The promise was extraordinary. Artificial intelligence would elevate human output, sharpen our thinking, accelerate discovery. Instead, according to a growing body of research, AI is doing something far more mundane and potentially more consequential: it’s turning in work that’s just good enough to pass.
A study from MIT researchers, reported by Fortune, found that AI systems overwhelmingly produce what the researchers termed “minimally sufficient work” — output that meets the basic requirements of a task without exceeding them in meaningful ways. Not wrong. Not brilliant. Just adequate. The kind of work that earns a B-minus and moves on.
That finding should unsettle anyone who has bet their company’s strategy on generative AI as a force multiplier for quality. And a lot of people have made exactly that bet.
The MIT research examined outputs across a range of tasks — writing, analysis, coding, summarization — and consistently found the same pattern. AI-generated work satisfied surface-level criteria. It hit the right
This isn’t a fringe observation anymore. It tracks with what practitioners across industries have been saying quietly for months — and in some cases, loudly. Software engineers report that AI-generated code works but often introduces subtle inefficiencies or ignores edge cases that an experienced developer would catch instinctively. Lawyers say AI-drafted briefs read well on first pass but collapse under adversarial scrutiny. Marketing teams find that AI copy is publishable but forgettable. The pattern is consistent. Functional but flat.
So what’s actually happening inside these models?
Large language models are, at their core, prediction engines. They generate the most statistically probable next token based on training data. That architecture has a built-in bias toward the average. The most likely next word in a sentence is, by definition, the most common one. The most likely structure for a paragraph is the one that appears most frequently in the training corpus. This means that LLMs are architecturally inclined to produce output that resembles the median of their training data — not the top percentile, not the bottom, but the thick middle of the distribution curve.
And the median of human-produced text on the internet is, to put it plainly, not great.
The implications extend well beyond individual task quality. When organizations deploy AI at scale to handle knowledge work — drafting reports, answering customer inquiries, generating analyses — they’re effectively standardizing their output at that B-minus level. Every document starts to sound the same. Every analysis follows the same template. The variance disappears, and with it, the occasional flash of genuine insight that comes from a human expert having a particularly good day or an unusual idea.
There’s a counterargument, of course. Proponents of AI-assisted work point out that minimally sufficient output, delivered at massive scale and near-zero marginal cost, can still represent an enormous productivity gain. If a consultant previously spent eight hours producing a B-plus report, and AI can produce a B-minus draft in thirty seconds that the consultant then polishes for two hours, the net effect is positive. The floor has been raised. The median worker becomes more productive.
True. But that framing obscures a deeper problem.
The MIT findings suggest that when humans rely on AI-generated drafts as starting points, they tend to anchor on the AI’s output rather than substantially reworking it. The B-minus draft doesn’t get polished to an A. It gets lightly edited and shipped as a B. Over time, the human’s own standards drift downward to match the machine’s. The researchers observed this anchoring effect across multiple experimental conditions, and it proved remarkably persistent even when participants were explicitly warned about it.
This is the real risk. Not that AI produces bad work — it doesn’t. It produces adequate work. And adequate work, delivered frictionlessly and at scale, has a way of becoming the new standard. The gravitational pull of “good enough” is powerful, especially in organizations under pressure to move fast and cut costs.
Consider what this means for industries where quality differentials matter enormously. In financial analysis, the difference between a competent summary and a genuinely insightful one can be worth millions in investment decisions. In legal work, a brief that’s merely adequate might lose a case that a brilliant one would have won. In journalism — and I say this with some self-awareness — an article that hits all the expected points but offers no original perspective is just noise in an already deafening information environment.
The technology industry itself seems to be grappling with this tension. OpenAI, Anthropic, Google, and others continue to push model capabilities forward, but the improvements increasingly show up in benchmarks rather than in the qualitative experience of using the tools. GPT-5 scores higher than GPT-4 on standardized tests. It handles longer contexts. It follows instructions more reliably. But does it produce work that a discerning reader would call genuinely excellent? The MIT research suggests the answer is mostly no — and that the gap between benchmark performance and real-world output quality is wider than the industry acknowledges.
Some companies are trying to engineer their way around this limitation. Retrieval-augmented generation, or RAG, attempts to ground AI outputs in specific, high-quality source material rather than relying solely on the model’s general training. Fine-tuning on domain-specific data aims to shift the model’s output distribution toward expert-level work rather than internet-average work. Multi-agent architectures pit AI systems against each other in review loops, hoping that adversarial pressure will push quality upward.
These approaches help at the margins. They don’t solve the fundamental problem.
The fundamental problem is that excellence in knowledge work is not primarily about information retrieval or pattern matching. It’s about judgment. It’s about knowing which details matter and which don’t, sensing when a conventional analysis misses something important, having the courage to make an argument that contradicts the consensus. These are precisely the capabilities that statistical prediction engines lack, because they are, by construction, consensus machines.
I’ve spent enough years working with technology to recognize a familiar cycle here. A new tool arrives with enormous promise. Early adopters are dazzled by what it can do. Then, gradually, the limitations become clear — not as dramatic failures but as a persistent, quiet undertow of mediocrity that’s hard to quantify and easy to ignore. The tool becomes indispensable not because it’s excellent but because it’s fast and cheap, and the organizational memory of what excellent looked like before the tool arrived slowly fades.
We saw this with PowerPoint, which didn’t make presentations worse in any obvious way but gradually flattened complex ideas into bullet points. We saw it with email, which didn’t destroy communication but made it shallower and more reactive. AI-generated knowledge work may follow the same trajectory — not a catastrophe, but a slow erosion of standards that’s invisible in any given quarter and unmistakable over a decade.
The MIT researchers, to their credit, don’t frame their findings as an indictment of AI. They frame them as a calibration. Organizations that understand what AI actually produces — minimally sufficient work, reliably and at scale — can deploy it intelligently, using it for tasks where adequacy is genuinely sufficient and reserving human expertise for tasks where it isn’t. The problem arises when organizations fail to make that distinction, or when competitive pressure makes the distinction feel like a luxury.
And competitive pressure is intense. Companies that refuse to adopt AI risk falling behind on speed and cost. Companies that adopt it uncritically risk falling behind on quality. The smart play is somewhere in the middle, but “somewhere in the middle” is exactly the kind of nuanced strategic positioning that most organizations are bad at maintaining under pressure.
There’s also a labor market dimension to this that deserves attention. If AI can reliably produce B-minus work, then the economic value of a human who also produces B-minus work drops precipitously. The premium shifts to people who can consistently produce A-level work — the kind of work that AI can’t replicate. But here’s the uncomfortable truth: most knowledge workers, most of the time, produce work that’s closer to B-minus than to A. The bell curve is the bell curve. AI doesn’t just compete with the bottom of the talent distribution. It competes with the middle.
That competition will reshape hiring, compensation, and career development in ways we’re only beginning to see. Junior roles that served as training grounds — where young professionals learned by doing adequate work under supervision — may disappear if AI can do that adequate work faster and cheaper. But those junior roles were also the pipeline through which future experts developed. Cut the pipeline, and you eventually run out of experts. It’s a classic optimization trap: extracting short-term efficiency at the cost of long-term capability.
Some firms are already reporting this effect. Law firms that aggressively automated junior associate work are finding that their mid-level associates lack the foundational skills that come from years of doing that work manually. Consulting firms that use AI to generate first drafts are discovering that their analysts have weaker analytical instincts than previous cohorts. The cause-and-effect is hard to prove definitively — these are early observations, not controlled studies — but the pattern is suggestive.
None of this means AI is useless. Far from it. For tasks where speed and consistency matter more than originality — data formatting, routine correspondence, initial literature reviews, boilerplate code — AI is genuinely transformative. It frees up human attention for higher-value work. The question is whether organizations actually redirect that freed-up attention toward higher-value work, or whether they simply reduce headcount and pocket the savings.
The evidence so far leans toward the latter.
What the MIT research ultimately reveals is a gap between narrative and reality. The narrative says AI will make everyone’s work better. The reality says AI will make everyone’s work faster, and faster is not the same as better. In many contexts, faster-and-adequate is a perfectly reasonable trade-off. In others, it’s a slow-motion disaster. Knowing the difference requires exactly the kind of judgment that AI itself cannot provide.
So here we are. The machines are productive. They’re reliable. They’re affordable. And they are, with remarkable consistency, mediocre. The organizations that thrive in the coming years won’t be the ones that adopt AI most aggressively or the ones that resist it most stubbornly. They’ll be the ones that understand, with clear eyes, what a B-minus machine can and cannot do — and build their strategies accordingly.
That understanding is, itself, A-level work. No model can do it for you.
WebProNews is an iEntry Publication