For three years I have been sitting in meetings where someone says: AI is just a tool, the real value comes from the human who uses it.
Everybody nods. I nodded too, till 2024. They are still nodding.
Well, let’s correct.
There are tasks where the human does not add value to the AI’s work. The human takes value away. And it is time we had a name for that.
Negative value-add
You know added value. We write it into job descriptions, proposals, performance reviews. It is the reason we give for being in the room.
Economists have the opposite too: negative value-add. That is when a company’s output is worth less than what went in. You buy good wood, you work on it for a week, and the chair is worth less than the wood was.
Now put a person and an AI into the same picture.
Negative value-add is when the result with you is worse than the result without you. I do not mean you were slow. I do not mean you were unnecessary. I mean the thing got worse because you touched it.
MIT counted it
Do not take my word for it.
In 2024 three MIT researchers, Vaccaro, Almaatouq and Malone, published a meta-analysis in Nature Human Behaviour. They took 106 experiments and 370 results where the same task was done by a human alone, by an AI alone, and by the two together.
On average, the human and the AI together did worse than the better of the two alone.
Yes, you read that right.
We all expected 1+1=3. It did not even reach the bigger 1.
Before you forward this to your team with three exclamation marks, here are the fair details:
- Human plus AI usually beats the human alone. So AI helps you. The problem shows up when you compare the pair to the AI working alone.
- The losses were in decision tasks. Spotting deepfakes, forecasting demand, diagnosing medical cases. The AI gave an answer, the human had the last word, and the last word made it worse.
- Where the AI was better at the task than the human, adding the human pulled the result down. Where the human was better, adding AI helped.
- Creative work was the exception. Writing, making images - there the pair did well. MIT’s own summary says the same.
One more thing. The experiments in that analysis end in the middle of 2023. Think about what the models could do in 2023, and what they do now.
Read the manual (for once)
When did you last read a manual? Exactly.
So I did it for you, because the companies who build these models wrote it all down.
Anthropic’s official prompting guide tells you to give general instructions and skip the detailed list of steps. A plain “think thoroughly” often gets better reasoning than the step-by-step plan you wrote so proudly. And then comes my favorite sentence in any software documentation ever: “Claude’s reasoning frequently exceeds what a human would prescribe.”
They put that in the manual. Politely.
The same guide asks you to stop over-prompting, and to stop shouting at the model in capital letters.
OpenAI’s guide for its reasoning models says the same with other words. Keep the prompt simple and direct. Do not tell the model to think step by step, it already does that on its own. What they want from you is something else: say very clearly what the end goal is and what counts as success.
It is like a taxi. You say the address. You do not explain to the driver when to change gear.
And it goes beyond prompts. Anthropic also looked at how people actually work with its coding agent:
- Inside Anthropic, the success rate on the hardest tasks doubled within a few months. In the same months, human interventions per session went down from 5.4 to 3.3.
- New users let the agent run without approving every step in about 20% of sessions. Experienced users in more than 40%.
- And experienced users interrupt more often. They stopped clicking approve on everything and started watching for the moments where it matters.
What do they say about a human approving every single action? That it “will create friction without necessarily producing safety benefits.” In another paper they admit what we all know from cookie banners: ask people for approval all the time and they stop paying attention.
To be fair, neither company says throw the human out. The same Anthropic guide wants a person to approve anything that moves money, and tells the model to think about what can be undone. So the manual does not say zero humans. It says fewer touches, at the points where a human is really needed.
So how do we make it worse?
Mostly with habits we were praised for our whole career.
We overrule the right answer
The model says A. Your gut says B. You have 20 years of experience, so B it is.
Sometimes your gut is right. But when did you last go back and check who was right, you or the model? I’m guessing never.
We edit so that it becomes ours
You know the document that goes through three managers. Each one changes a few words, because approving something without touching it feels like you did not work.
Did the text get better? No. It got three more owners.
We are the queue
The AI is done in four minutes. Then the result waits four days until the boss has time to look at it. And if three people have to agree on the changes, there will be a meeting too.
The model does not need a meeting with itself.
We check things we cannot check
“Human in the loop” sounds so safe. But a tired person looking at the 60th output of the day is not quality control. They approve everything, or they change something just to feel useful.
Which one is better? Neither.
So yes, AI will take jobs
The comforting sentence of the last three years was: AI will not take your job, a person using AI will.
Nice sentence. It is getting old.
If the person with AI produces a worse result than the AI alone, why would a company pay for the person in that task? Would you pay extra to get less?
So yes, AI will take jobs. First where our contribution is around zero. Then where it is below zero. We can pretend it is not happening, and then we will have this conversation in a redundancy meeting.
And I say this as someone who is considered extremely pro-human, and who has spent the last ten years developing people. I am not happy about it. I just prefer counting to hoping.
The good news about a minus sign: it shows you where to stop. Take the human hours out of there and put them where we are still a plus. Deciding what we want. Saying what good looks like. Taking responsibility for the result. Dealing with other humans, who are upset, confused or brilliant, sometimes all three before lunch.
What to do next
- Measure three versions. Most companies compare “with AI” to “before AI”. Add the third one: AI alone. Until you do, you have no idea whether you are a plus or a minus.
- Check your own corrections. Take the last 20 times you changed an AI output. Better, or just more yours?
- Move people to the two ends. At the start: what do we want? At the end: did we get it, and who answers for it? In the middle, rewriting sentences and dictating steps - that is where we do the damage.
- Stop selling “human in the loop” as a quality stamp. It is a cost. Sometimes worth paying, for legal, ethical or trust reasons. Then call it that.
- Train people for the tasks where they still win. The MIT numbers show it: where the human is the stronger one, adding AI helps. Find those tasks in every role and invest there.
Added value is not something you have because you are human. It is something you can measure.
So measure it. Before your CFO does :)
Sources
- When combinations of humans and AI are useful: A systematic review and meta-analysis - Vaccaro, Almaatouq, Malone, Nature Human Behaviour, 2024
- Humans and AI: Do they work better together or alone? - MIT Sloan
- Prompting best practices - Anthropic documentation
- Reasoning best practices - OpenAI documentation
- Measuring AI agent autonomy in practice - Anthropic, February 2026
- Trustworthy agents in practice - Anthropic, April 2026