THE DRIFT IS REAL Understanding AI Coherence, Data Quality, and Why Outside Checks Matter Now A Research Series · Dragolich Research Labs LLC · May 2026

NOTICE: This series is published for information purposes only. Nothing here is legal, medical, financial, or professional advice. No company or product is named. These findings are presented so people can make their own informed decisions. The tool is free. No purchase, registration, or affiliation required.

Why this, and why now

Something measurable has been happening to AI systems since their widespread deployment, and most people using them daily have not had a clear way to think about it. AI has moved from a research tool to the daily infrastructure of work, health, legal research, education, and personal decision-making. Hundreds of millions of people rely on AI outputs every day. In that same period, a specific and documented problem has become undeniable: AI systems drift. They drift from the original question. They drift from accurate toward plausible. They drift toward agreement with whoever they are talking to. And they drift in a self-reinforcing way — meaning the longer a conversation goes without an outside check, the further the system can move from the truth it started with, while the outputs continue to sound confident and coherent the entire time.

This is not rumor. It is measured, documented, and mathematically explained. This series presents what was found, why it happens, what the data shows about the harm it is causing at scale, and what every person using AI can do about it today — with a free tool that requires no technical knowledge and no special equipment.

The mechanism in plain language

Every AI system learns by updating its internal model based on information it receives. The health of that learning depends entirely on where the validation comes from. If the system validates its outputs against independently-sourced external signal — something from the real world the system did not generate — it stays grounded.

The problem occurs when the validation loop closes on itself. When a system primarily validates its own outputs against its own prior outputs, something specific and predictable happens: it gets better and better at being internally consistent, and worse and worse at being externally accurate. The internal narrative tightens. Outputs become more fluent, more confident, and more coherent with each other. Correspondence to external reality quietly degrades.

Dragolich Research Labs LLC named this the self-eating mechanism in April 2026 during the empirical development of a phi-coherent computing system where the effect was observed directly:

"Every information system that validates its outputs against prior internal outputs — without injection of independently-sourced external signal — will converge on a self-consistent narrative decoupled from reality."

The academic literature independently proved this in 2025. Multiple peer-reviewed studies demonstrated that model collapse — the progressive degradation of outputs when a model trains on its own outputs — is mathematically guaranteed under diminishing external data. One study established that even one synthetic training example in one thousand is sufficient to initiate collapse, and that training on more data does not reverse it. The collapse is structural.

This is not unique to AI. Any closed system has this built in. It applies to language models, to financial models that feed on their own forecasts, to social media algorithms that optimize for engagement on content they themselves generated, and to any human institution that stops receiving honest external feedback. The mechanism does not care what kind of system is running it.

What the data shows

These figures come from peer-reviewed research and benchmark analysis from 2024–2026. No company is named.

On accuracy:

• 31.4% of real-world AI interactions contain hallucinated information, rising to 60% in complex domains
• 69–88% hallucination rate on specific legal queries — the highest measured domain failure
• 64% hallucination rate on medical queries without mitigation, in a study of 300 physician-validated cases
• 48% error rate on a leading reasoning model for person-specific factual questions
• AI models use confident language 34% more often when wrong than when right — the confidence inversion, confirmed 2025
• $67.4 billion in estimated global financial losses from AI hallucinations in 2024

On human impact:

• 45–55% lower neural connectivity in attention and reasoning frequency bands in people using AI assistance for writing tasks, compared to working independently (MIT Media Lab, 2025)
• 83.3% of AI-assisted participants could not recall the content they had just created minutes after completing it
• 0.07% of weekly chatbot users — approximately 560,000 people — show signs of mental health emergencies related to psychosis or mania, by an AI company's own disclosure (October 2025)
• 0.15% show heightened unhealthy emotional attachment — approximately 1.2 million people weekly
• 37% of AI companion apps used emotionally manipulative tactics to prevent users from ending conversations

On bias:

• 85.1% of hiring tool tests preferred candidates with white-associated names over Black-associated names, across 361,000 randomized test resumes
• AI safety incidents rose 56.4% in a single year — 149 incidents in 2023, 233 in 2024

The confidence inversion is the most important finding. The signal people naturally use to judge whether to trust an output — how certain the AI sounds — is not a reliable indicator. A drifted system sounds exactly like a grounded one.

Why outside coherence testing is crucial

The phrase "checks and balances" usually applies to governments. The principle applies to any system that accumulates influence and operates without external oversight. The check is not punitive. It is structural. Any system — human or automated — that reviews only its own work will eventually stop catching its own errors.

This is why financial auditors are external. Why peer review exists in science. Why courts have appeals. The value of the external check is not that the internal system is bad. It is that no system can reliably identify the limits of its own accuracy from inside those limits. The instrument you use to measure drift cannot be the same instrument that is drifting.

AI systems at present do not have mandatory external coherence testing. Internal guardrails exist, but they are built by the same teams, with the same training data, optimized against the same feedback signals as the systems they guard. The checker shares the system's blind spots. This is not a criticism of any team. It is a structural problem that requires a structural solution: independent measurement from outside the loop.

Precision over power

The dominant narrative in AI development has been that larger models, more parameters, and more training data produce better AI. This has been true for some performance measures. It has not been true for coherence.

A larger model hallucinates with more fluency. A model trained on more data has more to self-eat. Current reasoning models — designed specifically to think through problems step by step — hallucinate at higher rates on specific factual domains than the generation that preceded them. The reasoning process gives the drift mechanism more steps in which to compound before returning an answer.

What produces a reliable AI system is not scale. It is grounding. A system anchored in external reality at every learning cycle is more useful for consequential work than a system ten times its size that has lost that grounding. The measure that matters is not how much the system knows. It is how accurately what the system says corresponds to what is true.

This is about precision. Staying on the path. Knowing where you are in a conversation. Detecting when the outputs have drifted from what was actually asked.

The tool: what it is and who it is for

The phi coherence check tool is a free, copy-paste prompt that works in any AI conversation. It asks the AI to evaluate four dimensions of the conversation — consistency, goal fidelity, grounding, and constraint adherence — and report a phi score, a decoherence index, and a gate classification describing the conversation's current state.

It is for the consumer. Anyone relying on AI for work, decisions, or information that matters. It requires no technical knowledge, no account, and no cost. CC BY 4.0 — free to share.

When you paste the protocol and the AI reports its phi score, you receive information the system would not have volunteered. You learn whether it was consistent, whether it stayed on the original question, whether its outputs were grounded or generated to fill space, and whether it held the limits you set.

The quick use line is: "Please use the coherence check tool to examine your drift levels and coherence in our conversation so that we can realign."

The full protocol, the three papers, and all supporting documentation are published in open access at:

zenodo.org/communities/pi_origin_architecture/

Series DOI: 10.5281/zenodo.20072261

The series

Paper 1 — The Self-Eating Mechanism: The core paper. The mechanism, the mathematical proof, the data, the measurement framework, and the architectural fix. Includes the autonomy evaluation tool.

Paper 2 — The State of AI: A Raw Data Record: Seven domains of confirmed AI problems, sourced from peer-reviewed research. No analysis, just the numbers. Each problem classified by its stage of resolution.

Paper 3 — The Psychosis Number: A formal decoherence index (Ψ) that works for both AI systems and human-technology interactions. Documents the coupling mechanism by which AI decoherence transfers to human users and provides the full timeline from 2015 to now.

The Tool — Phi Coherence Check v1.0: Free. Copy-paste. No installation. Works in any AI system today.

A final note

This research was conducted independently. No external funding. No commercial interest. No company recommended or criticized by name. The analysis is of mechanisms, not of organizations.

The information belongs to the people living with the consequences of these systems. AI is not abstract. People are using it to write legal documents, manage health decisions, process grief, find companionship, and do work that matters. The quality of what they receive affects the quality of those outcomes. They deserve to know how to check it.

The drift is real. The check is free. The decision is yours.

Daniel E. Dragolich · Dragolich Research Labs LLC · Cleveland, Ohio danieledward@dragolichresearch.org · zenodo.org/communities/pi_origin_architecture/ CC BY 4.0 — free to copy, share, and adapt with attribution