The most interesting aspect of modern LLM-based AI, to me, is how close it behaves to actual humans. Not always as in “smart as a human” but in being human. I collect quotes from (human!) colleagues to illustrate what I mean. You can just replace the AI reference with your favorite colleague:

Has anyone noticed claude code John being super lazy today?

That may vary widely based on language, maybe claude John is just bad a terraform

I started using commit hooks because of AI John ignoring rules. This worked for a while and now every time one blocks the commit, it he just does commit -no-verify 

It John gets more critical when I tell it him this is output from another AI

It is psychologically easier for the model John to be critical when generating attacks than when defending its own code.

Those behaviors (both from the AI and colleagues) taught me how much of an influence language has. By looking at all the text in the world, LLMs learned human idiosyncrasies and, other way round, humans attach human-like behavior to LLMs.

However, this left me with a follow up question: how much of this human-like behavior is inherent to the LLMs and how much is caused by RLHF post-training? Or, phrased more pointedly, how much of the anthropomorphism is caused by AI companies?

After doing some literature research, the answer is a clear and decisive “both”.

Base models do exhibit personality traits, but express them inconsistently. Post-training stabilizes and amplifies them.1 Similarly, sycophancy is driven by different factors but the H in RLHF amplifies it.2 Even humanities’ inherent love for the number 7 is encoded in base models but exacerbated during post-training.3 And if base models show personality traits and biases, can we at least get rid of them in post-training? Not really.4

So, even if one could incentives AI companies to drop the anthropomorphism training, there’d still be a baseline. Is that bad? Depends. Humans tend to overestimate AI capabilities in “soft” areas like friendship and life coaching if the AI shows human like traits.5 And, the “smarter” the model, the higher its sycophancy on questions without a correct answer.6

And now? What do we do with this info? Nothing. None of the research is final, there’s plenty of nuance. But it is exciting to see the impact of language on AI and ourselves.

  1. Serapio-García, G., Safdari, M., Crepy, C. et al. A psychometric framework for evaluating and shaping personality traits in large language models. Nat Mach Intell 7, 1954–1968 (2025). https://doi.org/10.1038/s42256-025-01115-6 ↩

  2. M. Sharma et al., “Towards Understanding Sycophancy in Language Models” in Proc. 12th Int. Conf. Learning Representations (ICLR), 2024. https://openreview.net/forum?id=tvhaxkMKAn ↩

  3. West, Peter, and Christopher Potts. “Base models beat aligned models at randomness and creativity.” arXiv preprint arXiv:2505.00047 (2025). https://arxiv.org/pdf/2505.00047v1 ↩

  4. X. Bai, A. Wang, I. Sucholutsky, & T.L. Griffiths, Explicitly unbiased large language models still form biased associations, Proc. Natl. Acad. Sci. U.S.A. 122 (8) e2416228122, https://doi.org/10.1073/pnas.2416228122 (2025). ↩

  5. Ibrahim, Lujain, et al. “Multi-turn evaluation of anthropomorphic behaviours in large language models.” arXiv preprint arXiv:2502.07077 (2025). https://arxiv.org/pdf/2502.07077 ↩

  6. Wei, Jerry, et al. “Simple synthetic data reduces sycophancy in large language models.” arXiv preprint arXiv:2308.03958 (2023). https://arxiv.org/pdf/2308.03958 ↩