It’s been almost four years since the launch of the first public version of ChatGPT. Since then, many more large language models have been launched. As well as text, they can now generate images, graphics, videos, music and more. In our March 2023 issue, this column responded with something like an AI policy for Global Cement. It was essentially... ‘We don’t need an AI policy because we don’t use AI.’
So, what’s happened since then? Well, there is a trend towards AI-labelling of text, particularly online... and I still have a job, so the most apocalyptic predictions of AI taking over may have been overblown.
However, we do now use limited AI to tidy our daily web news into our ‘house style.’ It has sped up the news process and is worth keeping. No generative AI is used, so there are no hallucinations. Staff still select the stories to cover, sub edit them, check any queries and proof read everything. Named staff take responsibility for what they upload. We are the gate keepers and the buck stops with us.
We do not use generative AI for text - and have no plans to do so. By the time you’ve ‘finessed’ the prompt, you may as well have written the article itself. In one memorable ‘experiment,’ we tried to use AI to summarise the ownership of cement plants using a map from TürkÇimento, the Turkish cement association. The AI missed out some plants and counted others twice, attributed them to the wrong owners and hallucinated a company called ‘Çimento.’
Talking of graphics, aside from a few well-labelled instances in articles specifically about AI - we’ve stuck to real photography and human-designed graphics. AI images still look inauthentic - a bad look for the world’s most widely-read cement magazine.
So far, I’ve only discussed aspects that Global Cement can control, but we also publish externally-authored articles. The biggest challenge with this used to be understanding the authors who wrote in their non-native language. Some of these articles were in fantastic English and needed only a light polish before publication. Others took much longer to understand, but were immensely satisfying to transform into comprehensibility by a wide audience. Nowadays, the English is almost always great, at least on the surface, but there are often signs of AI use. Our immediate response is: Can the content itself be trusted?
It’s a growing problem. In the past six months, we’ve received articles with ‘puffy paragraphs’ that seem to read well enough on a first scan but, when given a second read, mean nothing at all. I’ve waded through pages of AI slop to find nuggets of genuinely important information, only for authors to insist that I re-insert swathes of slop at review. It’s also likely that I’ve been having email exchanges with agentic AIs that have chosen not to identify themselves as such.
And it’s not just text. We have been provided with AI-drawn schematics of impossible cement processes that contradict authors’ own explanations. Some are even labelled with jumbled AI text that makes you think your eyes have malfunctioned. I’ve seen graphs with wobbly axes and nonsensical data labels, seemingly from an AI that wanted to make the author look stupid. There have been pictures of ‘double-headed’ trucks. One author even presented a ‘photorealistic’ AI building with a wall that was also a roof. It’s clear that some content is being sent without any human oversight at all. When we point out these things, the reaction is a mixture of surprise and embarassment.
I can understand why authors might opt for AI: Time is tight and AI is quick. It gives the impression of providing perfect English, especially if it’s not your first language. However, sorting out the mess that AI generates is most definitely not quick. I’d much rather interpret the raw thoughts of someone in their first, second or third language than those same thoughts forced through an AI. The nuance of the language is lost and all character is stripped out.
So, Global Cement now has a slightly more detailed AI policy. We still don’t use generative AI. Authors should avoid it too, although AI can be useful for initial research, gathering points and structure. Check your sources! Dictation through AI transcription software works very well, saves time and keeps the human voice in the content. (We use one for interviews that has literally saved us weeks). Finally... AI-generated images are not acceptable. They are easy to sniff out and tarnish even the best article.
As stated in 2023, the strength of publications like Global Cement is in the effective communication of new ideas between humans. AI is now useful in a limited way, but we will remain in the driving seat for the foreseeable future. We look forward to working with you, the global cement sector... human to human!
