Garbage In, Garbage Out: Now With AI Attached

Garbage in, garbage out. It's one of the oldest phrases in computing, older than most of the people repeating it, and by now it's fair to call it a cliché. Clichés earn that status by being true often enough that people eventually stop really listening to them. This one deserves a second look, because it has never mattered more than it does right now.

There's a version of AI adoption that goes like this: buy the tool, point it at the business, wait for the insight. It's the pitch in most vendor decks, and it's wrong in a very specific way. AI doesn't fix bad data. It just makes bad data faster, more confident, and considerably harder to spot.

That's not a hunch, it's showing up in the numbers. A recent industry report found that three in four businesses say their AI initiatives are failing to deliver. And when leaders were asked why, the answer wasn't poor technology. It was poor data: outdated, fragmented, unstructured, incomplete. The exact ingredients that made analysts distrust their own dashboards for years are now being fed straight into models people are trusting to make decisions on their behalf.

What's changed isn't the principle. It's the scale of the garbage, and the speed at which it now gets out the other side.

The old GIGO problem was slow. This one isn't.

When a human analyst worked with bad data, the damage was bounded by how much a person could process in a day. A flawed spreadsheet might mislead one report, one meeting, one decision. Painful, but contained.

AI removes that ceiling. A model trained on inconsistent, unstructured or biased data doesn't produce one bad answer. It produces a pattern, applied at scale, delivered with the same fluent confidence whether it's right or wrong. It can churn through thousands of documents a minute. It doesn't pause to wonder whether the source data made sense. It just extrapolates from what it was given, faster than any human could sanity-check it.

That's the part that should worry more people than it does. A recent survey of UK business leaders found 72% are worried that outdated data may be used to make vital decisions, and 70% are worried that a lack of data oversight is already letting biased insights shape the calls being made in their business. Those aren't hypothetical fears about some future AI rollout. They're describing systems already running.

Most of what's being fed in was never designed to be read by a machine

Part of the problem is structural. The same research puts the share of organisational data that's unstructured (emails, PDFs, scanned forms, free-text fields, call notes) at somewhere between 80 and 90%. Most of it was created for a human to glance at once, not for a model to learn patterns from. It's inconsistent by nature: different people filling in the same field differently, different systems logging the same event in different formats, years of accumulated shortcuts and workarounds baked into how the data actually looks.

Feed that into a model and you're not just risking wrong answers. You're risking wrong answers that look exactly as polished and certain as the right ones. That's the real shift AI brings to GIGO: it used to produce visibly broken output. Now it produces output that reads as authoritative regardless of whether it deserves to be.

"Fix the data" is not a satisfying answer, but it's the right one

Nobody wants to hear that the unglamorous, years-long slog of cleaning, structuring and governing data has to happen before the exciting part starts. It's much more appealing to believe that a good enough model can compensate for a messy foundation. It can't. Every attempt to paper over bad data with better modelling just moves the error further downstream, to a point where it's harder to trace and more expensive to unwind.

The organisations getting real value out of AI right now are, disproportionately, the ones that treated data quality as the actual project, not a checkbox before the "real" work of deploying a model. That means knowing where data comes from, whether it's current, whether it's consistent, and whether anyone would notice if it quietly went wrong. Which is a less exciting sentence than "we're rolling out AI," but it's the sentence that determines whether the rollout does anything useful.

The uncomfortable bit

If the data feeding your models is the same data your analysts already didn't fully trust, adding AI doesn't neutralise that problem. It industrialises it. The question worth asking before the next AI initiative gets greenlit isn't "what can this model do." It's "what is this model actually being trained and run on, and would I have trusted that data to make this decision without it?"

If the honest answer is no, the model won't change that. It'll just make the wrong answer arrive faster, and dressed more convincingly.

This is the second in a short series on why more data isn't always better data. Next: why the same word means three different things in three different departments, and why that's quietly wrecking every integration project it touches.