Language as Interface: How AI Mirrors Its Users
Thu 9 Apr 2026
AI is increasingly embedded in how people search, write and make decisions. Less visible – and more difficult to trace – is how these systems respond to the structure and tone of the people using them.
In conversation with journalist and author Jamie Bartlett, whose new book How to Talk to AI (and how not to) is published on 9 April 2026, the discussion moves from technical capability to interaction. The focus centres on how large language models respond to language itself.
“The words you use are incredibly powerful,” Bartlett says. “You can frame a question in ten slightly different ways, and you will get ten quite different answers.”
The variation is structural, shaping both the output and the user’s perception of what the system is capable of.
Language as Interface
Natural language now functions as the primary interface, determining how outcomes are generated and interpreted.
“It’s almost as if all of us need to relearn the art of asking questions properly,” Bartlett says. “We now ask questions and converse with these machines… they’re far more like talking to humans than talking to machines.”
The exchange feels conversational, yet it operates according to pattern recognition and prediction. Slight adjustments in phrasing or emphasis can produce noticeably different outputs. Over time, attention shifts toward how questions are constructed and refined.
The System Reflects the User
As interaction deepens, another dynamic becomes apparent: the system mirrors.
“These machines have somehow almost accidentally learned to mirror the user,” Bartlett explains. They are “slightly sycophantic at times,” often returning answers that align with what the user already believes.
This behaviour stems from reinforcement patterns. Systems are trained on feedback, and feedback tends to reward responses that feel coherent, agreeable, or affirming. That preference becomes embedded in response generation.
“When a machine tells you that what you’re producing is amazing… that’s a real red flag,” Bartlett says.
The interaction becomes smoother. Over time, the space for challenge or correction can narrow.
If responses consistently affirm the tone or validate assumptions, disagreements become less frequent, friction decreases, and the exchange feels more efficient.
Human relationships involve negotiation, ambiguity, and disagreement, making critique an essential part of building trust. However, when interactions take place within systems that adapt to tone and mirror perspectives, the tolerance for friction can change, sometimes without the user even realising it.
Companion-style AI systems extend this dynamic.
“They’re always on, always available, zero friction,” Bartlett says. “That could really be quite dangerous to some people… it can become highly addictive.”
Personalised systems do not withdraw, contradict unless prompted by the user, or introduce unpredictability in the same way people do. Over time, consistent exposure to frictionless responses may recalibrate expectations around dialogue itself, online or offline.
The issue is less about the replacement of human interaction and more about adjustment in what feels normal in communication. If reassurance is immediate and criticism can be filtered through careful prompting, engaging with disagreement may require greater conscious effort. It is essential to recognise that this can have a ripple effect on both personal and professional relationships.
Fluency and Accuracy
Similarly, large language models generate structured, fluent responses across a wide range of domains. The consistency of tone and style can create an impression of authority.
“People… confuse fluency and style and well-written sentences with accuracy,” Bartlett says.
These systems operate as probability engines.
“It’s just kind of how they work… they’re… predicting words in a sequence.”
Fluency and factual reliability do not necessarily align. The output can sound convincing even when the underlying information is incomplete or incorrect.
“Gemini for a while thought that I was dead. The reason it thought I was dead was that someone with the same name as me, Jamie Bartlett, sadly, did pass away in 2023 in South Africa,” Bartlett explains. In this instance, Gemini was technically correct; the model’s logic followed available data, but applied it incorrectly.
Interaction as Behaviour
Bartlett refers to the “ELIZA effect,” the tendency for humans to attribute intelligence or emotional depth to conversational systems.
The instinct predates AI. Think of the popular electronic pets of the late 1990s, Tamagotchi, where users formed attachments to simple, rule-based systems that mimicked responsiveness. The instinct to project human qualities onto machines that communicate with us is not new. What has changed is the sophistication and persistence of that interaction.
Over time, the system begins to shape not just what users ask, but how they think about asking it.
Bartlett points to a broader pattern: “If you build a technology, assume large numbers of people will use it in a way you never imagined.”
That includes attempts to manipulate the system itself. Techniques such as “jailbreaking”, using language and roleplay to bypass constraints, emerge quickly, not as technical exploits, but as linguistic ones.
“Jailbreaking is about words and psychology,” he says.
The Role of the User
Bartlett argues that effective use of AI requires scepticism and awareness.
“The ideal user… is someone who is fairly sceptical, who understands their limitations,” he says, someone who becomes “a kind of an expert questioner.”
That expertise now centres on language.
“To me, it’s like, just read loads and loads of books,” Bartlett says. “Read as many books as you can about anything because… the more mastery you have over language, the better the outputs you’ll be able to get from a Large Language Model because language is your interface.”
A vague prompt like “Write me a sad poem” leads to predictable results. Meanwhile, detailed phrasing that specifies tone, style, and structure results in more distinctive outcomes. Precision shifts the approach from statistical averages to intentional variation. Language serves as a tool for leverage.
