Why "As a Language Model" Keeps Showing Up in 2026
AI tone can be configured more easily than it looks. Review what a rewrite means before you trust how it sounds.
September 27, 2026
If an AI assistant keeps opening with "As a language model," do not read that sentence as a deep statement about the system. A paper that hit Hacker News this week argues that the wording can be pushed up or down by something much more ordinary: the chat template wrapped around the model.[1][2]
That sounds like a niche implementation detail. It is not, if you use AI to clean up emails, notes, drafts, or support replies. The opening tone of generated text often gets mistaken for judgment. A formal disclaimer can make a response feel cautious. A first-person phrase can make it feel personal. Neither is a reliable sign that the output understood your situation.
What the new paper found
The paper, "As a Language Model...", tested eight open-source instruction models up to 9B parameters. Its author reports that adding a chat template increased disclaimer-style language and reduced experiential phrases such as "I feel." Removing the template moved the language in the other direction.[1]
The study also tested an activation direction in three models. Adding that direction made models disclaim more. Removing it reduced the disclaimer voice. The paper's central warning is simple: a model's self-description is partly shaped by how it is deployed, so researchers should not treat those lines as straightforward evidence about what the model "knows" about itself.[1]
That is a narrower claim than "the model is faking everything." The work does not settle whether models understand their own behavior. It shows that the wording around that question is easier to steer than people may assume.
Why the wording matters in everyday work
People do not only ask AI systems for facts. They use them to tighten an awkward paragraph, summarize a meeting, reply to a customer, or turn rough notes into a checklist. In those moments, voice can change the decision a reader makes.
A disclaimer-heavy rewrite can sound defensive. A smooth, confident rewrite can sound more certain than the source material deserves. Both can hide the actual question: did the assistant preserve the facts, the request, and the person speaking?
The practical test is not whether the output sounds humble, warm, or impressively careful. Check whether it did the job you asked for. If you asked for a shorter email, compare the before and after. Did it keep the date, the number, the ask, and the boundary? If you asked for notes, did it separate decisions from guesses? That is more useful than trying to read the model's personality from a stock phrase.
Do not use tone as a quality signal
There is a habit worth dropping: accepting a rewrite because it sounds polished. Polished text can still delete the one sentence that matters. It can make a soft suggestion sound like a commitment. It can turn "can I send an example?" into "send it over," which flips who is asking whom for something.
A short review pass catches most of this:
- Read the first and last sentence side by side with your source.
- Check every date, price, name, link, and next step.
- Look for a change in certainty: may, should, must, and will are not interchangeable.
- Remove any generic opener that does not help the reader.
- Keep your own decision in the final version, especially before sending it.
This is not a call to make every draft stiff. It is a call to stop treating style as evidence. The paper gives a technical reason for that instinct: a deployment-level setting can change the self-referential voice without changing the underlying task.[1]
Where voice dictation helps
Voice input has a useful role here because it keeps the first draft close to the person responsible for it. Speaking a rough update usually captures the qualifiers people forget when they type fast: "I think," "after we confirm," "not for this customer," or "please do not promise that yet."
DictaFlow is built for that first step. Hold a hotkey, say the thought, and release to put it into the app in front of you. Its cleanup should stay narrow: fix transcription friction, punctuation, fillers, and obvious repeats without quietly turning your point of view into generic assistant prose. That matters most in support, sales, clinical-adjacent documentation, and any situation where a soft change in tone changes the meaning.
The best workflow is boring in a good way. Speak the real draft. Use AI for limited cleanup. Read the final ask once before it leaves your screen. DictaFlow's comparison guide explains the system-wide dictation workflow; its Citrix and remote-desktop guide covers the harder environments where ordinary text insertion can fail.
A better way to read AI output
The Hacker News discussion had 70 points and 69 comments when checked, which makes sense. The paper pokes at a familiar oddity in AI writing, then shows that part of the oddity may come from the wrapper rather than a mysterious inner voice.[2]
That does not mean templates are bad. They are part of what makes a model usable. It means the wrapper is part of the behavior you are seeing. When an AI draft sounds unusually cautious, personal, or robotic, judge the content before you judge the tone.
The next time "As a language model" shows up in a draft, delete it if it adds nothing. Then check the sentence that follows. That is usually where the real work starts.