AI & Honesty

AI¹ tools are really good at producing stuff that looks like it’s high-effort/high-quality when it’s really not. This can be mitigated by putting in that effort by closely checking the work, editing it, actually putting in the effort, but that takes as much time as doing it yourself, without the AI tools. But who’s checking? Nobody, and we all know it. We all signed a legal agreement when we started using Claude, promising that we’re personally responsible for any output, that we check over every line. But who read that agreement? Nobody. We know we’re not being judged by the actual quality of our work because our managers don’t have the time to read code, let alone read it closely. They do have time to see that you’re making a bunch of stuff that looks like hard work and that’s what really matters, that’s what the system optimizes for.

I think a lot of people have figured out that they’re never going to be judged on the quality of the product, dutifully followed the pressures of the systems, and given up on producing high quality work. This, to me, is fundamentally dishonest, and embodies a real lack of respect for the other people around you. It’s like throwing your trash out of a car window: it’s Someone Else’s problem now, and people seem to think that they’ll never have to deal with being someone else’s Someone Else.

When I’ve probed other people about why and when they use AI tools, it all boils down to some form of “I don’t care about that stuff” or “I only care if it works”, etc. I think there’s actually a reasonable case for using AI tools for stuff you “don’t care about” as long as that stuff doesn’t impact other people, e.g. one-off tools, scripts, queries etc. As soon as it does impact other people, saying “I don’t care” is handing off responsibility to anyone who does care, and making them shoulder the load.

AI tools are often incorrect, sometimes subtly, sometimes not. AI users often pass on these errors unchecked, lending their credibility to the AI output. Two main aspects are important to how I interpret someone’s participation in this process: first, if the user knows that there is (or might be) an error, and second, if they think I’ll catch it.

For the first aspect, did the user even read what they’re sharing? Was there an honest mistake in the review, and they just missed the error? Or did the user not even read the output. If they didn’t read it, I’d interpret this charitably by assuming that they trust the output to be reasonably correct, or uncharitably as them not caring whether the output contained errors.

Second, we need to consider their mental model of the reader, too. If they think the reader will honestly read it and catch errors, perhaps there is no penalty for making such a mistake, and pawning off this work to Someone Else saves them time. However, it seems more likely that the reader won’t catch the error, and the AI user understands this. When we see work that seems to have all the signs of high quality and high effort, (i.e. all the polish that AI is good at) it’s natural to assume that the underlying work is high quality too! In practice this means that most of the AI output is merely being skimmed for surface level characteristics, and nobody is actually reading it!

There are a few important combinations of these scenarios, and there is formal philosophical terminology for these classes of deception. When someone says something that they know is false, with the intent to deceive someone, that’s called a lie. There are severe social penalties for lying, or rather for being caught lying, and it’s a serious thing to accuse someone of lying. A lie requires the teller to intend to deceive, and establishing that intent can be a tricky thing. A common interpretation of this behavior falls into a sub-category of lying, the bald-faced lie. In a bald-faced lie, the lie is so obvious, so brazen, that the listener instinctually assumes that there must be no intent to deceive. Here, we find it hard to fault the user, because they are the victim of the AI’s deception, and the lie isn’t really a lie and they are innocent. I don’t agree with this view. I believe that my coworkers are capable, intelligent individuals. I know that they are aware of the possible shortcoming of the AI models. Instead of maliciously representing a falsehood, I think that they simply don’t care about the truth of their statements. The core deception is in pretending to care about the truth, not any specific deception. In philosophy, the name for this is bullshit.