A couple of weeks ago, I wrote about how some of the errors we see with written content are going away after the advent of generative AI. I argued that an error (especially in written text content like emails, social media posts etc.) feels different now, as AI is eliminating most or all of them today. I also wrote that some of these errors that we make in content define the humanness of it. A small deviation signals it is not artificial, that there was a human behind the scenes.

Today I want to expand on how human errors are transforming into something different, and impactful (specifically in knowledge work context where AI is rapidly going mainstream). Essentially, human errors are undergoing a transformation in terms of size and shape.

The scale of errors is changing

AI gives you a lot of leverage now.

Let’s consider an example in software development. Today, you can write 8x (or more) the code you used to write before AI. Essentially you have agents writing code for you and you are limited by the tokens you can use and the time you can spend on coding, although the former attribute has a higher weightage. And what happens when you deliver 8x of what you did earlier?

A few of the dynamics need to change or adapt:

  • There is potential for new errors to increase. Not exactly by 8x, but definitely not zero. While AI is good at creating code, it might not (yet) create something that is free of errors. These may not be technical errors; they could be the way the result works, which in turn deviates from the expectation.
  • Testing needs to grow as fast as generating code. If you are writing at 8x speed, you need to be able to ship at roughly that speed to keep up. And for that to happen reliably, you need to test newer use cases and possibilities.
  • Human in the loop: AI is still not reliable enough to act fully autonomously for mission critical jobs. Let’s say there’s a new humanoid robot that administers medicines to old and unwell people. You cannot rely on this unless it’s a purely deterministic workflow. Generative AI can hallucinate and could potentially cause catastrophic results here. So when the output of these use cases grows, the need for humans also grows in these areas.

The shape of the errors is changing

Like I mentioned in my earlier post, we are no longer seeing typos and grammatical errors. Our errors are shifting towards more impactful areas, for instance:

Wrong judgement

Wrong judgement is saying yes to an AI response, when you should have reviewed it better and asked for a change. AI can increase your cognitive load, eventually draining your attention span in a task. I have personally faced it. When you have said yes the last three times, there’s inertia to keep saying yes. And the fourth time, perhaps the AI is wrong, but you didn’t notice it. The shape of the error depends on the gravity of the task in hand here.

Not asking what could go wrong

This is more of a planning error. If you let AI write a thesis for you, and you never asked what could go wrong, you are essentially playing Russian roulette, perhaps a mild version of it. Think about those companies that shifted to AI chatbots for handling support, but never asked about what could go wrong. User experience declined, complaints were closed before being properly addressed, and eventually, loyal customers were lost.

Trusting a recommendation that is inherently broken

This is similar to wrong judgement, but it starts with us. We should first ask: how much can we trust a recommendation from an AI model? I would say very much. AI is super helpful. LLMs are super helpful. But there’s a thin line between where it can be fully trusted and where you have to rethink before trusting it. The more you know about that small gap, the better you are at working with it.

A long-term perspective on fully outsourcing our thinking

I don’t want to go astray from the original topic, but this is worth considering. Humans of today have been trained on their experience and environments, which fairly demands thinking. Decisions that we have been taking for a long time have been a result of thinking, sometimes deep. Our neurons fire and connect inside our brains to come up with a better response or decision when we think. But now, that is going to change. Now we can outsource thinking to AI, which is not bad in general. It makes our lives easier, helps us do things faster, and live more. However, what would happen to our ability to think after 10 years of very high dependence on AI? We don’t know yet. Now, if your thinking is not superior (or functioning optimally) in the long-term, how can you make sure you are doing the right thing when the AI suggests something? What would be your reasoning when you cannot reason well?

While the above might sound like a pessimistic or overly cautious take, the bottom line is this: we should embrace AI, but also make sure it is not detrimental to our cognitive abilities over the years.

Our brains haven’t evolved enough to easily realize that regular mental and physical effort is essential for our health and well-being.

When we use AI extensively, it can make us drown in cognitive debt, to the point where we can’t reason well enough to tell what is an error and what is not. Our brains haven’t evolved enough to easily realize that regular mental and physical effort is essential for our health and well-being. If that was not the case, then nobody would have any trouble going to the gym, or having a difficult conversation that would make us better people. Now, in 10 years’ time, with that kind of advancement and ’less thinking’ as a default, the possibility of error becomes higher, unless your definition of error changes. Unless the AI is self-correcting, the quality of our decisions degrades and eventually, ‘quality of your decisions’ = ‘quality of your AI’.

And unless your AI is self-correcting to the right degree, the results could be potentially harmful to humanity at that point. But we should think about factoring in our thinking and our roots as we develop this technology more. Perhaps all the AI pioneers are thinking this way. Will the AI journey going forward be more structured and controlled? Will our lives keep only improving because of the advancements in AI? Will we see lesser errors after AI? We have to wait and see. A balanced adoption approach might be the answer.

References and Further Reading