Signal and Noise: Volume Three
This is the third in an occasional series called Signal and Noise - essays on the principles that have guided 46 years of building software, and why they matter more in the age of AI.
The maxim
Garbage in, garbage out.
Six syllables. First recorded in print in 1957. Possibly the oldest axiom in computing. So familiar it has become invisible - one of those truths people acknowledge without thinking about, like "measure twice, cut once" or "look before you leap."
I have been saying it for as long as I have been writing software - forty-six years now. Clients nod. They agree. And then they feed garbage into their systems and act surprised when garbage comes out.
The difference now is that the garbage has learned to dress well.
1957 versus 2026
When the phrase was coined, the problem was relatively contained. A programmer punched bad data into a card. The computer produced a wrong answer. The wrong answer was obviously wrong - a payroll figure with too many zeroes, a trajectory calculation that pointed at the ground. You could spot it. You could fix it.
The computers of 1957 were honest about their limitations. They did not pretend. Feed them nonsense and they produced nonsense, and the nonsense looked like nonsense. There was a transparency to the failure.
That transparency is gone.
Modern AI - large language models in particular - has a property that makes GIGO exponentially more dangerous than it was in the punch-card era. It does not just process bad input. It launders it. It takes incomplete, biased, contradictory, or simply wrong data and produces output that reads like it was written by a thoughtful analyst who spent a week on the research. The formatting is clean. The paragraphs flow. The conclusions are stated with quiet confidence.
The garbage is still garbage. But now it is wearing a suit.
The laundering problem
This is what I call the laundering problem, and it is the single most underappreciated risk in the current wave of AI adoption.
Here is how it works. A company has customer data. The data is incomplete - 40% of records have no industry classification, half the contact information is three years out of date, the notes fields are a mixture of shorthand, speculation, and copy-pasted email threads. Everyone in the company knows the data is unreliable. They work around it. They have learned not to trust the CRM at face value.
Then someone plugs that data into an AI tool and asks for a market segmentation analysis.
The AI produces a beautifully structured report. Four customer segments, clearly defined, with buying patterns, churn risk indicators, and recommended strategies for each. The language is precise. The framework is coherent. The output looks like something a Big Four consultancy would charge six figures for.
Except it is built on data that the company's own sales team does not trust to be accurate.
The AI has not cleaned the data. It has not flagged the gaps. It has not said "41% of your customer records lack industry classification, so any segmentation I produce will be unreliable." It has simply done its best with what it was given and presented the result with the same confidence it would apply if the data were perfect.
This is not a hypothetical. I have seen versions of this in real engagements. A financial services firm asked an AI assistant to summarise client risk profiles based on internal notes. The notes were inconsistent - different advisors used different terminology, some were years old, some contradicted each other. The AI produced clean, authoritative-sounding summaries. The summaries were presented to a compliance review. Nobody checked them against source material until an auditor flagged discrepancies six weeks later.
The damage was not that the AI made things up. It is that it made unreliable data look reliable - and people downstream treated it accordingly.
Confident garbage at scale
The problem compounds with scale. One bad analysis from a human analyst is a contained risk. It sits in a report. Someone reads it, maybe challenges it, maybe asks for the source data. The chain of reasoning is traceable.
AI-generated analysis at scale is a different proposition entirely. When you can produce a hundred market reports, a thousand customer summaries, ten thousand risk assessments - all in the time it used to take to produce one - the volume of confident garbage can overwhelm any organisation's capacity to verify it.
Consider what happened when a New York law firm submitted a brief to court containing six fabricated case citations, all generated by ChatGPT. The cases did not exist. The citations looked perfect - correct formatting, plausible names, reasonable-sounding legal reasoning. The lawyers trusted the output because it looked like output they would trust from a junior associate. The judge did not trust it, but only because opposing counsel checked.
That is the visible version of the problem. The invisible version is worse: AI-generated analysis that is not fabricated but is built on bad foundations, and nobody checks because the output looks too polished to question.
McKinsey reported in 2024 that organisations using generative AI saw a 40% increase in the volume of analytical outputs. What they did not report - and what I suspect they could not measure - is whether the quality of inputs kept pace. If it did not, that 40% increase is not productivity. It is 40% more confident garbage circulating through decision-making pipelines.
The requirements connection
There is a deeper truth here that goes beyond data quality into requirements quality.
Einstein - or at least the version of Einstein that lives in management consultancy slide decks - said that if he had an hour to solve a problem, he would spend fifty-five minutes defining the problem and five minutes solving it. Whether Einstein actually said this is beside the point. The principle is sound and it is directly connected to GIGO.
The quality of an AI system's output is determined by three things: the quality of the training data, the quality of the input data, and the quality of the question you ask. Two of those three are entirely within the organisation's control. The third - training data - is rapidly becoming a regulatory concern, but even today, the most common failure mode is not bad training data. It is bad input data and bad questions.
"Analyse our customer data and give us insights" is a garbage question. It will produce garbage insights. Not because the AI is stupid, but because the question is too vague to produce anything meaningful. The AI will find patterns - it always finds patterns - but whether those patterns matter, whether they reflect reality, whether they are actionable: that depends on whether you defined the problem before you asked the machine to solve it.
This is the same discipline that has driven effective systems delivery for decades. Understand the problem. Define the requirements. Specify what "good" looks like. Then - and only then - build the system. The technology has changed. The discipline has not.
Why this is more dangerous now than it has ever been
In 1957, GIGO was a technical problem. Bad data produced bad output. The fix was better data.
In 2026, GIGO is an organisational risk. Bad data produces authoritative-looking output that influences decisions, shapes strategy, and allocates resources. The fix is not just better data - it is a fundamental change in how organisations think about the relationship between input quality and output trust.
Three things make the current moment uniquely dangerous:
Speed. AI produces output fast enough that there is no natural pause for human review. When a report takes a week to produce, someone reads it carefully. When it takes thirty seconds, it gets skimmed, forwarded, and acted upon.
Volume. The ability to generate analysis at scale means organisations are drowning in output. More output does not mean more insight. It often means more noise - and noise that looks like signal is worse than silence.
Trust transfer. People have learned to trust well-formatted, well-written text. Decades of reading reports from analysts, consultants, and experts have trained us to associate polished prose with reliable thinking. AI exploits that trust transfer mercilessly. The prose is polished. The thinking may not be there at all.
What to do about it
The solution is not to avoid AI. The solution is to treat input quality as the most important variable in any AI deployment - more important than model selection, more important than prompt engineering, more important than the user interface.
This means:
Audit your data before you automate it. If your CRM is a mess, no AI tool will unmess it. It will just produce confident-sounding analysis of your mess. Clean the data first or be honest about the limitations of any output.
Define the question before you ask the machine. Spend the fifty-five minutes. What specifically are you trying to learn? What would a good answer look like? What data would you need to answer this question reliably? If the data does not exist, the answer does not exist - no matter how convincing the AI's attempt looks.
Build verification into the pipeline. Every AI output should come with a provenance chain. What data went in? How complete was it? What assumptions were made? If the system cannot tell you where its conclusions came from, you should not trust its conclusions.
Treat confidence as a warning sign, not a reassurance. The more confident an AI output sounds, the more carefully you should check it. Genuine expertise comes with caveats. AI-generated analysis almost never does.
Distil before you decide
This is the principle behind Rubicon Hopper, the intake engine we have built at Rubicon. Hopper does not try to do everything with every piece of data it receives. It distils first. It classifies, extracts what matters, identifies what is missing, and presents a clean, verified picture before any downstream system - human or AI - makes a decision based on it.
The philosophy is simple: the most valuable thing you can do with incoming information is make sure it is worth acting on before you act on it. That means checking sources, flagging gaps, and being honest about confidence levels. It means treating data quality not as a hygiene task that someone else handles, but as the foundational discipline on which everything else depends.
Garbage in, garbage out. The oldest truth in computing. More dangerous now than at any point in its sixty-nine-year history.
The organisations that will get value from AI are not the ones with the best models or the biggest budgets. They are the ones that take the time to ensure what goes in is worth the analysis that comes out.
Everything else is just well-dressed garbage.