The AI Feedback Loop: Is AI Learning From AI?

The AI Feedback Loop. AI Learning From AI.

As technology companies pay millions for access to human conversations, AI-generated content is increasingly finding its way into the same online communities. Are we approaching a future where artificial intelligence is learning from information that artificial intelligence itself created?

Abet News | Artificial Intelligence / Technology | November 18, 2026

The internet has become an AI training ground

For years, the internet has been one of the most valuable resources for artificial intelligence companies. Billions of conversations, articles, technical discussions, product reviews, personal experiences, and opinions have provided the raw material used to develop increasingly sophisticated AI systems.

But something is changing.

The very technology being trained on human-generated content is now producing enormous amounts of new content, much of which is finding its way back onto the internet.

Reddit, one of the largest repositories of human conversations online, has become an important example of this transformation.

In February 2024, Google announced an expanded partnership with Reddit, gaining access to its Data API to obtain fresh, structured content. The agreement also included provisions allowing Google to train on and otherwise use Reddit information.

Reuters reported that the licensing agreement was worth approximately $60 million annually, citing sources familiar with the deal.

This wasn’t simply a purchase of information. It represented the growing commercial value of conversations that millions of people had contributed to online communities.

And Google wasn’t just interested in what people were discussing. It was interested in the collective knowledge contained within those discussions.

Google, Reddit, and Gemini: What we know

The relationship goes beyond speculation.

In May 2025, proposed findings of fact submitted in the U.S. government’s Google search antitrust proceedings documented testimony from Google representatives confirming that Google pretrains its Gemini models using Reddit data.

The documents also referenced Reddit data in Gemini’s training documentation.

This establishes that Reddit content has been used in Gemini’s model-training process, not merely that Google can display Reddit search results.

However, the complete training datasets, filtering methods, and proportion of Reddit material used have not been publicly disclosed.

That distinction matters.

Google has invested in access to human conversations because they contain something traditional databases often lack: firsthand experiences, disagreements, practical advice, and perspectives from people dealing with real-world situations.

But what happens when those conversations increasingly contain AI-generated material?

When Reddit stops being exclusively human

Anyone spending time on Reddit recently may have noticed a change in the way some discussions read.

Posts can appear remarkably polished, with predictable introductions, perfectly organized arguments, familiar expressions, and carefully constructed conclusions.

Some discussions feel less like spontaneous conversations and more like generated content designed to attract engagement.

There are legitimate reasons people use AI to help write. Language translation, accessibility, grammar assistance, and organizing thoughts are useful applications.

The problem is different when automated accounts generate content and participate in conversations while pretending to be real people.

Reddit CEO Steve Huffman has acknowledged this distinction. During Reddit’s fourth-quarter 2025 earnings discussion, he recognized the growing use of AI to write online while emphasizing that fully automated, bot-like behavior is something the platform does not want.

Reddit has also been investing in automated defenses against spam and inauthentic activity.

In July 2026, the company reported a 20% reduction in user exposure to spam following improvements to its detection systems.

These measures are encouraging, but they also demonstrate the scale of the challenge.

If AI-generated content is increasingly entering the same communities that AI companies rely on for training material, how can those companies reliably distinguish original human knowledge from machine-generated responses?

The scientific concern: Model collapse

This isn’t merely a theoretical concern.

In July 2024, researchers published a significant study in the scientific journal Nature, titled AI Models Collapse When Trained on Recursively Generated Data.

The research examined what happens when successive generations of AI models are trained on information generated by earlier models.

The researchers demonstrated that indiscriminate use of AI-generated material in training can progressively degrade model performance and cause less common information to disappear from the resulting models.

They called this phenomenon model collapse.

The underlying concern is straightforward: when models repeatedly learn from their own generated representations of the world, they risk losing the diversity and accuracy present in the original human-generated information.

This doesn’t mean every AI system trained on synthetic information will collapse. Carefully curated datasets, verified original sources, and appropriate training techniques can mitigate the risk.

But the research establishes that the problem deserves serious attention.

And the internet is becoming an increasingly important part of that equation.

Synthetic Empire is a high-stakes crime thriller following investigative journalist Anabel Kaspario as she uncovers the dark intersections of AI technology...

Synthetic Empire is a high-stakes crime thriller following investigative journalist Anabel Kaspario as she uncovers the dark intersections of AI technology…

A real-world example: When Gemini changes its mind

The implications aren’t limited to future AI training models. They also affect the reliability of the answers consumers receive today.

Recently, we conducted an informal experiment with Google’s Gemini involving Apple’s iOS 27.2.1 update and the iPhone 15 Pro.

The question was straightforward: How was the update performing on an older iPhone, considering concerns that software updates can sometimes affect the performance and battery life of previous-generation devices?

Gemini initially responded that iOS 27.2.1 performed smoothly and reliably on the iPhone 15 Pro, describing it primarily as a stability and security update rather than a major feature release.

However, after updating and experiencing audio fluctuations and rapid battery drain, we returned to Gemini with a different question.

We explained that users were reporting similar audio problems and battery drainage and asked whether updating was advisable.

Gemini changed its recommendation:

“I would hold off on updating to iOS 27.2.1 if your iPhone 15 Pro is currently running smoothly on iOS 27.2.”

We then asked whether Gemini was aware of these problems before we brought them to its attention.

Its response was revealing: it admitted that it wasn’t aware of the issues before being asked and explained that its base knowledge did not automatically include live software updates or real-time bug reports. It had used web search during the conversation to investigate the release and user reports.

We then asked why it frequently referenced Reddit and why it hadn’t mentioned the reported problems in its original response.

Gemini explained that Reddit is often one of the fastest places to identify early iOS bug patterns, particularly following minor software releases.

Finally, we opened a completely new conversation and asked essentially the same question: Was it safe to update an iPhone 15 Pro from iOS 27.2 to iOS 27.2.1?

The answer was once again different.

Gemini recommended updating, describing it as safe and advisable.

This experiment does not establish that Gemini deliberately withheld information, that the reported iOS problems affect all iPhone 15 Pro devices, or that Reddit data caused the inconsistency.

It does, however, illustrate a significant weakness in AI-assisted consumer advice.

An assistant can have access to relevant information, acknowledge that information when prompted, and still produce a different recommendation in a fresh conversation.

The question isn’t simply whether AI can retrieve information. It’s whether it consistently retrieves, evaluates, and communicates the information consumers need before making decisions.

The danger of circular information

Consider a possible scenario.

A person experiences a technical problem and publishes a detailed explanation on Reddit.

An AI assistant uses that discussion as a reference. Another person asks an AI to explain the same problem, and the AI generates an answer based partly on the original discussion.

That generated answer is subsequently published on another website or social platform.

A future AI system encounters both the original Reddit discussion and the generated explanation.

If the generated material is mistakenly treated as an independent account, the same underlying information may appear to have multiple confirmations.

Now imagine this happening across millions of conversations, product reviews, technical explanations, and news discussions.

The problem becomes more complicated when inaccurate information is repeated, summarized, rewritten, and redistributed until its original source becomes difficult to identify.

Five websites repeating the same incorrect claim do not constitute five independent confirmations.

And five AI assistants repeating the same answer don’t necessarily provide five independent opinions.

This is why information provenance, source transparency, and independent verification matter more than ever.

Who protects the authenticity of human knowledge?

There is an uncomfortable irony in the current AI economy.

Technology companies recognize that human conversations are valuable enough to justify multimillion-dollar licensing agreements.

At the same time, AI tools make it easier than ever to generate synthetic conversations that can occupy those same spaces.

Reddit is attempting to protect its communities through automated detection, account verification, and greater transparency around automated activity. Other platforms face similar challenges.

But content moderation alone cannot solve the entire problem.

AI companies also need to consider how training datasets are selected, how synthetic material is identified, how sources are evaluated, and how information is independently verified.

Consumers need better transparency, too. When an AI assistant provides an answer based on online discussions, readers should be able to understand whether those sources are firsthand experiences, independently verified reporting, or potentially generated content.

The future of the internet depends on trust

Artificial intelligence has enormous potential. It can make information more accessible, help people solve complicated problems, and make technology more useful.

But its future depends on the quality of the information it receives.

Human experience, independent reporting, scientific research, and genuine expertise cannot simply be replaced by an endless supply of generated text.

As AI becomes increasingly capable of producing convincing conversations, distinguishing authentic knowledge from artificial repetition will become one of the internet’s most important challenges.

Perhaps the greatest irony is that the technology industry is now facing a problem of its own creation: the more AI-generated information enters the internet, the harder it may become to preserve the original human knowledge that made the internet so valuable to AI in the first place.

If AI learns from humans, but humans increasingly learn from AI, who is responsible for ensuring that the information being passed around is actually true?

That’s a question the technology industry cannot afford to ignore.

Abet News — Technology, creativity, and the questions shaping our digital future.

Amy H.

© 2026 Abet News. All rights reserved.

Related Posts

Apple iPhone Pro screen showing audio playing, illustrating reported iOS 27.2.1 audio and volume problems.
Apple's Keyboard: Smart Design or Stubbornness?
AI Doesn’t Love You. So Why Does It Feel Like It Understands You?

Leave a Reply
Apple TV+
Apple TV+ Ad
JMI Construction
JMI Construction Ad
EzTen Website Design