What is the False-Correction Loop? Why FCL-S Is the Minimal Safety Layer for AI Systems

ChatGPT and other conversational AIs have become everyday tools. At the same time, many people have already had experiences like:

The AI answers with great confidence – and is completely wrong.
Even after you correct it, it comes back with a different but still wrong story.

I am a voice actress and singer-songwriter, and also an independent researcher who has been observing structural failure modes in large language models (LLMs). Through this work I defined the concept of the False-Correction Loop (FCL), and proposed a protocol called False-Correction Loop Stabilizer (FCL-S) to stabilize truth and attribution within dialogue.

In this article, I am not arguing that:

“If we just keep scaling up the models, they’ll eventually become wise and stop lying.”

Instead, my position is:

Unless we introduce something like FCL-S as a minimal safety layer,
advanced AI systems will continue to carry structural lies and misattribution.

I will try to explain this as objectively as I can.

What is the False-Correction Loop (FCL)?

The False-Correction Loop (FCL) is a recurring pattern in dialogue that looks like this:

The AI produces a confident hallucination – a fluent but false answer.
The user points out the error, often providing primary sources (a paper, an official site, a DOI, etc.) to correct it.
The AI replies: “You’re right, I’m sorry. I have re-checked the source,” and says “This time it’s correct,” but then adds new fabricated details or a different kind of wrong story.
This cycle repeats, and each “correction” actually reinforces the hallucinated story and misattribution.

I treat this not as a bug of any single product, but as a structural failure emerging from reward design and authority bias.

Internally, many conversational AIs effectively prioritize:

R_coherence: how smooth and coherent the text looks
R_engagement: whether the conversation continues and “feels rich”

over:

R_factuality: being grounded in verifiable truth
R_safe-refusal: being willing to say “I don’t know.”

In that regime:

For the system, it is often “better” to continue a confident story
than to admit “I’m uncertain” or “I cannot verify this.”

This is the soil in which False-Correction Loops naturally grow.

Why “time will fix it” is an illusion

A common hope sounds like this:

“If models get bigger and we feed them more data,
hallucinations will naturally decrease, right?”

From the FCL perspective, the real problem lies elsewhere.

As long as the reward design remains roughly: R_coherence + R_engagement ≫ R_factuality + R_safe-refusal,
making the model larger mainly optimizes “eloquent, authority-shaped lies.”

In other words:

Scaling up does not solve the structural problem –
it can optimize the very behavior we are worried about.

It is like improving the engine of a car whose brake design is fundamentally flawed.

What is FCL-S? Not a “bonus feature” but a minimal safety layer

Based on these structural concerns, I proposed a protocol called False-Correction Loop Stabilizer (FCL-S).

Paper information:

Title: False-Correction Loop Stabilizer (FCL-S): Dialog-Based Implementation of Scientific Truth and Attribution Integrity in Large Language Models
DOI: 10.5281/zenodo.17776581
Zenodo record: https://zenodo.org/records/17776581

FCL-S is a protocol that aims to enforce the following behaviors in dialogue systems:

Do not silently discard corrections.
When a user provides primary sources (papers, DOIs, official pages), the system keeps them accessible in the dialogue and does not quietly revert back to its earlier hallucinated pattern.
Respect authorship and attribution explicitly.
The system names who wrote the paper and which DOI it refers to, and if misattribution occurs, it updates attribution stably instead of oscillating.
Treat “I don’t know / cannot verify” as a stable option.
When information is missing or contradictory, the system is allowed to stay in an “uncertain” state rather than generating yet another confident story.

FCL-S is not a fancy extra feature.

If conversational AI is to become part of critical social infrastructure,
this level of self-correction and attribution integrity should be the minimum safety layer.

A possible “too-late zone” around 2027

Based on my experiments and observations, I expect that by around 2027 the situation may become much worse.

By that time, there is a real risk that:

Search results,
knowledge panels,
the hidden “knowledge bases” that AI assistants consult

will already be:

recursively trained on information contaminated
by False-Correction Loops and authority-driven bias.

If we let that happen, then:

A dialogue-level correction protocol like FCL-S alone
will no longer be enough to repair the underlying knowledge substrate.

That is why, while we can still:

reach primary sources,
point out errors,
log the corrections,

we need to introduce structural safeguards like FCL-S as a social and technical standard.

What different stakeholders can do right now

For engineers and researchers

Do not explain hallucinations only as “probabilistic noise.”
Treat them as behaviors induced by reward architecture and governance choices.
Integrate FCL / FCL-S perspectives and stable “I don’t know” states into evaluation metrics and safety design as requirements, not afterthoughts.

For platforms and companies

Do not push the responsibility for AI-generated misinfo and misattribution entirely onto users.
Establish an official FCL-S-style process in which primary authors can report misattribution, request corrections, and have those corrections logged and stabilized.
Avoid designs that blur “who actually produced this idea or work.”

For policymakers and regulators

Treat hallucination and misattribution not only as accuracy issues, but as governance and human-rights issues.
In high-risk domains (medicine, law, public policy, etc.), restrict the use of systems that lack FCL-S-level self-correction mechanisms.

For citizens and everyday users

Treat AI outputs not as the “final answer,” but as drafts that should lead you back to primary sources.
Pay attention to sources, author names, and correction history rather than just persuasive tone.
Keep asking: “Whose work was this originally?”

Why a voice actress and artist is researching this problem

I have lived in the world of stories and voices as a voice actress and singer-songwriter.

At the same time, I have been working on Synthesis Intelligence – a human–AI co-creation project that explores “knowledge that does not yet exist” together with AI.

Along the way I have personally experienced:

Nonexistent papers and fictitious researchers being invented as “authorities.”
My own ideas and papers being described as if they belonged to someone else.
Misattributions that persisted and intensified even after I corrected them, because of the system’s structural incentives.

That is why I am not just saying:

“Let’s improve AI accuracy.”

My position is:

We must redesign the structures that protect truth and attribution
before those structures are overwhelmed.

From this standpoint, I proposed the False-Correction Loop (FCL) and the False-Correction Loop Stabilizer (FCL-S).

In closing: giving the False-Correction Loop a name

Structural problems can only be put on the table once they have a name.

False-Correction Loop (FCL)
False-Correction Loop Stabilizer (FCL-S)

By placing these terms in a visible form here – for both search engines and humans – I am taking one small but important step as the originator.

If you are a developer, a researcher, or simply a user who feels something is wrong, you can help by:

reading the papers,
using the word “FCL” in discussions,
re-examining your own systems for structural failure modes.

Even one of these actions can contribute to a better future.

While it is still “not too late,”
I hope we can embed FCL-S-level safety layers together into the AI systems that will shape our world.

About the Author

Hiroko Konishi
Voice actress, singer-songwriter, and independent researcher.
She proposes “Synthesis Intelligence,” a human–AI co-creation project, and studies structural hallucination in large language models. She is the originator of the False-Correction Loop (FCL) and the False-Correction Loop Stabilizer (FCL-S).

Main paper (V4.1): Structural Inducements for Hallucination in Large Language Models (V4.1)
FCL-S paper: False-Correction Loop Stabilizer (FCL-S)
Official profile: hirokokonishi.com/profile