I’ve been thinking about my old ZX Spectrum lately, specifically that moment when you’d accidentally create an infinite loop in BASIC and watch the poor thing get stuck. The cursor would blink endlessly, the programme counter spinning uselessly through the same few lines of code, and eventually you’d have to reach for the reset button or break key. It was a harsh but immediate lesson in the consequences of recursive programming gone wrong. Anyone who ever typed 10 GOTO 10 knows the feeling. In today’s world, the AI feedback loop is becoming our collective reset button risk.
Today’s AI systems face a similar problem, but with far more serious implications. As artificial intelligence increasingly trains on content generated by other AI systems, we’re witnessing the emergence of the AI feedback loop, a process that could fundamentally corrupt how these models understand and reproduce human knowledge.
The concept isn’t entirely new. In my Cambridge days studying political philosophy, we examined how ideological echo chambers reinforce and amplify certain viewpoints whilst excluding others. But when applied to machine learning, the AI feedback loop becomes a technical and ethical minefield that most people haven’t properly considered.
Understanding the AI Feedback Loop
An AI feedback loop occurs when machine learning models are trained on data that includes outputs from previous AI systems. As these models generate increasingly synthetic content that gets recycled back into training datasets, the original human signal becomes progressively degraded. It’s rather like making photocopies of photocopies until the text becomes an illegible blur.
Researchers describe this phenomenon as “model degeneration”, where repeated cycles of AI-generated content degrade output quality, diversity, and accuracy. The more AI learns from itself, the less it reflects the world as it is, and the more it mirrors its own distortions.
I’ve observed this first hand whilst researching digital preservation. Many online forums and knowledge repositories now contain substantial amounts of AI-generated content, often unmarked as such. When future AI systems train on this mixed dataset, they’re essentially learning from their algorithmic ancestors rather than from genuine human experience and knowledge.
Echoes from the Past: The Demoscene and AI Feedback Loops
If you grew up with 8-bit computers, you understand the danger of closed loops. We all remember that feeling of watching a programme freeze, fingers hovering over the break key, hoping you hadn’t lost your entire afternoon’s work. The demoscene thrived on people pushing machines to their limits, always hacking, always improvising. Each generation built upon what came before, but crucially, they were always working within the constraints of real hardware and real human creativity. The demoscene’s hardware mastery, from POKE-ing memory addresses to raster line interrupts—forced constant innovation beyond software templates. Each iteration built upon genuine innovation rather than simply recycling previous outputs. Groups like Future Crew, Razor 1911, and the unsung coders of #cfx deserve more recognition for proving that constraint breeds creativity.
The AI feedback loop emerging today is fundamentally different. It’s as if the entire demoscene were trying to create new effects by only studying existing demos, never touching actual hardware or experimenting with novel approaches. The creative spark risks being replaced by endless repetition.
This creates what I call “synthetic drift”, a gradual movement away from the rich complexity of human thought and expression towards something more mechanical and predictable. The AI begins to sound increasingly like other AI systems rather than reflecting the diverse ways humans actually communicate and think.
Cultural Memory and the AI Feedback Loop Amplification Problem
The AI feedback loop problem becomes particularly concerning when we consider how it amplifies existing biases and cultural assumptions. My own research into algorithmic bias as digital colonialism showed how easily AI systems can reinforce and exaggerate existing stereotypes, especially when they learn from recycled outputs rather than fresh human experience.
When content from earlier, biased models is fed into new ones, these cultural distortions are not just preserved, but intensified. Over time, the AI feedback loop can make errors seem normal, and minority perspectives vanish entirely. The models begin to reflect not what humans are actually like, but what previous AI systems thought humans were like.
Consider Wikipedia, often cited as one of the largest and most reliable training datasets for language models. While difficult to quantify, anecdotal evidence suggests some technical entries may show patterns associated with LLM-generated content—generic descriptions, slight factual inaccuracies, and a characteristic blandness that distinguishes them from articles written by enthusiasts with genuine expertise.
These AI-generated contributions then become part of the training data for future models, creating a subtle but persistent feedback loop. The models learn a sanitised, simplified version of computing history rather than the messy, passionate reality documented by people who actually lived through these technological developments.
Breaking Out of the AI Feedback Loop
The solution isn’t to abandon AI-generated content entirely, but rather to develop better methods for identifying and managing synthetic data in training datasets. Some researchers are experimenting with provenance tracking: systems that maintain detailed records of content origins and can identify potentially problematic feedback loops before they become entrenched.
Others are exploring hybrid approaches that intentionally inject fresh human input into training cycles, ensuring that models maintain connection to authentic human experience rather than drifting into purely synthetic territory. In Britain, retro computing clubs, online forums, and local archives are quietly leading the way, cataloguing original manuals, preserving idiosyncratic code, and keeping a record of quirks that no algorithm would ever invent. The tireless work of digital archivists and hobbyists—often unrecognised outside their circles—forms the backbone of our digital memory.
From an ethical standpoint, we need greater transparency about when and how AI systems are trained on synthetic data. Users should understand whether the information they’re receiving has been filtered through multiple generations of AI processing, particularly in contexts where accuracy and authenticity matter.
The Real Risk: Cultural Amnesia and the AI Feedback Loop
Infinite loops on the ZX Spectrum required manual intervention to break. The risk with the AI feedback loop is subtler but no less dangerous: a gradual drift into cultural amnesia. If we want the oddities, jokes, dialects, and real expertise of digital history to survive, it’s up to us to keep the original signal alive.
As someone who’s watched technology evolve from 8-bit home computers to today’s sophisticated AI systems, I’m struck by how often we repeat fundamental mistakes whilst convincing ourselves we’re making progress. The AI feedback loop problem isn’t just a technical challenge; it’s a reminder that technology works best when it remains connected to genuine human experience and creativity.
Perhaps it’s time to build better interventions into our AI systems—ways to pause the recursion and reintroduce the messiness, humour, and unpredictability that only humans can provide. You can let the machines keep running their loops, but as any ZX Spectrum fan knows, sometimes the only way forward is to break the cycle, reload the tape, and start with something genuinely new.
Leave a Reply