“No, I don’t want to hear what ChatGPT thinks, I want to hear what you think”
Said a software testing friend recently, when in answer to a question I responded that I’d researched this very topic on OpenAI for an upcoming interview and wrote down the salient bits on my notepad. It stung a bit, because I thought I’d been clever to get a second opinion, but it was a hard no on their part. The agency of thought had to come from me, what a computer had to say wasn’t what they were interested in.
Maybe in times to come we’ll look back on those kinds of exchanges wistfully, and Peter Kay-esque comedy sketches will be written (by computers of course) saying “remember when folk used to actually care what a human thought about the best way to do something!”.
Its been fascinating to see both the level of interest and the level of distaste emerging around this new tool within the QA Community. People are happy to take a look at it, and then very quickly seem to come to an irrevocable marmite-esque conclusion that either:-
“This is the future! It’s incredible what this thing can do – look, it helped me do X, Y and Z, and with far less pass agg than Stack Overflow. Love it”
“This tool is dangerous. We should avoid it at all costs, and if we do use it, treat it with extreme care. Relying on the information it produces (which is often incorrect) without the ability to critically evaluate it will lead to some terrible results. Look – I asked it X,Y and Z and it came up with some absolute mansplaining tosh that sounded great but there was absolutely no factually correct substance to it. Hate it.”
For those of you unfamiliar with Harry Potter, the evil Lord Voldemort was considered so powerful , terrible and omnipresent, that to even utter his name was something shocking. Everyone thought about him, and knew about him, but only those with incredible skills as wizards would dare to mention his name.
OpenAI and other AI tools such as Lensa may quickly become tools-that-must-not-be-named within professional tech circles. In other words, tools that a lot of people actually use, but don’t openly acknowledge for fear of retribution. A bit like a company advertising for a “manual tester”, or the quarter final of the FIFA World Cup being the most watched TV event of the year (in the UK). Just what the testing world needs, another thing to argue about interminably – hurrah!
As testers, I love that people are using their noggins to evaluate a new tool. And it has genuinely educated me to learn some of the more negative sides of the AI world – and there are plenty, so I am glad there are people out there who are talking about that stuff!
I also believe in confirmation bias, so we will look for information to justify our inherent beliefs and place less importance on things that seem to cause us cognitive dissonance by diminishing or trivialising them. We, at least, are still human after all.
So what can we acceptably use Open AI for?
However, I believe in shades of grey. I believe that there is a middle ground, and a set of acceptable use cases for this suite of models in particular which will evolve, many of which I am already finding myself forming the habit of using:-
AI as an explainer
Tech is full of acronyms, weird expressions with several meanings (hello Lambda!) and difference of thought. So much so that it is baffling to outsiders, or people trying to enter the industry, or even those of us who have several years under our belts in all honesty!
These smoke and mirror linguistics can feel gatekeepy, and it’s exclusionary to say that people have to learn everything through experience only, or somehow magically know all the same things you do. I recently met up with a group of new software tester recruiters who were overwhelmed with the amount of buzzwords, do’s and don’ts and terminology they had to get their head around.
I think for basic definitions such as this one, the AI is probably good enough to be reasonably accurate – at least as accurate as a google or stack overflow search. However, I’d be careful using it for more detailed information, or information about recent events as the data is only as good as the data set – which in OpenAI’s case currently ends in 2021. At the very least, when asking a tricky question you should try and cross-reference the key facts the engine is giving you elsewhere – we’ve all had that icky moment when we’ve realised a Wikipedia entry has been modified and we’ve already relied on the definition!
AI for experimentation (and fun!)
I think as testers, we love learning new stuff. We are as magpies to the shiny free tech glimmery gold. And, its got to be said, there is joy to be had from asking a computer to write a poem about something tech related in the style of a gruff yorkshireman/robot/sarcastic salesperson. There just is.
I used Postman Flows (anyone who reads this blog knows what a huge fan I am of that feature) to automate a workflow that checked an OpenAI auto-complete phrase and then output the results to a Slack channel using their API. In my case, “give me the top 3 headlines this week on Postman” – this could easily be leveraged into a scheduled run each week using the new scheduler on the collection runner.
There is so much to evolve and iron out here, with images in particular. For example, I’d caution against uploading any pictures of yourself, as the rights to what AI does to that data stop being yours (noone needs nude deepfakes). If you have a strong conviction against AI generated art (or art which has been moderated by AI without acknowledging or compensating its originator) then probably steer clear altogether. We will be having to ask ourselves moving forward with pretty much everything we see – could this be fake?
I remain just one of a multitude of opinions on this subject – mine more uninformed than many. So read this blog and take its advice with the same critically applied evaluation that you apply to the rest of your testing life. I will still be interested, curious and open to hearing and learning about the complex and evolving opinions on this topic.