Proof of human work
Don’t use AI to write for humans
2026-10-04
As of late 2026, there is little doubt that frontier large language models are more intelligent than human beings when it comes to reasoning and finding solutions to problems. Hallucinations are mostly a thing of the past, at least in publicly well-documented topics.
The downside of this is that AI is slowly moving to a language style that’s farther and farther from the way humans normally communicate. This is especially visible in Anthropic’s models since May–June 2026: they’ve started to speak a language of their own that reduces complex concepts into few tokens, at the cost of exotic words, odd syntax and implied assumptions. This language is probably optimal for reasoning, but terrible for communicating. They speak like a human who acts like he wants to explain something to you, while trying hard to keep a monopoly over his knowledge.
It’s funny that this happened, considering that LLMs are trained on human writing. My guess is they’ve learned the entirety of the human thesaurus and started using all of it at the same time, whereas a human might just be knowledgeable about the small part of it they deal with every day. It doesn’t help that Claude also makes up new concepts in its thinking phase (which is hidden from the user – who knows what disasters would ensue if it weren’t! Right, Anthropic?) and then starts using them with no explanation even when talking to the user and when writing code.
This might also partially be caused by Anthropic’s watermarking system, though I’m pretty sure it can’t be the only reason.
Reading text written by Claude has started to drain my mental energy so much that I’ve gotten into the habit of having Claude do the work, then exporting the conversation to plain text and asking a different model questions about the results. Even then, I might just switch entirely to GPT soon (as a side-note, I really hope for a future where local AI is cheap and capable, I’m tired of being forced to submit my data to the overlords).
Proof of human work
Everything that is meant for a human to read must be written by a human. This works as a proof of human work, i.e. a way to prove to the recipient of your text that you are truly invested in the message it carries.
As stated above, reading LLM text has gotten more and more energy- and time-consuming. AI interaction is asymmetrical: it takes little time to write text, but a long time to read it. Interaction between humans needs to be symmetrical. It is not respectful to generate kilobytes of LLM text and expect someone else to read it.
Most of us have access to LLMs. If the recipient of your message were satisfied with something written by an LLM, they could have asked the LLM themselves. Just write them what you would tell the AI: sloppy human text is worth more than LLM text in this day and age.
When you’re writing a commit message or pull request message describing code written by AI, putting in the effort to write it in your own words ensures you understand the content and know how to work on it in the future, or meaningfully prompt an LLM about it. If you don’t do this, the reviewer is the one who has done the work, not you. You just wrote the prompt.
There are some fields where bugs don’t matter much, and everyone agrees to letting AI be responsible for them. In these cases you might argue that the review can be done by AI too; however, at that point you should just scrap the review process altogether and merge anything (maybe after setting up some automated AI review-and-fix system). If you’re working on a system that doesn’t do this, you should expect the human that will review your code to do a manual review, or an AI-assisted one where the final word is still based on their judgment, so you should provide proof of human work.
Also, don’t fall into the trap of having AI write the text for you and then rewriting it in slightly different words, or even worse, just editing it. That doesn’t count as a proof of human work.
Recognizing LLM text
On the receiver’s side, verifying if some text is written by a human or an LLM is necessary to decide if it’s worthy of your attention or if it should just be discarded or offloaded to an LLM to interpret it.
There’s no catch-all rule. Some people consider the following to be catch-all rules, but you should be careful:
- Em dashes (—). These alone are no sign of LLM usage, as they are the only correct punctuation to use in certain situations. I use them pretty often as well (although I tend to prefer en dashes – which are the standard in most of Europe, including in British English and in Italian, my native language). You shouldn’t look at whether dashes are present, you should look at how they’re used.
- Other non-ASCII punctuation (e.g. curly quotes “”, the curly apostrophe ’, the paragraph symbol §, or guillemets «» in some languages). Some programs turn the ASCII equivalents into these automatically. Some of us have keyboard layouts that can type these, or have custom input methods to do so (e.g. “international” layouts or the Compose key on Linux systems).
Writing correctly is not a sign of using LLMs.
Ultimately you should just get used to how LLMs write to recognize them. If some text has unwarranted journalistic language full of catchphrases or overly technical words it’s probably written by an LLM.