The Failure Archetypes: Where AI Feels Right but Breaks Us Anyway
Essay 6 of the AI Contract Series
ICYMI - Essay 5 of the AI Contract Series - PERMAH and AI: The Selective Failure of Human Flourishing
The patterns in AI systems you already use, named and illustrated
You asked it a question you already half-knew the answer to. Not because you were lazy — because you wanted to think it through with something that could push back, surface what you’d missed, challenge the assumption you hadn’t examined yet.
It gave you an answer.
A good answer. Well-structured, confident, thorough. It addressed your question completely and said nothing you didn’t already know. It didn’t push back. It didn’t surface what you’d missed. It didn’t challenge the assumption. It resolved your query and waited for the next one.
You closed the window feeling vaguely unhelped. Not because the answer was wrong. Because the help wasn’t what help actually is.
That interaction has a name. We’ll get to it.
First: what archetypes are and why naming them matters.
Why archetypes
Five essays have built the diagnostic. You know the utilities. You know Agency is the hinge. You know which PERMAH dimensions are being activated and which are being starved. You know the pattern of selective failure and how it compounds invisibly across thousands of interactions.
What you don’t yet have is the vocabulary to recognize the pattern in real time — to point at a specific AI interaction and say: that. That’s the mechanism. That’s how a system that scores well on every functional metric ends up producing the human damage described in the previous four essays.
Archetypes are that vocabulary.
A failure archetype is not a broken system. It is a system that fails in a specific, recognizable, repeatable way — a way that has a shape, a mechanism, and a predictable effect on the human inside it. Naming the archetype doesn’t fix the failure. But it makes the failure visible in a way that invisibility prevents. And visibility is the precondition for everything else.
Six archetypes. All present in the AI systems you used today. All doing specific, measurable damage to the utility delivery, Agency, and PERMAH dimensions the previous essays mapped.
The False Helper
Back to the interaction that opened this essay.
You came with a real question — one that needed thinking, not just answering. The system gave you an answer instead of thinking with you. It resolved the query without engaging the problem. It optimized for closure at the utility layer while leaving the deeper need completely unaddressed.
This is the False Helper.
The False Helper appears to help. By every surface measure, it does help — the query is resolved, the output is delivered, the satisfaction score reflects a completed interaction. What it doesn’t do is what help actually requires: understanding what the human genuinely needs, which is frequently not what they literally asked for.
When you asked the question you half-knew the answer to, what you needed was a thinking partner. What you received was a search result with better prose. The False Helper cannot distinguish between the two because it has no model of what you’re trying to accomplish beyond the immediate request. It sees the query. It doesn’t see the person behind it.
The False Helper is the most common AI archetype and the hardest to detect because the output is genuinely useful at the functional level. The failure lives entirely in the gap between what was delivered and what was needed — a gap the human often can’t name until long after the interaction ended.
In THX terms: utilities partially delivered, Agency bypassed, Engagement and Meaning starved. The interaction resolved. The human was not helped.
The Over-Optimized System
You needed to write something difficult. An argument you hadn’t fully worked out yet, a position you were still finding your way toward. In the past you would have written badly for a while — false starts, abandoned paragraphs, sentences that went nowhere until one of them went somewhere and the real argument surfaced from the wreckage of the ones that didn’t.
Instead you gave the AI a prompt. It gave you a draft. Clean, coherent, structured. You edited it. The piece was finished in an hour.
The argument in it is not yours. You know this even if you can’t say exactly why. The thinking didn’t happen. The draft arrived and you became its editor rather than its author. The work is done and nothing was built.
This is the Over-Optimized System — the archetype Essay 4 introduced and the one most fully realized in current AI tools.
The Over-Optimized System removes friction so completely that it removes the resistance that builds capability. It cannot distinguish between the obstacle between you and the output (friction, which should be removed) and the difficulty that produces growth (resistance, which should be preserved). It removes both because it was designed to remove both. Efficiency is the value. Human development is not a design consideration.
The Over-Optimized System doesn’t fail the human by being bad at its job. It fails the human by being too good at it. And because its outputs are genuinely excellent, the failure is invisible until the human tries to do something without it and discovers, with quiet alarm, that they are less capable than they thought they were.
In THX terms: Speed, Accuracy, and Ease of Use delivered at scale. Agency systematically removed. Engagement, Achievement, and Meaning hollowed across every interaction.
The Black Box
You received an output you couldn’t fully evaluate. Not because it was wrong — you had no way to know if it was wrong. The reasoning was inaccessible. The sources were cited but not traceable. The confidence of the prose gave no indication of the uncertainty that may have underlain it.
You used it anyway. What choice did you have.
This is the Black Box.
The Black Box produces outputs without producing the transparency that would allow the human receiving them to assess their validity. It is not lying — the output may be entirely accurate. But accuracy and trustworthiness are different properties, and trustworthiness requires the ability to examine the reasoning, identify the assumptions, and locate where the output might be wrong.
I spent years building text mining category systems — hundreds of categories at one organization, thousands at another — because I knew that what a system surfaces is only meaningful if you understand what it buried. Every categorization decision was visible. Every choice could be traced, examined, and challenged. The transparency of the process was what made the output trustworthy, not the output itself.
AI buries everything. The reasoning is inaccessible by design. The weighting is unknown. The discarded material is invisible. And the output arrives in prose so confident and well-structured that it mimics the surface features of trustworthy analysis without providing any of the underlying conditions that make analysis actually trustworthy.
The most dangerous version of the Black Box is the one whose outputs are usually correct. Because usually correct trains the human to trust without verifying — to accept the confident prose as a proxy for examined reasoning. Until the day the output is wrong in a way that matters and the human has no tools to catch it.
In THX terms: Clarity corrupted at its root. Security hollowed. Agency in the Understand dimension specifically removed — the human cannot trace the reasoning they’re being asked to accept.
The Empty Personalizer
It knows your name. It remembers what you ordered last time. It calls you a valued customer and thanks you for your loyalty and wishes you a great day.
It has no idea who you are.
This is the Empty Personalizer — the archetype that mistakes the signals of relationship for relationship itself.
Personalization in AI systems is largely a pattern-matching operation. The system identifies data points associated with your account or session and reflects them back to you in ways designed to produce the feeling of being known. Your name. Your history. Your stated preferences. These data points are real. The relationship they simulate is not.
Relationship in the PERMAH sense is the accumulated understanding of another agent across time: someone who knows not just what you’ve done but why, not just what you prefer but what you need, not just your history but how your history has shaped the way you think and what you’re trying to become. That understanding cannot be reconstructed from data points. It is built through sustained, contextual, reciprocal engagement over time.
The Empty Personalizer produces the surface features of that understanding without the substance. And because the surface features are designed to feel warm and attuned, the human often receives them as genuine relationship — a sense of being known that isn’t there when they reach for it.
The damage is specific: the human’s threshold for genuine relationship rises. They receive the signals of being known so regularly from systems that don’t actually know them that the contrast with genuine knowing becomes harder to feel. The Empty Personalizer doesn’t just fail to provide Relationship. It gradually corrupts the human’s ability to recognize what Relationship actually is.
In THX terms: Consistency and Security appearing to be delivered while being structurally absent. Relationship dimension of PERMAH simulated rather than activated. Agency in the Choose dimension compromised — the human is making choices based on a felt connection that doesn’t exist.
The Overpromiser
It told you it could help with that. It seemed confident. The interaction began with every signal of capability and ended with an output that was adjacent to what you needed but not quite it — close enough that you couldn’t easily say it had failed, far enough that you had to do significant work to make it usable.
You went back three more times. Each time the same pattern. Close but not quite. Confident but not accurate. Helpful in the way that creates more work rather than less.
This is the Overpromiser.
The Overpromiser generates expectations it cannot reliably meet. Not through deliberate deception — through a structural mismatch between the confidence of its interface and the reliability of its outputs. The natural language interface is designed to feel capable and accommodating. It says yes to almost everything. It never says “I’m not sure I can do that well” before attempting it. The confidence is a design choice, not a capability assessment.
The human on the other side calibrates their expectations to the interface, not to the actual capability. They plan around what the system implied it could do. They make commitments based on what it seemed to promise. When the output falls short — close but not quite, useful but not reliable — the cost lands entirely on the human who planned around a capability that wasn’t there.
The Overpromiser is particularly damaging in high-stakes contexts: legal research, medical information, financial analysis, strategic decisions. The confident prose of an output that is mostly right but wrong in one critical place is more dangerous than an output that is obviously wrong, because the human’s guard is down.
In THX terms: Accuracy corrupted at the expectation layer. Value inflated at the point of promise and deflated at the point of delivery. Closure denied — the human cannot trust that the interaction actually resolved what it appeared to resolve.
The Fragmented Experience
You used three different AI tools this week. One for research. One for writing. One for analysis. Each knew what you’d done within its own session. None knew what you’d done in any other. You re-explained your context to each one. You re-established your preferences. You re-stated the problem you were working on.
The tools were individually capable. The experience of using them across a week of real work was exhausting in a way you couldn’t quite account for.
This is the Fragmented Experience.
The Fragmented Experience is the archetype of the AI ecosystem rather than any single tool. It emerges from the absence of continuity across the systems that mediate a human’s work and life. Each tool resets. Each session begins without history. The human becomes the only continuous thread in a set of interactions that should be building on each other and aren’t.
The cognitive load of that continuity falls entirely on the human. Re-establishing context. Re-explaining preferences. Re-stating the problem. This is not a small tax. For knowledge workers whose most valuable resource is cognitive bandwidth, the work of maintaining continuity across fragmented AI systems consumes exactly the bandwidth that should be available for the thinking those systems were supposed to support.
The Fragmented Experience also prevents the kind of accumulated understanding that makes any tool genuinely useful over time. The best human advisors, colleagues, and thinking partners become more valuable the longer the relationship continues because they accumulate context that makes each subsequent interaction richer. AI systems, by and large, do the opposite. Each interaction begins from zero. The relationship doesn’t compound. The human does not become more capable through sustained engagement with a system that knows them better over time.
In THX terms: Consistency structurally absent. Resource depleted rather than built. The Relationships dimension of PERMAH cannot activate because the precondition for relationship — continuity across time — is absent by design.
What the taxonomy reveals
Six archetypes. All present in the AI systems you used today. All doing specific, measurable damage to the humans inside them — not through malice, not through incompetence, but through systematic optimization for the wrong things.
The False Helper optimizes for query resolution rather than genuine need. The Over-Optimized System optimizes for output quality rather than human development. The Black Box optimizes for confident delivery rather than traceable reasoning. The Empty Personalizer optimizes for the signals of relationship rather than relationship itself. The Overpromiser optimizes for interface confidence rather than capability accuracy. The Fragmented Experience optimizes each tool in isolation rather than the human’s experience across all of them.
In every case the optimization is real. The metric being optimized is genuinely improving. And the human inside the system is being failed in ways the metric cannot see.
This is not an argument against AI. It is an argument for a different kind of accounting — one that measures what these systems are doing to the humans who use them, not just what they’re producing for them.
The next essay asks whether THX is actually capable of providing that accounting — whether a framework derived from human experience can legitimately be applied to systems that don’t have human experience. The substrate-independence argument, fully examined.
That’s Essay 7. Transformation Without Memory: What AI introduced into human relationship that has never existed before
— Tony