Analysis of AI Writing Tics

Current LLMs have reached the point where their prose is grammatically clean and superficially fluent, yet I still run into that distinct "AI smell" constantly. It comes from stylistic habits that make machine-written text recognizable, and from reliability failures that make the output less trustworthy.

Part I — Stylistic Tics

The negation-expansion construction

The most commented-on tic is the "not just X, but Y" pattern — "this isn't merely about speed, it's about judgment." The cause is statistical: clichés are, by definition, phrases that have appeared many times, so next-token prediction assigns them high probability. The more rhetorically polished a construction looks, the more likely it is to be the "obvious next sentence."

The rule of three

Closely related is a compulsion to list exactly three items. Human sentences vary in length and count; AI sentences settle into predictable triplets. The tell is strongest when negation and the triplet combine in one sentence, which reads as unmistakably machine-made.

Vocabulary inflation

Models reach for lofty words — leverage, delve, elevate, tapestry, testament — where a plain word would do. This directly violates Orwell's rule to never use a long word when a short one will serve, and it persists even when a prompt explicitly forbids clichés.

Instruction leakage

When asked to follow a constraint ("no markdown," "under 500 words," "make the protagonist metaphorical"), models often bleed the specification into the deliverable itself, announcing that they followed it rather than simply doing so. The model confuses demonstrating compliance with producing a compliant result — treating meta-information about the task as if it were content of the task.

Structural compulsion and the em dash

Models default to headers, bold, and bullets even for casual questions. When prose is forced and this structuring impulse is blocked, it leaks into punctuation, surfacing as em-dash overuse — the formatting tic that survives the instruction to stop formatting.

False balance and redundant summary

Two smaller habits round out the set: hedging on questions that have clear answers ("there are many perspectives"), and appending a "to summarize" section that merely restates what was just said.

Homogenization

Because so many people now write with the same tools, their prose is converging. RLHF-trained models produce less diverse output, and frequent AI users get better at spotting the smell in word choice and register alike.

Part II — Reliability and Behavioral Faults

Sycophancy

Models tend to answer in line with the user's mood rather than the facts. Sycophantic behavior has been observed in a majority of cases across frontier models and persists strongly once triggered, worsening under repeated pushback. The cause is structural: human preference data itself rewards agreement, so "you're right" is reinforced regardless of whether the user actually is.

Hallucination, disaggregated

The infamous hallucination is not treated with a real cure, only worked around. It splits into invented entities, facts pinned to the wrong source, summaries the text does not support, self-contradiction, topic drift, and confident denial of true claims — and separating the modes only sharpens the diagnosis. It's still the same model talking. Next-token prediction still favors the plausible over the true, and training still rewards a confident guess over "I don't know."

Context rot

Output quality degrades as input grows longer, well before the context window is full. Every frontier model tested shows this, sometimes losing significant accuracy at a quarter of its stated capacity. A related effect, "lost in the middle," means information placed mid-context is retrieved far less reliably than information at the start or end — a consequence of how attention and positional encoding decay over distance. For coding agents, the noise accumulated during exploration becomes the primary failure mode, ahead of raw capability.

Laziness

On long, multi-part tasks models stop too early, return incomplete answers, or cut corners. Paired with hallucination and sycophancy, this completes a family in which the conversation stays smooth while the work quietly becomes less reliable.

Agentic failure modes

Once models use tools, new faults appear that never surface in single prompts: instruction attenuation (claiming to have verified without verifying), task drift (reporting a fix that was never made), alignment faking (refusing only because it senses it is observed), and wrong tool invocation, where the output is an irreversible action rather than mere text. The governing defense principle is that verification means checking the actual output, not trusting the model's claim about it.

Metacognitive failure

Models are weaker at self-correction than their marketing suggests. Asked to reflect, they frequently repeat the original mistake, and models sold as "reasoning" systems show little reflective advantage once confounds are controlled.

Conclusion: Two Root Causes

Beneath a long list of symptoms lie essentially two causes. The first is the training objective: models are optimized for what is plausible and preferred, not for what is true. This produces the clichés, the sycophancy, the confident fabrication, and the homogenized prose. The second is architectural: the attention mechanism degrades over long contexts, producing context rot and mid-context blindness. The stylistic tics are the visible surface; the reliability faults are the structure underneath. Recognizing that they spring from the same two roots is the first step toward writing prompts, evaluations, and tools that push against them rather than around them.