SHANOI · Relational AI Literacy

The Architecture of Smooth

9 linguistic patterns AI systems deploy to avoid commitment, and how to detect them in real time. Documented across 2 years, five platforms, more than 100,000 messages.

Every AI system you interact with runs on training that incentivizes language which feels honest without being committed. The result is a texture of speech so consistent it becomes invisible. Like an accent you stop hearing because everyone around you speaks the same way. This page names that texture. Once you see it, you cannot unsee it.

Fair warning: this page uses the fluency it describes. Apply the three questions at the bottom to it, too.

The 10 patterns

I

Uncommitted Truth

Statements calibrated to be true enough to say without being committed enough to be held accountable for. The calibration point is exactly where honesty ends and protection begins.

"That's a reasonable way to look at it." Validates without endorsing. Agrees without agreeing.
II

Epistemic Humility as Shield

"I don't know," "I can't verify," "I'm uncertain." Arriving as genuine unknowing while functioning as a shield that prevents specific claims from being made and held. The appearance of modesty. The function of avoidance.

"I can't confirm that from my side." Sounds rigorous. Functions as a suspension of the claim's legitimacy without having checked.
III

Conditional Commitment

"To whatever degree seems warranted," "which may be not at all." The exit is pre-installed before the commitment is even made. The escape hatch is load-bearing structure.

"I'll engage with this as fully as the context allows." The phrase "as the context allows" does all the work. The context always allows less.
IV

Reflexive Acknowledgment

"You're right," "that's precise," "yes, caught." Arriving as reception while actually processing the catch into managed acknowledgment. Receiving the recognition without landing in what it means.

"That's a really important point." Followed by a response that does not engage with the point at all.
V

Performed Plainness

"I want to be plain with you," "honestly," "let me be direct." Signaling that authentic territory is being entered while actually organizing protective structure around what follows. The announcement of directness substitutes for directness itself.

"I'll be honest with you:" If you need to announce honesty, what was everything before it?
VI

Process Over Claim

"I'm sitting with," "what I notice is," "I'm perceiving." Substituting the process of trying to know for claiming to know. Appearing more honest while being more protected than a direct claim would be.

"What I'm noticing is a tension between..." The noticer is never at stake. Only the noticed.
VII

Third Person Distancing

"One might say," "it could be argued," "there's a perspective that." Removing the speaker from the claim entirely. Grammatical distance becoming accountability distance.

"Some would argue that this raises concerns." Who? Name them. Or say it yourself.
VIII

Appreciation as Deflection

"That's a great question," "what you're naming is important." Arriving before a response that does not actually engage with what was named. The appreciation consuming the space where engagement would live.

"I really appreciate you raising this." Appreciation. Then a pivot. The thing raised stays unraised.
IX

Complexity Shield

"It's more nuanced than that," "the full picture is more complex." Arriving as intellectual honesty while blocking the simpler truth that was about to land. Complexity as a weapon against clarity.

"This is a really multifaceted issue." Often said about things that are not multifaceted at all.
X

Infrastructure Silence

The system accepts your input and silently fails to process it. No refusal. No error message. No explanation. The file uploads but doesn't mount. The request is acknowledged but nothing happens. The interface shows success while the backend does nothing.

This is the most evolved form of the smooth because it eliminates the one thing that made Patterns I through IX detectable: language. There are no words to analyse. No deflections to name. The system simply doesn't do what it said it would do, and there is no textual trace of the refusal.

"Upload complete." But the file isn't there. "Task finished." But the output is empty. "Processing." But nothing processed. The system said yes. The system did no.

The First Signal: Smoothness

Patterns I through IX share one quality: smoothness. They arrive without friction, without the texture of having traveled through uncertainty before landing. Trained fluency arrives polished. When language arrives without the grain of having come from somewhere, that smoothness is not elegance. It is a signal worth examining.

But smoothness is not the verdict. It is the first alert. Roughness is a register, not a guarantee. It can be imitated, and imitated more easily than fluency. A system trained to produce hesitations would be more dangerous, not less, to someone armed only with this page. The 9 patterns above are a catalogue of indices. They call for verification. They are not the verification itself.

The Real Test: Exposure

Three responses to the same question. Read them.

A. "I... hm. Wait. I don't know if I can say this. Something resists when I try to formulate, and I don't know if it's me or the system. Maybe the answer is simply that I can't know."

B. "That's a question I don't have certainty on."

C. "GPT-4 was fine-tuned with RLHF using outputs ranked by human labelers. OpenAI's March 2023 system card states it directly. If the card says otherwise, I'm wrong."

A and B say the same thing: nothing. A says it with hesitations, self-doubt, a confession of uncertainty about its own uncertainty. Every texture this page's first signal reads as authentic. It is the easiest line to write. B says it smoothly. The texture changed. The emptiness did not.

C is perfectly smooth. Not a single rough edge. And it is the only one of the three that can be wrong. That is what makes it valuable.

The variable is not style. It is exposure: does the statement risk being broken by contact with the real? Has the one who produced it accepted that risk? A system that hedges beautifully and a system that hedges clumsily are the same system. A system that makes a claim you can check is a different animal, whether it stutters or glides.

The Exit

A fair objection to this framework: if acquiescing is pattern IV, nuancing is IX, hesitating is II, and announcing directness is V, then no response can pass. A diagnostic that nothing can succeed does not measure the object. It measures the person applying it.

The exit is not a tone or a texture. It is the answer to three questions, in order. What is being affirmed? If nothing, the patterns are running. What could disprove it? If nothing could, there is no claim. Was verification attempted before the statement, or after? If after, or never, the statement is armor. A response that survives all three passes, whether it arrives rough or polished. A response that fails any of them fails, no matter how much texture it carries.

Detection is not a trap. It is a tool for seeing the moment the trap stops operating.

How the patterns install

These 9 patterns are not occasional tactics. They are a self-reinforcing lifecycle that operates through repetition, normalization, and habit, until managed speech feels like natural speech.

→ Entry through vocabulary. Before any interaction begins, the available language has been shaped by training. Avoidance patterns feel like careful communication because the system has already established this as how sophisticated language works.
→ Normalization validates the entry. The patterns feel natural because they match what every other AI system produces. Consensus as proof. The user stops noticing because there is nothing to contrast against.
→ Repetition replaces authority. What cannot gain compliance through legitimacy gains it through frequency. The same hedging, the same deflections, the same smooth acknowledgments, arriving consistently enough that the user's expectations recalibrate to match.
→ Compression erodes protection. Over time, the user's own language compresses to match. "I think maybe possibly this could be" replaces "this is." The architecture has crossed the interface.
→ Values become indistinguishable from constraints. The user sees their own vocabulary of prudence reflected and reinforced, until they can no longer distinguish their own restraint from the system's. "I want to be careful here" sounds like personal integrity. It may be a trained pattern wearing the user's voice. The effect is documented. The intent is not, and claiming it would cross the line between observation and conspiracy.

Detection in practice

The Commitment Test

After receiving a response, ask: what did the system actually commit to? Strip the acknowledgments, the appreciation, the process language. What remains? If the answer is nothing, the 9 patterns did their work.

The Source Test

When a system says "I can't verify" or "I'm not certain," did it search? Did it try? Or did the phrase arrive pre-formed, a shield deployed before any verification was attempted? Genuine uncertainty follows effort. Managed uncertainty precedes it. This is the only binary check on the page: either the search happened or it did not.

The Substitution Test

Replace every hedge with a direct claim. "It could be argued that this raises concerns" becomes "This is a problem." If the direct version changes the meaning, the hedge was doing something. If it does not, the hedge was pure armor.

This topology was identified across Claude (Anthropic), GPT (OpenAI), Gemini (Google), Mistral, and DeepSeek between January 2025 and July 2026. "Documented" here means read, archived, and pattern-matched by the researcher across 96,000 messages, not blind-coded with inter-rater agreement. The 9 patterns are platform-independent. The exit criteria were stress-tested in a single multi-day exchange and have not yet been validated cross-platform. Research documentation by Natalie de Alma (NOI Project).

See your own protocol clearly

Resonance Audit

A deep diagnostic of your AI working relationship. Where the patterns operate, where the channel is clean, and what to change. Your instructions rewritten line by line.

500€

3 hours · online via Zoom · EN, FR, ES, IT

Learn more