On GenAI

Generative artificial intelligence (GenAI) has become a dominant factor in many of my areas of interest — such as research, programming, and education. However, the routine use of GenAI in these areas is currently a non-starter for me. Here, I’d like to consider how to navigate the consequences of this position.

I am fortunate to have the autonomy, for now at least, to limit my own level of GenAI use. However, I still need to collaborate with and rely upon those who use GenAI (to varying extents) in their programming, writing, research, and communication. I also need to be present in spaces where GenAI is taking resources, dominating discourse and strategic direction, and generally polluting the digital environment.

I need to do so without being a pain, both to myself and to others.

I’d like to respond to this challenge by seeking out and prioritising people, projects, technologies, and spaces where GenAI is minimised. I’d also like to emphasise approaches to my own work that are opposed to those that I consider to be associated with GenAI: simple instead of complicated; succinct instead of verbose; clear instead of diffuse; personal instead of impersonal; bespoke instead of generic; appropriately uncertain instead of over-confident; deep instead of shallow; responsibility taken instead of shirked.

But these responses are largely derived from avoidance and withdrawal, rather than the compromise and tolerance that is required in an environment dominated by GenAI. How best to do that is still unclear to me.

In developing a sustainable GenAI-minimised path, it will be helpful for me to be mindful of the foundations of my reluctance to use GenAI. These include:

  • No data provenance. We don’t know what data were used to construct the models, but we can be almost certain that much of it was gathered without authorisation, agreement, or credit.
  • Non-transparent operation. We don’t know the specifics of how GenAI models produce their output, or the criteria by which their internals have been optimised and tailored.
  • Corporate ownership. The power and wealth that is involved with the control of leading GenAI models is concentrated within a small number of companies — and I have no confidence that such power will be wielded responsibly.
  • Digital ecosystem destruction. The free and open-source software that we have become used to, such as Linux and Python, has enormously lowered the barriers to learning and utilising computational methods. That is being eroded by computation becoming dependent on which GenAI models you can access and how much money you can spend on tokens.
  • Natural resource usage. GenAI is resource-intensive and inefficient for routine tasks, with their requirement for data centres and electricity being opposite to the desirable direction of natural resource usage.

The technical foundations of GenAI are amazing, though, and I hope that my identified objections can be overcome in the future. For example, I think it can be transformative for delivering adaptive learning at scale in its capacity to act as an ‘intelligent tutor’, which can potentially achieve outcomes approaching or exceeding those obtained with human individual tutoring (e.g., VanLehn, 2011). It supports improved communication among those speaking different languages, which is vital for reducing the immense advantages given to native English speakers. I would be particularly keen to use it to improve the quality and reliability of code (and prose, potentially) through informed review mechanisms.

For now though, I will aim to persevere with my personal policy of minimal GenAI usage, hopefully supported by (and supporting) others with similar perspectives, while being productive and collegial in my interactions where GenAI is an intermediary.

The views described here are mine and should not be taken as representing those of my employer.

References

  1. VanLehn, K (2011) The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.