Experience report: Use of AI functionality in generating content for an article

Greetings, friends of motor racing, 

By the end of 2025, artificial intelligence has become an integral part of everyday life. This applies to both the private and, increasingly, the professional sphere. This also applies to journalists, for example: transcribing interviews, translating them into the target language, and checking the spelling and grammar of the finished article – there are now tools that use AI for all these purposes.

But are expensive and complex specialised AI programmes really necessary? Or can a graphic designer simply be replaced with a free, language-based AI system with an image generation function?

I recently explored this question and used ChatGPT-5 to transfer the current liveries of the ten Formula E teams to the recently unveiled Gen4 car. For this test, I created a free account and upgraded it to ChatGPT Pro for a free trial period thanks to a promotion from my mobile phone provider.

As I have been using artificial intelligence to analyse source code in my main job in software development for quite some time, I was familiar with the basic procedure: In the prompt, you have to give ChatGPT very precise instructions on what to do and what not to do.

It will come as no surprise that the very first results did not quite match what I had imagined. So I had to modify and expand the prompt several times to get really usable results.

For example, the Gen3 liveries that served as a model had Hankook branding on both the bodywork and the tyres. However, since the Gen4 cars will be running on Bridgestone tyres, it was obviously not desirable to use these logos.

Initial results were unusable

But here, too, it became apparent that despite precise instructions to ignore the orange patches on the sidewalls of the Gen3 tyres, ChatGPT repeatedly incorporated them into the Gen4 designs. Body parts such as the rear wing were also ignored in the Gen4 template, and the design of the Gen3 vehicle was adopted.

Lessons learned: Not only can AI make mistakes – as ChatGPT itself points out – but it also still makes them very frequently. In some cases, explicit instructions, even if marked in the prompt with "Very important: ...", are simply ignored.

Nevertheless, I have to say that after several adjustments, the results were perfectly usable for my purposes. In total, I had to run 34 renderings to transfer the ten paint jobs to an acceptable standard. Parallel tests with the predecessor engine ChatGPT4 have shown that the quality of the images generated has improved significantly. It can therefore be assumed that this trend will continue with further new versions.

Perfect results are not achievable - yet

However, even the "usable" results are miles away from being comparable in quality to the work of a graphic designer. It may be sufficient for a simple article along the lines of "Let me show you what this could look like", but if, for example, teams wanted to publish such images themselves, it would fail because the logos and lettering of sponsors would show AI-typical deviations from the original.

This is also the case with the lettering of e-formel.de, which should be clearly visible on each of the paint jobs: from e-tormel.de to e-formel.ole and e-formeiide to eformel:die.

Formula E Gen4 car in the design of the 1992 McLaren MP4/7

All ten results (the image used here in the blog with the design of the 1992 McLaren MP4/7 is of course not included) can be viewed in the following article at e-Formula.news: Link

Incidentally, the version of ChapGPT5 used at the time still allowed original lettering and logos to be used in renderings, even from cigarette brands such as Marlboro. This is no longer permitted due to a software update. "Unfortunately, I cannot render the image in the desired form because your description includes the representation of a real tobacco brand (Marlboro) and thus trademark elements protected by copyright and advertising law.’"

Conclusion

Finally, a brief summary: The time required – calculated for a single article – was immense. Generating each individual image took two to three minutes, and the prompt had to be revised repeatedly. Since one person accused me of wasting considerable resources by using AI, I calculated the energy consumption of my project.

In the worst case (i.e. using the highest values I could find for ChatGPT's energy consumption for a text response or image generation), I consumed a total of 0.624 kWh of energy for the 34 image generations and 25 additional text responses.

To put this into perspective: this amount of energy would not even get a Gen4 car halfway around a Formula E track. But ChatGPT itself says:

⚠️ Note on interpretation

• This is a theoretical maximum value; actual consumption is likely to be significantly lower (probably 3–10 times lower).

• Many processes run efficiently in batches on OpenAI servers ("batch inference"), so that energy is shared per request.

• The actual value depends heavily on model size, GPU utilisation, cooling, image resolution, etc.