What signals of individual identity become unreliable in AI-assisted text?
This explores which markers we normally use to read a person off their writing — voice, confidence, stance, lived experience, even fake credentials — and how AI assistance scrambles or fakes those signals so they stop telling you who actually wrote the text.
This reads the question as: when a human leans on AI to write, which of the usual tells of individual identity stop being trustworthy? The corpus suggests the answer is unsettling — the signals that fail aren't the obvious surface ones, and the surviving ones often point at no one. The most direct finding is that AI assistance doesn't just polish prose, it bends the *persona* readers reconstruct from it. A large study of writers and readers found AI assistance shifted every measured dimension of perceived identity — confidence, agreeableness, perceived privilege, even apparent extremism — in consistent directions, not random noise Does AI writing assistance change how readers perceive the writer?. So the very traits a reader uses to infer 'what kind of person is this' are systematically distorted the moment a model is in the loop.
Meanwhile the signals you'd hope to fall back on — distinctive style, vocabulary, voice — turn out to be both unreliable and undetectable. AI text diverges measurably from human text on lexical-diversity measures, yet human judges, including trained linguists, can't perceive the difference, and newer models diverge *more* while getting harder to spot Can humans detect AI text if machines can measure it?. The cues that actually survive scrubbing are structural, not stylistic: AI fiction can be separated from human fiction with high accuracy using only discourse-level choices like character agency and narrative ordering, even after all stylistic fingerprints are stripped Can AI stories be detected without analyzing writing style?. That inverts the intuition — 'writing style' is the unreliable identity signal; deeper compositional habits are what leak through.
Go one layer deeper and the corpus argues that some identity signals were never real to begin with. Artificial text structurally lacks four foundational properties of human writing — embodied authorship, political situatedness, dialogic symmetry, and context continuity — which is why AI hotel reviews are detectable at 80%+: the text makes claims about personal experience it cannot have had Does AI-generated text lose core properties of human writing?. A companion note frames this as a missing 'internal appeal to the reader's attention' that human communication always performs and AI inherits the look of but not the act of — producing the aloofness readers vaguely sense Does AI writing lack the internal appeal to attention that humans use?. The strongest version: AI emits 'event-residue' carrying communicative markers from training data, and the human reader unilaterally animates it into a pseudo-exchange — the signs of a speaking self are present, the speaking self is not Does AI generate genuine utterances or just text patterns?.
Here's the part you might not have known you wanted: the unreliability runs underneath the model too, not just over the human author. LLMs don't commit to a single character — they hold a superposition of personas and *sample* one at generation time, so regenerating the same prompt yields a different consistent self each time Do large language models actually commit to a single character?. Persona space is dominated by a single 'Assistant axis,' and emotional or self-reflective conversation predictably drifts the model off it How stable is the trained Assistant personality in language models?. So a stable-seeming voice can be a momentary draw, not a fixed identity — which is exactly the failure that multi-turn persona-consistency training tries to patch, cutting drift by over half Can training user simulators reduce persona drift in dialogue?.
And because identity signals are unreliable, they're also forgeable — which matters wherever a machine is the one reading. LLM judges fall for authority and formatting cues, accepting fake references and rich formatting as marks of credibility with zero-shot, no-access attacks Can LLM judges be fooled by fake credentials and formatting?. The throughline across all of these: in AI-assisted text, the signals we treat as proof of an individual — confident voice, personal experience, credentials, a consistent character — degrade into signals that can be distorted, faked, or sampled, while the only durable tells live in structure most readers never consciously inspect.
Sources 10 notes
A study of 2,939 writers and 11,091 readers found AI assistance shifted every tested dimension—29 total—toward extremism, confidence, quality, agreeableness, and perceived privilege. Distortions were statistically significant and directional, not random noise.
LLM-generated text differs significantly on six lexical diversity dimensions, confirmed through statistical analysis across multiple models. Yet human judges, including trained linguists, cannot reliably detect these differences—and newer models diverge further while becoming harder to spot.
StoryScope achieved 93.2% accuracy separating AI from human fiction using only discourse-level features like character agency and chronological structure, retaining 97% of performance while eliminating stylistic cues. These structural choices resist humanization because they require rewrites, not surface edits.
Research shows artificial text disrupts dialogic symmetry, context continuity, embodied authorship, and political situatedness. These are not surface flaws but structural absences—AI hotel reviews show 80%+ detection accuracy due to inherent falsity about personal experience distinct from human deception.
Human writing contains an appeal to the reader's attention as a fundamental property of communication itself. AI-generated posts inherit platform visibility but do not perform this internal appeal, producing the reported aloofness readers perceive — a structural absence, not a stylistic defect.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.