You’re Comparing AI to a Human Standard That Doesn’t Exist

Rama Ramakrishnan, professor at MIT Sloan, made a point in a 2026 MIT Sloan outlook piece that should be required reading for anyone running an enterprise AI procurement conversation. The point, paraphrased: when an LLM’s accuracy on a given enterprise task is judged “not good enough,” the comparison is almost always being made against the wrong baseline. “Rather than comparing the LLM’s accuracy to the best possible (i.e., 100%),” he wrote, “it is better to compare it to the accuracy of humans doing the work right now and to track the changing human-LLM accuracy gap for that task.”

Read that sentence twice. It is a small claim with enormous operational consequences. Most enterprise AI conversations in 2026 (in my experience and in the published surveys) implicitly assume the human comparison point is 100%. The AI gets a number. The number gets compared to perfect. The gap gets called the AI’s failure. At no point in the conversation does anyone ask whether the humans the AI is supposedly replacing or assisting actually get the answer right 100% of the time. They don’t. In most knowledge work, they don’t even come close. And the procurement decisions being made against the assumption that they do (that the human alternative is the gold standard the AI must match) are the decisions doing more damage to enterprise AI adoption than every model failure combined.

Here is the pattern. The AI got this wrong, here is the failure. Somebody points at the wrong output. The team agrees the output was wrong. The argument moves to whether the AI needs to be upgraded, replaced, retrained, or shelved. At no point does anyone ask the only question that actually matters: what would the human alternative have done with this case, and how often does the human get it right? The conversation assumes the human always does. It is mathematically wrong about that, and the decisions it produces get expensive AI products shelved for being competitive with the humans they were supposed to assist.

What follows: the actual human accuracy baselines in domains where AI is being deployed, why your eval setup is probably comparing against a fictional baseline, and the audit I now run to catch the mistake before it reaches procurement.

The strongest version of “humans are the right standard”

The case for using human accuracy as the baseline for AI quality is real, and worth engaging with before I take it apart.

The argument goes like this. The point of an AI system is to do work humans currently do, so the operational target is human-comparable performance. If the AI matches or beats the humans it replaces, the system succeeds; if it doesn’t, the system fails. Human accuracy is the right benchmark because human-comparable accuracy is the actual product requirement.

This is correct, and I concede it completely. The operational target for any AI deployment is human-comparable performance in the domain being deployed against. There is no other defensible standard.

The part that doesn’t survive contact with how the standard gets applied in practice is the assumption that “human accuracy” is a single number, that it’s high, and that it’s known. None of those three are true. Human accuracy in most knowledge work is a distribution, it varies enormously by domain and task, and almost no enterprise has actually measured it for the work their AI is being benchmarked against. The procurement conversation that says “the AI is X% accurate, that’s not good enough” is making three errors in the same sentence: it’s treating human accuracy as a point estimate, it’s treating that point estimate as 100%, and it’s comparing the AI to a baseline that nobody on the team has ever actually quantified. The argument is structurally invalid before the AI’s performance is even relevant.

Stated honestly: the steelman is right about the standard and catastrophically wrong about how the standard is operationalized. Human-comparable is the goal. The thing being compared to is fictional in most enterprises doing the comparison.

What the human baselines actually look like

The empirical picture: the published literature on inter-rater agreement in domains where AI is now being deployed is broadly consistent. The exact percentages vary by task, methodology, and the difficulty of the specific cases sampled, but the order-of-magnitude ranges are well-attested across decades of research.

Radiology. The inter-radiologist agreement literature (across mammography, chest X-ray, and CT interpretation) has consistently reported agreement on the primary finding in the rough range of 70-85%, depending on the task. On more subjective judgments such as staging, differential diagnosis between similar conditions, and BI-RADS scoring, the literature places agreement somewhat lower. The 92% AI performance numbers that radiology AI vendors cite on stage are being compared against a human baseline that the literature places below 100% by a meaningful margin.

Pathology. The inter-pathologist agreement literature reports concordance on cancer staging typically in the 75-85% range for major-staging decisions, with substantially lower agreement on fine-grained grading judgments such as Gleason scoring, where two pathologists looking at the same slide can disagree on the precise grade in a non-trivial fraction of cases even when they agree on the broader diagnosis.

Legal contract review. Studies of attorney agreement on contract risk and clause classification consistently find that experienced lawyers reviewing the same documents disagree on a meaningful fraction of flagging decisions. The legal-AI vendors that benchmark against “lawyer accuracy” are usually benchmarking against single-lawyer judgments rather than the messier inter-lawyer agreement distribution that is documented in the underlying research.

Code review. The software engineering literature has long documented that multiple senior reviewers given the same pull request flag substantially different issues. Inter-reviewer agreement on bug identification is commonly reported below the threshold most engineers would assume from how code review is treated culturally. Code review is not the gold-standard activity it is sometimes treated as in AI-for-code evaluations.

Resume screening. Studies of inter-reviewer agreement on hiring decisions show substantial variability, with the bias literature suggesting the variation is not random; it correlates with reviewer fatigue, time of day, and recent prior decisions. Resume scoring is one of the most documented examples of inter-rater unreliability in white-collar work.

Medical diagnosis at the primary-care level. Second-opinion research consistently shows that a second physician disagrees with the first physician’s diagnosis on review in a meaningful fraction of cases, depending on specialty and condition. The disagreement is not necessarily a measure of error; it is a measure of the inherent variability of medical judgment under uncertainty.

Annotation tasks more broadly. The NLP and ML literatures have decades of inter-annotator agreement research. Even on tasks treated as “subjective but tractable” (sentiment classification, for example) annotators commonly disagree on a non-trivial fraction of binary judgments, with agreement dropping considerably on fine-grained scales. Harder tasks (argument quality, factual accuracy, ethical concern in writing) document still lower agreement.

None of these numbers are catastrophic. None of them are evidence that humans can’t do the work. They are evidence that the work humans do is not the deterministic 100%-correct activity it is usually treated as in AI procurement conversations. The humans being benchmarked against are themselves operating in a distribution of “reasonable disagreement,” and the AI is being asked to outperform a point estimate that nobody has positioned within that distribution.

The category error

The move that costs enterprise AI deployments more capital than any single technical mistake I can name works like this. Somebody picks a hard task that humans currently do (diagnosis, contract review, fraud detection, code analysis) and assembles a test set of “correct” answers from a single domain expert. The single expert’s answers become the gold standard. The AI is evaluated against it and scores, say, 85%. The team then concludes “the AI is 15% worse than a human.”

This is not what the data says. The data says the AI agrees with one particular human 85% of the time. A second human, given the same test set, would also have agreed with the first human roughly 75-85% of the time, depending on the domain. The AI’s 85% is not 15% behind “human accuracy.” It is, in many cases, equal to or better than the inter-human agreement rate on the same task. The “AI is worse than a human” conclusion is comparing the AI to a human who agrees with herself 100% of the time, which describes no human in any domain.

The fix is not to lower your standards. The fix is to measure the standard you’re actually evaluating against. If your single expert is 75% in agreement with a second expert, your AI is “human-level” when it agrees with the single expert 75% of the time. If it agrees 85% of the time, it is better than the second human would have been. That is the real finding that almost no enterprise AI procurement conversation surfaces.

I do not want to overstate this. There are tasks where humans achieve very high agreement: basic arithmetic, well-defined classification with clear category boundaries, simple factual recall. In those tasks, the human baseline is genuinely near 100%, and an 85% AI is meaningfully worse. The category error is treating all tasks as if they’re in the high-agreement regime, when most knowledge work is in the moderate-agreement regime.

A baseline calibration audit

The audit I now run with any team evaluating an AI system has six binary checks. Score yourself. Most enterprises pass zero or one. The teams that pass four-plus make defensible procurement decisions; the teams below that make decisions against fictional baselines, and those decisions are almost always biased against the AI in ways that cost real money.

One. Have you measured inter-rater agreement on your test set? Take your gold-standard test set. Have a second qualified domain expert re-label it without seeing the first expert’s answers. Measure how often they agree. If the answer is “less than 90%,” your AI cannot be expected to do better, because the test itself contains that much disagreement. This single measurement reframes every quality conversation downstream.

Two. Is your gold standard a consensus of multiple experts, or one person’s judgment? A single expert’s labels are that expert’s labels. They are not “ground truth,” they are a sample of one. A defensible gold standard is built from multiple experts with documented disagreement resolution. Most enterprise gold standards are built from one person on a Tuesday.

Three. Are your accuracy targets calibrated to the human baseline in your specific domain? If radiologists agree 80% on the task, your AI’s target should be 80%, not 95%. The target should be set against the measured baseline, not against a generic “high accuracy” wish.

Four. Are disagreements recorded as “valid range” rather than “errors”? When two experts disagree, the system should distinguish “the AI is in a region where humans also disagree” from “the AI is wrong by all human standards.” These are completely different failure modes. The first is acceptable; the second is not. Most eval systems collapse them into a single “incorrect” label.

Five. Does the eval rubric distinguish “wrong” from “different reasonable answer”? Some questions have one right answer. Many do not. The eval should know which is which, and score accordingly. A reasonable alternative answer should not be coded as a failure; it should be coded as a “valid disagreement,” which is a different category of result.

Six. When you compare AI performance to humans, do you compare to a specific human alternative, not to perfection? The right comparison is “how often does the AI agree with our best human” or “how often does the AI agree with the consensus of three humans,” not “how often is the AI correct in the abstract.” Abstract correctness is not a category that exists for most knowledge work. Comparison to a specific human alternative is what an honest evaluation looks like.

Six items. Most enterprise AI evals pass zero or one. The teams that pass four or more make decisions about AI quality that hold up; the teams that don’t, make decisions that get expensive products shelved for being competitive with the humans they were procured to replace.

What this means for the work

What changes when a team internalizes that humans aren’t the gold standard, specifically:

Eval design changes. Gold standards get built from multiple expert annotations, not one. Disagreement gets recorded and analyzed, not averaged out. Accuracy targets get calibrated against the measured inter-human baseline for the specific task, not against an aspirational 95%. Faithfulness scoring distinguishes “the AI made up a fact” from “the AI gave one of several reasonable answers to an ambiguous question.”

Procurement conversations change. The question stops being “what’s the AI’s accuracy?” and becomes “what’s the AI’s accuracy relative to the human baseline we measured for this task?” The board memo says “the AI agrees with our reference experts X% of the time, and the experts agree with each other Y% of the time.” That is a complete sentence. It is also a sentence that almost no procurement memo currently contains.

Domain expert involvement changes. The SME is not the source of “the right answer.” The SME is the source of one expert’s answer, and the eval design has to acknowledge that. The benevolent-dictator pattern I argued for in the previous essay still holds (one person owns the criteria) but the criteria themselves need to include an explicit accounting for inter-expert disagreement. The SME’s role is to set the rubric, not to be the rubric.

Confidence calibration changes too, but that’s the bridge to the next essay in this set, and worth its own piece.

What this requires architecturally

The audit above is unrunnable without the right infrastructure. To answer “what’s the inter-rater agreement on our test set” you need a system that supports multi-annotator labeling. To distinguish “valid disagreement” from “AI was wrong” you need an eval pipeline that records labels alongside per-annotator metadata. To compare AI performance to a specific human alternative rather than abstract correctness, you need to capture the human alternative (what the team would have done without the AI) and that requires logging the human reviewer’s decisions alongside the AI’s on the same cases during a shadow-mode rollout.

None of this is exotic. It is basic eval infrastructure done with one specific discipline: the human side of the evaluation is captured at the same level of rigor as the AI side. Most teams capture the AI side carefully and capture the human side casually, which is exactly the asymmetry that produces fictional-baseline decisions.

On the build path, this is the same gateway-as-instrumentation argument I have made through this set. AOCore’s observability layer captures every model call, every prompt, every retrieved document, every output, and every user feedback signal, and the same architecture extends naturally to capturing human reviewer judgments during shadow rollouts. The data the audit requires is the data the gateway already produces; the only addition is per-annotator labels alongside the existing call traces. Teams that try to bolt human-baseline measurement onto a system that doesn’t already capture the AI side at this level of granularity are doing two engineering projects in parallel and usually finishing neither.

For organizations that aren’t going to staff this kind of infrastructure themselves, AODex inherits the audit-trail posture from the gateway beneath it. The platform’s quality work has been done against multi-expert evaluation rather than single-expert gold standards: the criteria for the included personas, the citation thresholds, the abstention behavior, were tuned against the kind of inter-rater calibration this essay argues for. The buying organization does not need to build the audit infrastructure to verify they’re getting human-calibrated quality; the calibration was done before the customer logged in.

The deepest lesson from three years of watching enterprise AI deployments is that the largest single source of misallocated procurement capital is the comparison between AI and a human baseline that nobody measured. The AI’s failures are visible and counted. The human’s failures are invisible and uncounted. The conversation that compares one to the other is structurally biased before any data is reviewed, and the systems that pay the price are systems that were never given a fair test.

Measure the human baseline. Compare against it honestly. The AI might still be worse, in which case you have a real finding. Or the AI might be exactly as good as the humans it replaces, while costing 80% less per case to operate, in which case you have a different real finding. The difference between the two is the difference between a procurement decision based on data and a procurement decision based on the assumption that the people in the room were the gold standard all along.

They weren’t. They never were. Nobody is. The sooner enterprise AI conversations stop pretending otherwise, the sooner the actually-good AI systems stop getting shelved for being competitive with the humans they were procured to replace.