Garbage In, Gold Out

What the unreasonable effectiveness of data in large language models means for enterprise AI strategy, and data governance. GIGO, the most durable rule in enterprise computing is now its most expensive habit — and what replaces it in the age of AI

tl;dr — In an effort to augment rigid determinism with latent space, I present an argument that the classic computing principle of “garbage in, garbage out” is an obsolete hurdle that prevents modern companies from fully utilizing artificial intelligence. Unlike older, passive software that required perfectly formatted data, contemporary large language models possess the semantic depth to interpret messy or unstructured information. By leveraging the unreasonable effectiveness of data and the emergent capabilities found in scaled models, organizations can now extract value from previously unusable sources like legacy code or free-form text. Business leaders should stop over-policing input quality and instead focus their engineering efforts on verifying outputs at the point of consequence. Ultimately, the source contends that insisting on flawless data hygiene is a costly mistake that allows more agile competitors to move faster. Those who embrace this new logic will likely dominate the future economy by turning imperfect inputs into valuable insights.

Complimentary Reading Link

“No, we can’t do this because Our Data Isn’t Clean Enough”, is quite possibly the Most Expensive Sentence in Enterprise AI.

Hot take: “Garbage in, garbage out.” maxim was coined for systems that processed punch cards in batch jobs, and it described them perfectly. Applied to modern AI, it has become the favorite intellectual cover of the most timid faction in every strategy meeting — the one that keeps usable signal locked behind imaginary cleanliness gates while competitors ship. It is among the most reliable predictors of which organizations will spend the next decade watching others lap them.

The claim I want to defend is narrow but strong: GIGO was true of a specific class of systems, that class is no longer the relevant one, and continuing to reason as if it were is a category error with a measurable price tag.

What GIGO actually described

GIGO was never a law of computation. It was a property of epistemically passive systems — systems that map inputs to outputs with no independent knowledge of the world. A regular expression does not reason. A SQL query has no world model. A rules engine has exactly the semantic depth of the engineer who wrote the rules, and that engineer left in 2014. For systems like these, the maxim potentially holds, because the system contributes nothing the input did not already contain.

Two findings from the last fifteen years dissolve that assumption for modern models, and they are the load-bearing evidence for everything that follows.

The first is what Halevy, Norvig, and Pereira, writing from Google in 2009, called the unreasonable effectiveness of data — itself a deliberate echo of the physicist Eugene Wigner’s 1960 essay on the unreasonable effectiveness of mathematics. Their observation, drawn from machine translation and web-scale learning, was that simple models trained on enormous, messy corpora reliably outperform elegant models trained on small, clean ones. The structure lives in the scale of the data, not in the tidiness of any individual record. This was the first serious crack in the GIGO frame: the field’s own evidence showed that more imperfect data beat less perfect data, repeatedly.

The second is emergence. Wei and colleagues documented in 2022 that certain capabilities — multi-step reasoning, instruction following, in-context learning — are effectively absent in small models and appear, sometimes sharply, above particular thresholds of scale. A model large enough to have internalized statistical regularities across hundreds of billions of tokens does not merely interpolate its training data; it brings priors to a new input. It can notice that an input is wrong, flag it, repair it, or ask. That is the definition of an epistemically active system, and it is precisely the property GIGO assumes away.

I will be honest about the seam here, because rigor demands it and because the steelman is more persuasive than the cartoon. Schaeffer, Miranda, and Koyejo argued in 2023 that some “emergent” abilities are partly an artifact of discontinuous evaluation metrics — measure with a smoother yardstick and the cliff becomes a slope. The right reading is not that emergence is fake; it is that we should be disciplined about the mechanism and resist treating capability as magic. The argument I am making survives the deflation cleanly, because it does not require sorcery. It requires only that a model trained at scale is no longer a passive function — and on that, scaling-law work from Kaplan and from the Chinchilla team leaves little doubt.

The failure nobody books: unattempted work

Here is the trade the GIGO mindset hides. Traditional systems failed constantly — they simply failed in familiar ways. They accepted bad data into clean schemas and propagated it with full referential integrity. They returned empty result sets no one investigated. They encoded the wrong business logic and ran it at scale for a decade before anyone noticed. We built an entire audit and compliance apparatus around those failure modes and called the apparatus “good engineering.” The failures became invisible because they were ours.

What modern systems change is that they attempt work the old pipeline simply refused. And that refusal — unattempted work — is the most expensive failure of all, precisely because it never appears on a dashboard. Consider three cases that every large enterprise is sitting on right now.

Free text and scanned documents. A customer complaint in free prose, a contract delivered as a scanned PDF, a decade of support transcripts in cold storage — the rules engine routed these to /dev/null and the OCR pipeline butchered them, so they sat unread. The same inputs are now labeled, categorized, and extracted with provenance. The traditional system’s response to all of this was a null set. The model’s is a draft.

Undocumented Legacy Codebases. Industry estimates put somewhere between 220 billion and 800 billion lines of COBOL still in production, much of it running core banking, insurance, and government systems, written by developers two retirements gone. (Both figures circulate widely — the lower from a 2017 Reuters report, the higher from a 2022 vendor survey — and both should be treated as order-of-magnitude, not gospel.) The traditional response to a critical module with no specification and no living author was manual archaeology or, more often, nothing — the code stayed frozen because no one dared touch it. A modern model produces from that source a coherent specification. Not a perfect one. A coherent and improvable one, where the baseline was zero. The honest framing is not “the AI understood the code”; it is “the AI moved the starting line from nothing to a draft an engineer can verify.” That is the entire game.

Rules engines. A Drools or JRules deployment built on Rete or PHREAK is brittle in a specific way: its behavior is bounded by what someone could enumerate in advance, and its intent evaporates the moment the authors disperse. Ten thousand rules accreted over fifteen years become a system no single person understands and no one will modify. A model can read the rules and the surrounding code, comments, and tickets, and reconstruct the intent behind the mechanism — turning a frozen liability into something a team can reason about again. The rules engine was epistemically passive by design. That was its virtue in 1995 and its trap in 2025.

In every case the input quality is identical. What changed is whether anyone could do anything with it.

The reweighting, not the abolition, of input quality

I am not claiming input quality is irrelevant. I am claiming it has been radically reweighted, and the failure to make that distinction is where most enterprise AI strategy goes wrong.

Surface noise — typos, format drift, missing fields, inconsistent schemas — has moved from blocking issue to rounding error. The model handles it; this is the direct, practical cash value of the unreasonable-effectiveness result. Demanding that every input arrive pre-sanitized is an input-hygiene tax levied uniformly across every workflow regardless of its value, and that flat tax has been the silent killer of adoption for three years.

Structural problems are entirely different and entirely real: biased training data, distributional shift, poisoned or untrusted retrieval sources, prompt injection. But notice that none of these is an input-tidiness problem. They are system-design problems, addressed with evaluation harnesses, verification loops, retrieval architecture, and human oversight placed where the stakes warrant it. Conflating “our spreadsheets are messy” with “our retrieval corpus could be adversarial” is the conceptual mistake that lets a data-governance committee spend its seventh meeting on the customer master file while the actual risks go unmodeled.

Move the verification cost to where the stakes are

This is the inversion that matters. The old paradigm spent enormous effort validating inputs and then trusted outputs by default. The new paradigm tolerates inputs and validates outputs in proportion to consequence. That is a more honest place to put the cost, because it scales with what is actually at risk. A low-stakes internal summarization does not warrant the scrutiny of a regulatory filing; GIGO could not draw that line, because it taxed the input boundary identically for both.

The two cautionary tales everyone cites make exactly this point, against themselves. In Mata v. Avianca (2023), a New York attorney submitted a brief citing cases that ChatGPT had fabricated and was sanctioned by the court. In Moffatt v. Air Canada (2024), a tribunal held the airline liable after its chatbot promised a bereavement-fare refund the airline did not honor — rejecting, memorably, the argument that the chatbot was a separate legal entity.

These are real, and instructive — but read what they actually demonstrate. The lawyer’s question was not malformed; he failed because he never checked the output. The airline’s failure was a missing verification gate on a customer-facing commitment. Neither is a refutation of the new paradigm. Both are the new paradigm telling you, precisely, where to spend your engineering: at the output boundary, at the points of consequence, where you would verify any high-stakes decision regardless of who or what produced it.

What the leaders are already doing

The organizations pulling ahead have internalized this and act on it. They deploy against the messy data they actually have rather than the clean data they wish they had. They treat evaluation as a first-class engineering discipline, run continuously, not as a launch-gate afterthought. They are extracting value from the documents, transcripts, logs, and free text that have been depreciating in cold storage for years. And they are moving while their competitors debate, in committee, whether the data is ready.

GIGO had its day. It was a useful caution for a world in which computers were dumb, deterministic, and unforgiving — which is to say, epistemically passive. We do not live in that world.

The systems we have now read context, infer intent, repair errors, and turn genuinely poor inputs into surprisingly good drafts, because scale made the data unreasonably effective and made new capabilities emerge. Treating these systems as 1960s mainframes that must be fed only spotless data is a failure of imagination dressed as professional rigor.

Garbage in, gold out is an emperical observation about what these systems do every day in the hands of teams who stopped policing the input boundary and started designing for verification at the output boundary instead.

The companies that make that shift in the next eighteen months will look up in 2028 with a lead that is hard to close. The ones still quoting GIGO will be writing the postmortems.

Author’s Note: A note on what this is: a hot take, deliberately. The strong form of the claim below — that “garbage in, garbage out” is actively destroying value — is a provocation. But it is not an ungrounded one. It rests on two of the most consequential empirical findings of the last fifteen years: the unreasonable effectiveness of data, and the emergence of capabilities at scale.

References & Further Readings

• A. Halevy, P. Norvig, and F. Pereira, “The Unreasonable Effectiveness of Data,” IEEE Intelligent Systems, 2009.
• E. Wigner, “The Unreasonable Effectiveness of Mathematics in the Natural Sciences,” Communications on Pure and Applied Mathematics, 1960. (Origin of the framing borrowed above.)
• J. Wei et al., “Emergent Abilities of Large Language Models,” Transactions on Machine Learning Research, 2022.
• R. Schaeffer, B. Miranda, and S. Koyejo, “Are Emergent Abilities of Large Language Models a Mirage?” NeurIPS, 2023. (The principal counterargument, included deliberately.)
• J. Kaplan et al., “Scaling Laws for Neural Language Models,” 2020; J. Hoffmann et al., “Training Compute-Optimal Large Language Models” (Chinchilla), 2022.
• Mata v. Avianca, Inc., S.D.N.Y., 2023 — attorney sanctioned for citing AI-fabricated cases.
• Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 2024 — airline held liable for its chatbot’s representation.
• COBOL footprint estimates: Reuters (2017, ~220B lines) and a 2022 vendor-commissioned survey (~800B lines). Treat as order-of-magnitude.