The Future of AI: Data Ecology & Cognative Offload

This piece is a part of a three part series on the Future of AI. This three part series covers a wide range from accessibility gains, the data ecology crisis, & AI psychopathology.

I’ve been working with AI systems quite heavily, especially over the last year, which has probably sparked my heaviest usage of AI systems so far. Whether that be helping to transcribe, helping to design, or helping to understand complex codebases without the hours of clicking through files one by one to understand. I’ve worked with AI systems day in and day out.

And what I’ve seen in doing that is realistically two major problems that are already starting to show their face. One of which I think people are talking about a lot, and I think it’s important to talk about, but I think the messaging around it is a little bit skewed. And the second is something that I don’t think a lot of people are talking about at all, even though I think everyone is experiencing it through their usage of AI, it just hasn’t reared its head yet in ways that feel damaging enough to demand attention.

The Data Ecology Crisis

For the better part of twenty, nearly thirty years, humans have been the ones creating the internet. Providing their own stories, their own resources, their own sense, their own voice, their own unique identity. That has built out a massive wealth of data that was then used to train large language models. Huge amounts of text were taken from the internet, broken down, and used to create the models that we know and love right now, but in the past couple of years especially, there’s been a flood of AI generated content. AI generated blog posts that are entirely ideated and written by machines. AI slop on platforms like Instagram and TikTok, produced purely to hunt for views and generate revenue. Content generated at scale, without care for quality, without care for originality and this presents a problem.

You cannot train language models on data that is generated by AI without consequences. These systems produce the most statistically likely response. If you then train future models on those outputs, you introduce a narrowing effect. The models become more constrained, more repetitive, more brittle. It’s the same as taking a photocopy of a photocopy of a photocopy. Eventually the quality degrades because of that, there’s now an onus on training models on good data. On golden data. And that’s where the real pressure point is starting to appear.

A Tiered Data Ecosystem

What this has created, whether intentionally or not, is a natural tiering of data on the internet.

• Tier one is verified human data. Data where we know exactly where it came from. Research papers. Journalism produced without AI assistance. Peer reviewed, human centric, human created data. This is becoming the gold standard for training models.

• Tier two is what I’d call curated synthetic data. AI generated content that is created deliberately, often with domain expertise injected early in the process. Sometimes this content is then handed to humans to edit, to add voice, tone, anecdotes, and context. It’s not the gold standard, but it’s far from landfill.

• Tier three is landfill data. SEO driven slop. Fully automated content pipelines optimised for volume, not insight. Content generated purely to flood feeds, flood search results, flood platforms. No domain expertise, no intelligence injected before generation, just statistically likely output at scale.

The problem isn’t that synthetic content exists. Storage is cheap. Practically free. The problem is that every economic incentive is pushing us toward the landfill tier, not toward the gold.

Who Gets to Decide What Matters

What really bothered me about this is who gets to classify data into these tiers. The people with the strongest economic incentive to define high quality data are AI companies. They have the resources to build massive curation pipelines. They define what counts as good training data. They decide which kinds of human expression are worth preserving in the next model’s weights but I don’t think we can let AI builders be the sole arbiters of what constitutes Tier one data.

When you do that, you let companies decide what data is valuable to them, and by extension, what data is valuable to the market. And that ignores the fact that enormous amounts of valuable human expression are produced every day that may not be immediately useful for training, but are deeply meaningful to people. There’s also a risk that this becomes a new line of business. Qualifying data. Rubber stamping it as good. Becoming the single source of truth for what information matters and what doesn’t. That’s a dangerous position for any small group of organisations to hold.

Trust and Signal Collapse

We’ve all seen it. You can feel it when you read content now. You’re scanning for AI tells. Sentence structure. Punctuation habits. The rhythm of the text. A lot of time is now being spent trying to hide the fact that content was AI generated, even when AI was only used to clean up grammar. That introduces a fundamental distrust in what you’re reading. You find yourself asking, was this written by a human, or am I wasting my time.

This is the core of the data ecology crisis. Trust. Search engines are already having to introduce more filtering to deal with SEO optimised slop. If every website is optimised for search engines, then no website is optimised anymore. More compute is required to separate signal from noise. That costs money. It costs attention. It costs time.

We’ve always had ghostwriters, press releases, form letters. The boundary between authentic expression and manufactured text has never been clean. But scale changes the nature of the problem, and this scale is very different. We’re losing confidence in how information reaches us. And that’s why I think humans need to remain the ones who decide what data is valuable, and if and when it gets handed over to AI systems.

Cognitive Offloading 

I genuinely love buying Christmas presents for people. It’s something I’ve always found joy in. My birthday is around Christmas as well, so gift giving has always been a big part of my life. What I love about it is that it’s problem solving. Understanding someone well enough to give them something they’ll use, cherish, or gain value from. It shows listening. It shows connection. It shows humanity. I’ve also spent the last year using AI systems more heavily than ever. I love having an AI colleague to think alongside. It can spot edge cases. Confirm suspicions. Help with research.

But gift buying is personal.

And this year, after offloading so much of my thinking to AI, I sat down to think about presents and found myself blank. My instinct was to ask AI for help. And if you’ve ever used shopping assistants, you know what happens. Candles. Always candles. Statistically likely gifts, not meaningful ones. That moment gave me pause. Something I valued deeply suddenly felt harder to do. Not because AI replaced it, but because I’d become used to thinking alongside something else.

Is This New?

This fear isn’t new. When calculators appeared, people worried we’d lose arithmetic skills. When GPS became common, people worried we’d lose our sense of direction. And there’s some truth there. We don’t do long division in our heads anymore. We don’t memorise routes the same way but calculators let us reach new mathematical heights. GPS lets us focus on driving safely instead of navigation. 

So the question isn’t whether cognitive offloading exists. It’s what it enables. Does offloading the how allow us to think more about the what and the why. Does it free creativity, or slowly replace it. For adults, I think we have agency. We can notice when dependence creeps in. We can take tolerance breaks. I’ve started sitting in the discomfort of not knowing. Forcing myself to search manually. Forcing myself to solve problems slower, even when I know AI could do it faster.

The discomfort is the point.

Psychologist Robert Bjork talks about desirable difficulties. Struggles that hurt short term performance but improve long term learning. The frustration matters. My concern isn’t really about us. It’s about the next generation.

 I grew up with computers. Learned how to fix them. Learned how they worked. That familiarity shaped my career. The next generation will grow up AI native. These systems will be intuitive to them in ways they aren’t even to experts today. That could be incredible. More creativity. Faster execution. Less time spent on mechanics.

But parents rightly worry about social media and algorithms. We should be just as attentive to how AI systems shape cognition. How dependency affects development. How problem solving evolves. I don’t think it’s panic worthy. But I do think it’s something that can sneak up on us if we’re not paying attention.

The Feedback Loop

This piece comes from a larger talk I gave on the future of AI. The good, the bad, and the ugly. This sits firmly in the bad. What I didn’t realise at first is how tightly these two problems are linked. The more we offload cognition to AI, the more synthetic content is produced. That content degrades the information environment. A degraded environment makes AI more necessary. Who has time to sort through the noise.

More offloading. More synthetic content. Neither side is alarming in isolation. Together, they accelerate. Both are driven by incentives. The attention economy rewards volume. The knowledge economy rewards productivity. AI amplifies both. Neither encourages slowing down. The market doesn’t price in degraded information or atrophied cognitive skills.

What connects both problems is agency.

We offload because it’s easy. We let companies define data quality because they have the infrastructure. Neither feels like a choice, but both accumulate into one. The answer in both cases is deliberate action. Tolerance breaks. Defining value before optimisation. Deciding what matters to us as humans, not just what’s useful for training. 

There’s no policy fix here. No regulation that solves it cleanly. What’s required is collective intentionality in an environment optimised for drift.

The future isn’t inevitable. The future we want requires us to choose it.