Let’s talk about Artificial Intelligence: How Do We Use It Ethically? There’s been a lot of interest in this topic recently, as well as quite a lot of emotion and misinformation.

I wanted to set out how the Safety Artisan uses Artificial Intelligence, or AI, ethically. But to do that, we have to understand what the ethics are and the ethical questions around AI.

Many of the arguments about AI centre around two questions:

• First: are we infringing copyright, the intellectual property of other people, by using AI?  
• Second: by using AI, are we putting people out of work?

Copyright Law & AI Training

So, let’s look at the first issue. Before we can make any judgment about what AI does we need to understand Copyright law and how AI is trained.

First, we should note that not all published material is subject to copyright. Many US government publications are copyright-free on principle, because they were paid for by the American people and may be freely used by them. There is also lots of online material published under Creative Commons (CC) 3.0 or 4.0. CC 3.0 may be copied, changed and republished for non-commercial use, but CC 4.0 may be used in this way for commercial gain. Both usually require attribution to or acknowledgement of the original.

Copyright law (or is it just convention?) says that we can quote copyright, provided that we attribute the information to our source. (Let’s skip the issue that when we quote somebody, we might be quoting a quote; the true originator might be somebody else entirely, but we might not know that.)

If, however, we copied someone else’s work word for word without attribution, then that may be a civil offense. We have passed off somebody else’s work as ours, and we have infringed their copyright, their Intellectual Property (IP).

Interestingly, it is not an offense to look at somebody else’s work and to write a précis or summary, or reword it – to use those ideas but to put them in our own words. That has always been legal. Indeed, this is the basis of reportage and criticism. Every time we write a book, an article, or a blog post, we are probably taking information from many sources and putting it together in a unique way, using our own unique words. That’s normal and ethical practice.

What is curious is that different standards seem to be applied to AI. Perhaps it is because those who criticize it don’t understand how it works.

How Does Artificial Intelligence Really Work?

AIs, particularly Large Language Models (LLMs), are trained on vast data sets which are freely available online. The AI engine does not read and store or copy these sources, but learns from them. It is remarkably like the way humans learn, and sometimes deliberately so. AI adjusts itself depending on the information that it receives. Just as a brain is a complex neural net, AI is a complex set of algorithms; indeed, it may use an artificial neural net, deliberately aping the human brain.

So, we can put aside the notion that AI is copying other people’s work. When an LLM is trained on billions of web pages, it would simply not be feasible to store all that information for the AI to refer to when we ask a question. Instead, the AI interprets the questions that we ask it and then it responds using the algorithms which have been trained on the real data.

It looks like a reproduction because it is so lifelike, so incredibly realistic. We think it must be working from stored, i.e., copied, knowledge. But it isn’t.

AI’s uncanny ability to process data does not come from real intelligence. Rather, it is Machine Learning, and it is a dumb machine. What it is really doing is using brute-force power to make up for its lack of intelligence. A single microprocessor can carry out hundreds of thousands of operations in a second. An AI model may be run on hundreds or even thousands of such microprocessors in a server farm.

Using AI at The Safety Artisan

So, let’s look at some practical examples. How do we use AI at The Safety Artisan?

First, we use LLMs to help us generate some blog posts.

A typical process is that I make a video where I’m talking live to camera, often without a script. I’ve got the knowledge of the subject to be able to do that, with structure from presentation slides that I’ve prepared. However, prose that works in spoken form – when you can see my face, hear my voice and interpret my words – doesn’t work so well when it’s text on a page.

I use an AI to take my recorded video and extract my words into a raw transcript. I still use a tool called Pictory AI (https://pictory.ai) in order to edit my videos, which it does by manipulating the AI-generated transcript.

I then used to pay my daughter, who is an English literature graduate, to edit this transcript. All through the COVID epidemic, she edited my work and then sent it back to me. I paid her to do this, and it supplemented her furlough payments while she was laid off.

What I can do now is take that AI-extracted transcript and feed it to ChatGPT, an LLM (GPT stands for ‘General Purpose Transformer’). I ask it to improve the transcript and turn it into prose of whatever style I want (formal, conversational, whatever). It does an incredible job in seconds.

Lately, I’ve been asking ChatGPT to turn my work into simplified English in accordance with an international standard, ASD-STE100 (Simplified Technical English). I know of this standard because it is used internationally in engineering and maintenance publications when parsing unambiguous user instructions for tasks, which may be safety-related. Of course, we want such instructions to be clear and easy to read!

Another variation is to take an existing publication and to use ChatGPT to translate it. For example, I might take an excerpt from a safety standard or process guide, which may be written in dense and difficult language. I then ask ChatGPT to convert it into something more accessible and readable for a blog post article.

This is legal and ethical on several levels:

• First, the original source may be copyright-free, or it may be usable under CC 3.0 or 4.0. There is a huge amount of material online which is available in this way.
• Second, the LLM is not copying the material; it is making a reworded precis or summary of the material, which is not an infringement of copyright.
• Third, I give attribution to show where the information originally came from.

The last point is not always strictly necessary, but I like to do it anyway. It shows that I am drawing on an authoritative source, which is important given that I’m teaching people how to do safety in accordance with legal requirements or standards.

Another example is the use of Imagery

When I started The Safety Artisan in 2018, I paid a graphic designer (https://www.linkedin.com/in/samjusaitis/) I knew to come up with a logo and colour scheme for the brand. I was very pleased with the result, and his advice on website design – thanks, Sam!

Having done that, I might change that logo by feeding it again into ChatGPT and asking it to update the logo. Here’s the original and an example of ChatGPT content. I asked the LLM to come up with a ‘Millennial Minimalist’ version using Sam’s colour palette. (Why ‘Millennial Minimalist’? I asked Google what was popular with younger demographics, my target audience. How do I know what my target audience is? Because I did my homework.)

Similarly, a former employer paid Iain Bond Photography to take headshots of me and my colleagues. The images belong to the photographer, but the firm bought the right to use the photos. Ownership is quite a complex issue in Australia – see this guidance from The National Library of Australia. (N.B. I haven’t researched the law in other jurisdictions – it may be different where you are.)

I fed this original image into ChatGPT, which it manipulated. With mixed results, as you can see.

The original photograph is on the left, and the obviously stylised version in the middle. So far, so good. The version on the right makes me look like someone else (my wife says Ewan McGregor, my daughter says Andy Burnham or Ronnie Corbett!) and I find it faintly unsettling. This is illuminating, isn’t it? Something that is obviously derivative and artificial looks fine. Conversely, something realistic, but not quite right, looks false and wrong. I guess it’s down to authenticity.   

Another example is images I use in web pages and posts. I used to use pictures that I took myself, which were never very successful, as I’m not a very good photographer. Or I used to go to sites like Pexels.com and get free images there. I would then acknowledge the photographer who took the original pictures in the picture’s caption.

Now I can use ChatGPT and just feed it the text that I’m going to use. The LLM will generate a suitable image for me to use with the material. Sure, I’m not paying any artist, but as a small business I couldn’t afford to pay a graphic designer or photographer for such images anyway. I would have done it myself in the free version of Canva, or some other online tool.

Here are some examples of ChatGPT’s work.

All I had to do was give it the address of a recent blog post, “Supporting a Vision Worth Sharing”, and it did the rest. First, I asked for a Millennial Minimalist version, then an art deco version, and, finally, one done in the style of a vintage British railway poster. Each one took only seconds to generate, and they are all quite lovely. (Incidentally, the LLM made up a new Safety Artisan logo, which I later asked it to replace.)

Summary

If we are going to use AI ethically, then I dare to suggest the following principles:

• We need to understand copyright, Intellectual Property, Creative Commons Licences, and the conditions that the original creator imposes.
• We need to understand how AI (Machine Learning) works and what it does and does not do.
• Authenticity and openness about use of AI are key.
• Successful use of AI requires skilful direction, which comes from knowing the subject and doing your research.
• We still need to check the results from LLMs and other AIs.

What do you think?

Declaration: I dictated this article into my phone using Gmail voice recognition and edited it in Microsoft Word. I fed this article to ChatGPT asking it to suggest SEO hashtags for it. It did, but it also pointed out that I had incorrectly written “ASD-100STE”, whereas it is “ASD-STE100 (Simplified Technical English)”.