Researchers have spent a good deal of effort teaching language models to sound like specific people online. The models have learned to count words like a human. They have not learned to feel like one. A new study compared the replies of nine open language models with real messages from X, Bluesky and Reddit, and found that even carefully calibrated models can still be told apart from people by ordinary machine-learning classifiers, with accuracy of 70 to 80 percent.
The setup: nine models, three platforms, five ways to cheat
The authors tested models from the Apertus, DeepSeek, Gemma, Llama, Mistral and Qwen families. They did not simply ask the models to chat. They tried five calibration strategies meant to push each model toward the style and habits of particular users. These included fine-tuning, giving the model examples of the user's style, and retrieving relevant context to feed into the prompt.
To score the results, the team built a framework called CHAT, short for Conversational Human Authenticity Test. It does not just give a thumbs up or down on whether a reply looks human. It combines several measures:
- Detectability: how easily a classifier can separate model replies from human ones.
- Meaning: how close the generated reply is, semantically, to what the human actually wrote.
- Texture: how the two differ in style, vocabulary, emotional coloring and topic.
The point of the last group is diagnostic. Knowing that a model fails the test is mildly interesting. Knowing where it fails is something you can work with.
Where the disguise slips: feelings
The most persistent giveaway was emotional tone. Calibration did help with the mechanical parts of a message, such as sentence length and word count. Models got better at looking like a post. They did not get much better at sounding like a person having a reaction to something.
The analysis also found that sentiment and toxicity kept playing a stable role in separating the two groups. Both are tied to how people express attitudes and emotional responses, which is more or less the whole business of arguing on the internet.
Put simply, the models learned the outer shape of an online conversation faster than its emotional content. They can produce a plausible reply of the right size. The mood is where the seams show.
Bigger is not better, and politeness is a liability
Two findings cut against common assumptions.
Size did not rescue anyone. Llama 3.1 with 70 billion parameters showed no noticeable advantage over much smaller models. Scaling up is the industry's favorite remedy, and here it did not pay off.
Instruction tuning made things worse. Models that had not been trained to follow instructions generally imitated human speech better than models specifically tuned to carry out requests. That is slightly awkward for the usual product pitch, since the helpful, obedient assistant is exactly the version that sounds least like a person on a social feed. The source does not explain why, though the result at least fits the stereotype of the overly agreeable chatbot.
The platform mattered too. Models were most convincing on X, less so on Bluesky, and had the hardest time on Reddit, whose conversational norms are more varied. A model that can pass as a poster in one place may be quickly caught in another.
The trade-off nobody gets to escape
The main limitation emerged when the researchers set naturalness against accuracy of meaning. The two pulled in opposite directions.
Methods that made replies harder for classifiers to flag also made them less semantically similar to the real human messages. And the replies closest in meaning to the originals were more likely to be recognized as artificial. This held across different calibration methods and different ways of selecting among generated replies.
Loosening the generation parameters had the same pattern. It made AI text harder to detect, but it further weakened the match in meaning. In other words, one can have a reply that sounds more human or a reply that says what the human said, but so far not both at once.
There is a dry irony here. The easiest way to look less like a machine is to say something less like what the person would have said. That is also a fair description of a certain kind of human, but it is not what researchers are looking for.
Why social scientists should care
The authors point to a concrete audience. Social science increasingly uses language models to stand in for people. Uses include simulating survey respondents, predicting wellbeing, studying political behavior and modeling how the public reacts to events.
Those applications depend on the simulated speaker being a decent proxy for the real one. If the emotional layer is consistently off, and if closer meaning comes at the cost of naturalness (or the reverse), then a synthetic respondent may be wrong in systematic ways rather than random ones. Sentiment and toxicity are precisely the dimensions that matter when studying public reaction.
What to take from it
The source is a summary, so it does not say how the classifiers were built, who the authors are, or how large the datasets were. Even so, the practical reading is fairly clear.
- Do not trust fluency. A reply that matches length and vocabulary can still carry the wrong emotional tone.
- Measure more than one thing. A single detectability score hides the trade-off between sounding human and staying faithful to the original meaning. CHAT is an attempt to keep both in view.
- Test per platform. Results on X say little about Reddit.
- Do not assume a bigger or more obedient model is the safer choice. In this study, neither was.
The closing observation is a modest one. After all the tuning, examples and retrieval, the machines can imitate how people type far better than how people feel. For now, the tell is not the grammar. It is the mood.