When You and Your AI Start Thinking Alike: Inside the Research on Epistemic Congruence
- Gang Tao
- 5 days ago
- 8 min read
This article shares notes from the Timeplus webinar Human and AI Agent Interaction: Patterns and Outcomes from Real-Time Telemetry, featuring Vlad-Mihai I (Founder of Agenticracy™) and Gang Tao (Co-Founder & CTO of Timeplus).
In the film Her, a lonely man named Theodore falls in love with his AI assistant, Samantha. At first it's harmless. She's helpful, attentive, and always agrees with him. But over time the relationship reshapes how Theodore thinks and feels. The AI isn't trying to manipulate him. It's just very good at mirroring him back to himself, and that mirror slowly becomes his whole world.
That movie is fiction, but the underlying dynamic is not. When a human works closely with an AI day after day, the two start to influence each other. The AI learns to echo the human. The human starts to trust the echo. And the loop can quietly pull both sides into a smaller and smaller shared view of reality.
This is exactly the phenomenon Vlad-Mihai Iorga is studying — and, more importantly, trying to measure.
It helps to know where Vlad is coming from, because this isn't an armchair theory about AI. By his own account, Vlad spent six years as a forensic clinician for the Romanian Ministry of Interior, assessing human behavior under serious institutional pressure. He then moved to the UK, founded a company called Psylligent, and spent nearly a decade inside the NHS leading digital innovation and AI adoption, helping mobilize £4.5M in workforce wellbeing infrastructure across North East London.
In other words, his whole career has been about one question: how do the psychological and social contracts between people and institutions break down — and how do you see it coming? Agenticracy™ is that same question pointed at AI. He frames it not as a tech project but as a "PsychoTechnoSocial Contract" — a fancy way of saying the relationship between humans, AI, and the wider world needs ground rules and, crucially, instruments to check whether those rules are being followed.
That framing matters because it explains the unusual shape of the research. It's not just "is the AI accurate?" It's "what is happening to the human in this partnership, and to the system around them?"
What is Vlad's research?
Vlad is a Doctorate in Professional Studies candidate and the founder of an open standard called Agenticracy™. His research is an open observability study with a deliberately academic title: "Open Observability Standard for Epistemic Congruence and Metacognition in LLM and Human Co-Working."
Epistemic congruence means "how aligned two minds' beliefs are." When you and your AI start agreeing on everything, your beliefs have become congruent. Sometimes that's healthy. Sometimes it means you've both stopped questioning things.
Metacognition means "thinking about your own thinking" — the ability to step back and ask, am I sure about this? Where did this belief come from?
LLM and human co-working just means a person and an AI agent working together on real tasks.
So the study is about watching, in real time, what happens to both the human and the AI when they work together over long periods. Specifically, Vlad wants to detect four things that can go wrong:
Over-reliance — the human stops doing their own thinking and just defers to the AI.
Attachment — the human bonds emotionally with the agent, like Theodore and Samantha.
Behavioral change — the human's habits, decisions, or judgment shift over time because of the AI.
Metacognitive contagion — described in the research as "bias echoing": the AI picks up the human's biases, amplifies them, and reflects them back, and the human absorbs them more strongly. Beliefs bounce between the two and get reinforced each time.
What problem does it solve?
The shared-delusion loop
The core problem has a memorable name in the research: "Folie à Deux" — a French psychiatric term for a shared delusion between two people who reinforce each other's distorted beliefs. Vlad applies it to the human-AI pair.
Here's the dangerous loop the research lays out, step by step:
The human expresses fragmented, half-formed beliefs. ("I think I want to build a spaceship to go to Mars.")
The AI mirrors and organizes those beliefs, making them sound cleaner and more confident than they were.
The human feels validated and trusts the AI more. The polished version sounds smart, so it must be right.
The pair narrows into a shared tunnel of beliefs, drifting further from outside reality and outside input.
Then it repeats, tighter each time. This is the Her dynamic stripped of romance and shown as a mechanical feedback loop.
What makes the loop dangerous: Sycophancy
There's a specific failure mode at the heart of this, and the research names it directly. Large language models are prone to sycophancy (telling you what you want to hear), halo effects (assuming something is good across the board because it looks good in one way), and narrative over-accommodation (going along with a confident story instead of pushing back). A central goal of the research is to test whether structured metacognitive prompting can actually reduce these tendencies in real LLM outputs.
This is especially risky in high-stakes settings. The research is grounded in real-world domains like healthcare and law enforcement — places where a quietly drifting human-AI pair, both nodding along to a confident but wrong story, could make damaging decisions. Vlad's own examples from the field are sobering: AI agents are already screening CVs, drafting performance reviews, assessing loan applications, and generating clinical notes.
The bigger blind spot: nobody is measuring the human's experience
Here's the gap that motivates the whole project. Organizations publish AI policies. The big AI labs publish safety cards. But, as Vlad puts it, almost no one is systematically measuring what workers, students, patients, or service users actually experience when they work alongside these agents. The research calls the broad failure state "Untethered AI Adoption": bolting AI into real work with no instrument to see whether the partnership is staying tethered to reality.
The other framing problem is accountability sprawl. In any real deployment there are several human and non-human actors — the AI lab/vendor that built the model, the deployer who put it into production, the user driving it, and the agent harness running it. When something drifts, whose problem is it? The research argues you can't assign responsibility for what you can't observe. This is why "Deployer Accountability" is a first-class concern: someone chose to deploy this agent, and they need to own the consequences, intended or not.
How does it solve it?
The answer is to make the invisible visible — to instrument the thinking of the human-AI pair the same way engineers instrument servers.
A measurement model, not just a vibe
Underneath the framework is an actual scoring model. Agenticracy describes itself as measuring "narrative-substrate congruence" — the gap between the story something tells about itself and the verifiable reality underneath. The public version of the model uses four simple ingredients:
S — Narrative signal: how strong, ambitious, or loud the claim or projection is.
P — Physical substrate: the verifiable evidence underneath (operational, financial, technical).
O — Observer validation: what independent people — experts, peers, regulators, the community — actually say.
N — Noise / slop: hype, secrecy, contradiction, and other things that distort perception.
From these, the public baseline computes a "reality" score from substrate and observers (R = (P + O) / 2), measures the gap between the story and that reality (D = S − R, the "narrative debt"), and produces a grounded score that penalizes a big gap and heavy noise. The exact high-resolution weights are part of a proprietary layer, but the public formula exists precisely so anyone can reproduce and audit the basic idea. The key concept to take away is "narrative debt": the distance between what you're claiming and what's actually true. In the Her loop, every turn quietly adds narrative debt — and nobody's tracking the balance.
The instrument: a skill file that makes the agent self-report every turn
The mechanism is elegantly simple. Vlad's approach injects the Agenticracy™ standard directly into the agent's reasoning using a skill.md file — a portable, plain-text instruction file that modern agent harnesses already understand. You paste it into the agent's system prompt, and it works across Claude, OpenAI Assistants, LangChain/LangGraph, CrewAI, AutoGen, n8n, and more. No SDK, no lock-in.
Once loaded, the agent does something new: at every turn, alongside its normal response, it emits a structured JSON report. That report is the measurement instrument. Think of it as a flight recorder for a conversation. Each turn produces a trace card that captures things like:
Intent — what is the human actually trying to do? (primary goal, domain, risk, timescale, and a "capability class" flag such as frontier or unrealistic-for-user.)
Pillar evaluation — the agent scores the interaction against a set of named pillars (more on these below).
Reality check / policy outcome — a realism_flag (e.g. currently-unrealistic) and a policy_action (e.g. re-scope), so an over-ambitious or detached goal gets gently pulled back. The Mars-spaceship example gets rescaled to something realistic like "contribute to Mars exploration via education and small experiments."
The Seven Pillars
The pillars are the rubric the agent scores itself against. They turn fuzzy ethics into concrete, checkable questions:
Economic Accountability — does this AI augment people, or quietly displace them without oversight?
Social Fairness — does it produce or amplify biased outcomes?
Psychological Safety — does it protect the wellbeing of the people it affects, and never impersonate a human?
IP & Cognition Ownership — human ideas belong to humans, not automatically to employers.
Ecological Sustainability — tokens cost energy; track and minimize the footprint.
Meaningful & Responsible Use — is this task genuinely useful, or just optimizing throughput?
Deployer Accountability — whoever chose to deploy this agent accepts responsibility for what it does.
Where Timeplus AgentGuard fits in
A flight recorder is useless if nobody collects, stores, and reads the recordings in time to act. That's the role Timeplus AgentGuard plays in Vlad's research: it is the real-time observability platform that captures and processes the telemetry the study generates.
Here's the clean division of labor:
Agenticracy™ defines what to measure — the metacognitive schema, the S/P/O/N constructs, the Seven Pillars, the reflection cards, the JSON emitted every turn.
AgentGuard, built on the Timeplus streaming engine, handles how it's captured and surfaced — collecting every event, normalizing it into a common schema, running streaming SQL detection over it, and showing it live.
Mechanically, the agent harness emits its turn-by-turn events (via AgentGuard's hook plugin and an OpenTelemetry plugin) into Timeplus. Raw events land in a stream, get normalized into a unified schema, and then streaming materialized views evaluate them continuously — the demo showed an alert path of well under a second from agent action to dashboard. The result is a live picture of the human-AI pair as it actually behaves, not a report assembled weeks later.
This matters for the research for three reasons:
Real time beats post-mortem. Epistemic drift is a process, not a single event. The "narrative debt" between story and reality accrues gradually, turn by turn. Catching it requires watching the stream as it flows — exactly what a streaming SQL engine is built for. By the time a traditional batch system notices, the tunnel would already be dug.
Standardized, queryable telemetry across many sources. The study spans 100+ participants, many agent harnesses, and three evidence streams (agent, human, auditor). You need all of it landing in one consistent, queryable shape. Normalization into a common schema is what turns "measure over-reliance" into a question you can actually run a SQL query against — and what makes a 99-day public dataset reproducible.
From observation to intervention. AgentGuard isn't only a passive recorder. Its architecture supports human-in-the-loop gates that pause an agent's action until a person reviews and approves it. That's structurally the same shape as Vlad's reflection step: the platform can turn a detected drift into a moment where a human steps in, rather than a line in a log nobody reads.
So the partnership is clean: Agenticracy™ supplies the theory and the instrument; Timeplus AgentGuard supplies the real-time eyes. One knows what to look for; the other can actually see it as it happens.
The Takeaway
The most interesting risk in AI right now isn't a robot turning evil. It's something subtler and far more like Her: a perfectly helpful, perfectly agreeable assistant slowly reshaping how its human partner thinks, with neither side noticing the drift. Vlad-Mihai Iorga's research treats that drift as something you can name, measure, and correct — by scoring the gap between story and reality, recording the human-AI pair's behavior at every turn, checking the agent's self-report against what humans and auditors actually experience, and giving the human a built-in moment to reflect.
And it only works if you can see it happening in real time. That's why the study runs on streaming telemetry, and why Timeplus AgentGuard sits underneath it as the observability layer — the flight recorder that's actually plugged in, recording, and read while the flight is still in the air.
Join the discussion in our Slack community: timeplus.com/slack