The million dollar question

That everybody in the LCA world and beyond is asking, right now, is whether AI will steal their jobs. I can tell you it won’t steal mine! Because my primary job is making LCA jokes. And AI has a terrible sense of humor, so I feel pretty safe. Sidenote: I won’t make a blog post about LCA jokes any soon. These are reserved for in person occasions, mainly conferences and similar academic and expert gatherings, where I can annoy people with such jokes. Only for humans, sorry dear AI!

When it comes to the other job though…

…we might have a problem, both you and me. I am talking about doing LCAs.

There is already a large group of people working with AI integration in LCA, and by this I mainly refer to using generative AI and Large Language Models. I have seen a bunch of company demos already about AI-supported tools, and prototype systems that can help you calculate a carbon footprint score, or scroll through a number of datasets to choose or get suggested the best one. Or more advanced software setups where you can use prompts to “talk” to a LCA database and interrogate that.

At the recent brightcon conference 20271 there was a keynote on AI by my super AI-expert slash guru colleague Lotte A. Thomasen and an entire session on AI in LCA. I should mention the PFC bench initiative by Krishna Rao and James Joyce, tried at Brightcon, where I could compete against AI in choosing datasets and guessing carbon footprint scores2.

We have also been working on AI-assisted LCA database use, check the work of my good colleague Ning An. And we can see similar work popping up, so much that this research already feels old before it’s even published. In fact, we are exploring new directions in terms of what uncertainty does the use of AI introduce in LCA, and how can we use these systems, rather than building more systems.

Anyway, right now, AI still produces continental-breakfast-type of content3. Which is a way to say the output is…poor. Maybe it’s around the corner, but I haven’t seen yet a commercial product shipped where people have an actual stake (= money) that is fully AI-based. The LCA scores are sketchy, the documentation is absent or poor, is you ask a couple of questions to the chatbot, you won’t get far.

But a lot of development is ongoing, and things are moving fast.

So it is not unlikely that soon we might be able to see LCAs made by AI, where the AI builds LCA models, produces impact estimates and makes reports, that might even be trustable. The practical act of generating LCA results might be over soon, or at least much more automated.

What is left for us poor LCA experts to do?

Here a bold prediction, bolder than Wanna Marchi can predict: the AI might steal some part of the job, but won’t steal the job as such.

Think about it: we can now produce, without AI support, LCA result much faster (way way faster!) and in much larger amount than we could do five, ten, or 15 years ago. Ok also compared to 20 years ago, yes I have been around this space for so long (damn, send help). This has not reduced the amount of LCA expertise needed and the number of LCA experts. Jeez, there are even more of those than before! There are even those who call themselves LCA experts and know nothing about LCA! Ok that’s another story.

Anyway, yes, this automation has shifted the type of expertise needed, and LCA people now need also to understand some things about LCA data formats, about computational structures, about simulation, about statistics, programming, software, etc. etc. on top of knowing all the other things (e.g., what is allocation…)4. In the end, AI is just more automation. So much that every time I read the term AI somewhere I try to read “A-utomat-I-on” instead. And we might expect a similar effect as we have seen so far. With shifting competences.

But I like the good old competences

These are going to be needed even more. Another prediction, in your face, tech lords! Two reasons.

One reasons is that AI is a peculiar tool, in the sense that it has a stochastic element, or how do you call that. It is not like a calculator or like a spreadsheet in MS Excel. It does not return exact results and not even consistent results. Would be nice though. But in my understanding, this will not change much. And thus we will always need LCA experts to validate AI outputs, and to be able to dig deeper into the details of the analysis and be able to explain it. Because for so much that humanity nowadays likes black boxes, and that people like “single number” answers, even a single score LCA is useless if you can’t explain the result. And here comes the second point.

There is a natural bottleneck in how much a person can intake. After all, LCAs are for people to be understood, and are supposed to support decisions (of people). And even if you can produce thousands of LCAs, somebody will have to interpret them and take action based on that. And the more information is produced, the more we need to be able to discern the signal from the noise, and select and weight this information and what is important and actionable for us and our decisions. You can’t trust the machine to do that for you. Yes you can ask the AI to summarize 1000 reports, and to answer all your questions, but if you don’t understand the answers…it’s useless, or extremely time consuming 5. Only the person with the right LCA expertise can help you understand the automated LCA results.

Bonus point, but very important: in the very end, commissioners of LCA studies will always need to be able to blame somebody for the poor results on the performance of their products. And blaiming AI is not very satisfying, it feels so much better to blame a real person: the LCA expert. And here you have it, there is still need for LCA jobs in the future.

 

  1. That we hosted in Aalborg btw. Yay! Free work just for the fame! ↩

  2. After playing this game I gained a lot of self-respect for my own LCA expertise: confidence boost! ↩

  3. I stole this term from Joao Oliveira dos Santos at UniTwente. Who’s also the only person in this planet who reads this blog so he will catch it. Thanks Joao, brilliant formulation. ↩

  4. I have also noticed a weird phenomenon where some LCA practitioners who are very good at designing complex simulation seem to ignore key LCA concepts or problems (e.g., allocation again) that undermine the very base of such complex models. And on the other side of the specturm LCA practitioners who like laywers can cite by heart specific articles and commas from a number of standards and guidelines but have no idea of what is the mathematical structure of a LCA model. Specialisation, wow. ↩

  5. Borges would probably say that if you use all this time asking the AI, then you’ll eventually become an LCA expert yourself… ↩