks:

1. Surgical sonification could make use of interactive AI to learn from surgeons how to best target an anatomy and how to effectively sonify anatomical regions of interest.

2. Sonar information measured under the sea at different locations could be used to reconstruct and learn about marine animal migration behavior?

3. Can we use AI in an interactive architectural design process where both the architect and the AI draw floorplans, sections and elevations together?

4. Can we slow down global warming by creating a pro-climate biased AI that influences our daily decision-making process in climate friendly way?

5. Can we use AI for psychotherapy, for example as an aid to complete out-of-session homework?

6. AI could be useful to help us adhere to our circadian rhythm by providing us with the exact amount of light intensity at each time of the day. (e.g. through an AI-controlled lighting system)


Challenging existing conventions:

7. Instead of text-based prompts we could use thoughts as a user input to AI language models to reduce communication issues.

8. In situations where voice and key / button inputs fail us can we use gestures as an input to AI systems? This might be useful when rethinking AR and VR interactions that include interactive AI components.

9. Instead of obscure recommendation algorithms, can we make explicit when and from where we are taking user data as input to the machine learner and let the user decide at each instance which of their clicks or decisions, they want to transmit to the AI to alter recommendations?

10. Instead of training on large datasets, can we train on very small and incremental information that is stored and then altered what was learned based on new input (just like children are taught by their parents, slowly and over time)


The above is simply a list of rough ideas. I could see a lot of negative implications associated with each of the ideas, which are however not regarded within the scope of this critical response.