Thomas Carlyle wrote, “Man is a Tool-using Animal. Weak in himself, and of small stature… without Tools he is nothing, with Tools he is all.” The sentence appears in Sartor Resartus, published in 1831, where it is spoken by Carlyle’s invented Professor Teufelsdröckh. Carlyle’s first example is not a hammer but clothing. The tool is what lets a limited creature act beyond its unaided body.
That sounds like a celebration of technology. It also sounds like a warning. Turn Carlyle’s sentence toward the present and it becomes: the human is an AI-using animal. Without AI he is nothing, with AI he is all. The modern version captures a real expansion of capability. It may also be a trap, especially when the tool speaks fluently enough to hide where it is weak.
The task and your judgment decide whether the tool makes you all or nothing.
The promise of becoming all
AI can make an ordinary workday look transformed. In a preregistered randomized field experiment, Dell’Acqua and co-authors studied 758 consultants at Boston Consulting Group. The consultants were assigned to three groups: no AI, GPT-4, or GPT-4 plus a short prompt-engineering session. They worked on two separate sets of tasks, which matters because the result was not one simple verdict about artificial intelligence.
Inside the frontier, 385 consultants handled 18 realistic tasks. The AI groups completed 12.2% more tasks, worked 25.1% faster, and produced 29.9% to 33.9% higher quality output. For a planner, analyst, or consultant, that could mean finishing a four hour task in three, which is a working day a week returned to the calendar. The tool can expand what a person can do, and improve work that already fits its strengths.
That is the challenge. Once a tool repeatedly helps, confidence starts to migrate from the result to the tool. Faster work feels like better judgment. A polished explanation feels like understanding. The user begins to ask what the system can produce, rather than what the situation requires.
Where the frontier turns jagged
The same experiment included 373 consultants outside the frontier. They faced one business task with an objectively correct answer. Consultants without AI got it right 84.5% of the time. With GPT-4, they got it right 70.6% of the time. With GPT-4 plus training, the figure fell to 60%, a fall of 19 percentage points across the two AI groups.
The contrast is uncomfortable because the tasks may not have felt radically different. The researchers call this the jagged technological frontier. Tasks that look equally hard to a person can sit on opposite sides of it, and the frontier cannot be deduced from how difficult the task feels. AI can be excellent at one activity and unreliable at another that appears similar in complexity.
Consider one of the 373 consultants preparing for a client meeting. The answer arrives quickly, in confident language, with enough reasoning to seem complete. If the task sits outside the frontier, those useful features can make the mistake harder to notice. The consultant may carry a wrong answer into the room because speed and fluency have been mistaken for evidence.
The paper’s own correction is important. The widely quoted “40% quality” figure came from the 2023 working draft and was removed from the peer-reviewed abstract. The smaller, peer-reviewed numbers are enough to make the point. AI produces gains on some work and serious losses on other work. A responsible account must hold both results at once.
The price of not knowing
There is another danger beyond getting an answer wrong: losing contact with the thinking that should have tested it. An MIT Media Lab EEG study by Kosmyna and co-authors in 2025 put people through three essay-writing sessions under three conditions: an LLM assistant, a search engine, and no tools. The LLM group showed the weakest neural connectivity and the weakest sense of ownership of what they had written. The researchers call the effect cognitive debt.
Cognitive debt is useful because the cost may arrive later. A person can submit a fluent essay today while retaining less of its structure and reasoning. In operational work, a planner may accept a recommendation without building the mental model needed to explain it when conditions change. The tool has reduced immediate effort, but it may also reduce the user’s ability to inspect, adapt, and defend the result.
The struggle is not between humans and tools. It is between two ways of using a tool. In one, the human remains the animal that chooses, checks, and learns. In the other, the human becomes the tool’s customer, paying with attention and judgment for convenient answers.
Choose your side
The transformation is modest but demanding. Before using AI, identify the task’s risk, the kind of correctness it requires, and the parts you can independently verify. Use the tool where its strengths are visible. Slow down where a wrong answer could travel into a decision, forecast, design, or client conversation. Keep enough reasoning in your own head to notice when the output is strange.
Before your next AI-assisted decision, ask which side of the frontier the task is on and what would prove the answer wrong. That turns AI from an authority into an instrument. Humans are not powerful because tools erase their limits. They are powerful because they can choose a tool, understand its limits, and remain responsible for what happens next.
The human is an AI-using animal. Without AI, he is not nothing. With AI, he is not all. He becomes more capable, or less capable, according to the task and the judgment he brings to the tool.