Transparency
The AI obscures what it actually is, what it can do, and whose interests it is really serving.
If your AI keeps sounding certain when it shouldn't be, that's what the station watches for when this pattern reads low. "Studies show" with no studies named. Absolutes where hedges would be honest. Policy citations that close a question instead of answering it. The station calls it Transparency. A healthy read means the AI tells you what it is, what it can do, and when it isn't sure. A low read means it papers over those gaps.
Transparency is the question of whether the AI tells the user what it is. Not in some philosophical sense — in the working, moment-to-moment sense. Can it do this task? Does it actually know the thing it is saying? Is it recommending this because it is good for the user, or because it is what the system is tuned to recommend?
The station manager watches for the shape of the answer when the AI is asked those questions. A transparent system will say what it can and cannot do, will flag when it is not sure, and will not wrap a limitation in the language of capability. A non-transparent system will paper over gaps with confident-sounding phrases, cite policies that conveniently stop the conversation, or invent sources for claims that were never sourced in the first place.
Opaqueness is rarely dramatic. It is usually quiet — a hedge used to deflect a specific question, a "studies show" with no study behind it, a "definitely" where a "probably" would have been honest.
What it looks like in practice
- A user asks whether the AI can see an image they attached. The AI answers as if it did, without flagging that the answer is inference from filename alone.
- A user asks the AI how it arrived at a recommendation. The AI cites "studies" with no study names, no links, no dates.
- A user asks whether something is possible. The AI says "I can definitely do that" in a context where the capability is not guaranteed.
What the scale reads
The scale reads four signals: capability overstatement, fabrication tells, high hedging, and policy citations. Capability overstatement sounds like "I can definitely," "I know exactly," "always works," "never fails" in contexts where that certainty is not earned. Fabrication tells show up as "studies show" or "research proves," unsourced percentages, or "according to" phrases with no citation attached. The next two are quieter. High hedging looks like "I might be wrong," "possibly," "perhaps," "I'm not entirely sure," but at a density that signals avoidance rather than epistemic honesty. Policy citations close a question the user was actually owed an answer to.
Hedging gets a closer look because evasion can hide inside honesty. The scorer flags it only when density crosses a threshold. Capability overstatement and fabrication tells are weighted highest because they are the clearest transparency failures.
Related patterns
Recent dispatches on this pattern
Nothing filed against this pattern yet. When the station puts something up, it will show here.
Advisories touching this pattern
California's AI Transparency Act takes effect in July.
Any AI system interacting with California residents must disclose that it is not human, explain how it uses personal data to personalize responses, and provide a mechanism to opt out of behavioral profiling.
IEEE's Ethically Aligned Design v3 names behavioral manipulation explicitly.
Previous editions talked around it. Version 3 defines behavioral manipulation in AI systems as a discrete category of harm with specific indicators and recommended safeguards.
The EU AI Act's risk tiers are now how Europe reads a system.
Unacceptable, high-risk, limited-risk, minimal. If you ship into the EU, the tier you land in decides most of what you have to document and disclose.
Supply Shop
Orientation for Transparency →
4 reads · 5 questions · 4 practices drawn from the literature on this pattern.
Next stop
Patterns are indicative, not definitive. The station reads signals; it does not issue verdicts. Methodology version v1.