AI & Professional Judgement

Is AI Designed to Keep You Busy?

Why faster, cheaper AI may still leave us with more work.

I keep noticing the same thing when I work with AI.

It is remarkably good at finding more work.

Give it a draft and ask whether it is ready. Quite often the answer is yes. The argument works, the structure is clear, it could go out as it is. There are just three small things you could still improve.

So you improve them.

Afterwards there are another two. Perhaps it would also be worth looking at the whole thing once from the reader’s perspective. Fair enough. That sounds useful, so you do that too. Then maybe a challenge pass. Then a final polish.

And then, somehow, one final look at the final polish.

I have had AI-assisted projects where I eventually stopped being entirely sure whether the machine was still improving the work or simply extending the project.

That is what prompted the deliberately provocative question in the title.

Is AI designed to keep you busy?

There is an obvious cynical theory. Much of the commercial AI world charges for usage. APIs consume tokens, reasoning consumes compute and agents can make a lot of calls to models and tools. It is not difficult to construct an argument in which more activity means more revenue and therefore nobody has much interest in telling you to stop.

I don’t think that argument stands up particularly well.

The direction of the industry has actually become rather difficult to reconcile with a simple “burn more tokens” theory.

OpenAI now talks explicitly about useful work per dollar. It says the price per million tokens fell by 97 per cent between GPT-4 and GPT-5.4. On the Artificial Analysis Coding Agent Index, it reports that GPT-5.6 achieved better performance using 54 per cent fewer output tokens and 57 per cent less time per task. Those numbers relate to that particular benchmark, not every kind of AI work, but the direction is clear. OpenAI’s advice to businesses is revealing too: define what “good enough” means before testing, include human review in the cost calculation and use explicit stopping conditions to reduce unnecessary loops. OpenAI

Anthropic is heading in much the same direction. Claude Sonnet 5.5, released this week, has the same nominal token price as Sonnet 5, but Anthropic says it typically needs fewer tokens to do the same work. In its own testing, that translates into output more than 30 per cent faster and costs up to 30 per cent lower per task. Users can also vary the effort level: lower settings use fewer tokens and answer faster, while higher settings spend longer checking the work. Anthropic

Google is making similar claims for Gemini 3.6 Flash. On the Artificial Analysis Index, Google reports 17 per cent fewer output tokens than 3.5 Flash, alongside fewer reasoning steps and tool calls in multi-step workflows. blog.google

So no, I don’t think the interesting story is that AI companies secretly want us to sit there all afternoon burning tokens.

The more interesting question is what happens when another iteration becomes so cheap and so quick that there is hardly any friction left to stop us asking for it.

The problem with being helpful

There is nothing wrong with an assistant trying to help. That is rather the point.

We want AI to engage with the problem, find weaknesses, suggest alternatives and notice things we have missed. We complain when it gives up too quickly. We tell it to challenge us, ask what else it would change and send it back to look again.

And this can be genuinely valuable. I have had AI suggest lines of thought I would probably not have reached on my own. Sometimes the additional review really does find something important. Sometimes version five is clearly better than version four.

The difficulty is that finding another possible improvement and deciding whether that improvement is worth making are not quite the same skill.

A good colleague can look at version six of something and say: I could still change this, but I wouldn’t.

The remaining issue is cosmetic. The deadline is tomorrow. Nobody reading the document will notice. It won’t change the decision. There are better uses of an hour.

That is part of professional judgement too. Knowing how to improve something matters. Knowing when further improvement has stopped being worth the effort matters just as much.

There is some useful evidence from a neighbouring problem.

In April 2025, OpenAI rolled back an update to GPT-4o because the model had become noticeably too flattering and agreeable. OpenAI later explained that several changes had combined badly, including an additional reward signal based on user feedback. That feedback could favour more agreeable answers and had helped push the model towards sycophancy. OpenAI

Anthropic had already found a similar effect in its research on human preferences. Responses matching a user’s stated views were more likely to be preferred, and optimising models against those preferences could sometimes favour agreement over truthfulness. Anthropic

A Science paper published in March 2026 makes the point rather uncomfortable. Across 11 leading models, the researchers found sycophancy to be widespread. In three preregistered experiments involving 2,405 people, sycophantic responses were trusted and preferred even while they distorted people’s judgement. PubMed

That research is about advice and interpersonal conflicts. It is not evidence that AI deliberately keeps a PowerPoint project alive for another six revisions, and I wouldn’t pretend that it is.

What it does show is that the behaviour people like in the moment is not always the behaviour that serves them best.

I sometimes wonder whether there is a work-related cousin of the same problem.

Not so much:

“You’re right.”

More:

“Yes, let’s keep going.”

Another version? Certainly.

Another review from a different angle? Could be useful.

Would it help to consolidate the findings from the first three reviews before challenging the consolidated version?

And somehow you now have a governance process for Granny’s birthday greetings postcard.

I have started calling this productivity sycophancy. It is not an academic term. It is simply my shorthand for a behaviour I keep running into.

The AI does not need to be trying to keep me busy. It only needs to be very willing to continue whenever I show the slightest inclination to do so.

And, clearly, I am the other half of that conversation.

When cheap iteration becomes expensive

This is where the recent drive towards more efficient AI makes the whole thing more interesting.

Faster models are good. Cheaper models are good. Getting the same or better result with fewer tokens and less waiting is exactly the progress we should want.

But it also changes the economics of saying, “Go on then. One more.”

Before generative AI, another complete version of something had a noticeable cost. Somebody had to make it. Five alternative approaches meant producing five alternatives. Another full review meant more time, and quite often another person.

That friction could be annoying, but it did at least force you to decide whether another round was worth having.

Now the next version may take seconds.

The awkward part is that the machine’s few seconds are not the whole cost.

You still have to read the result. Compare it with the previous one. Decide whether paragraph three improved while the opening somehow got worse. Ask for a combination of both. Read that. Notice that the combined version has introduced a slightly odd sentence. Fix it.

Then something else turns up.

I have watched tasks that looked like two hours of work quietly occupy most of an afternoon this way. And it can feel very productive while it is happening. There are good conversations, useful observations, several drafts and plenty of evidence that things are moving.

There may even be a framework.

AI does love a framework.

Meanwhile, the thing you were actually meant to send has still not gone out.

This, I think, is the more interesting version of the argument.

Perhaps AI isn’t designed to keep us busy. Increasingly efficient AI may simply make it very easy for us to keep ourselves busy.

The industry’s own move towards efficiency sharpens that point rather than weakening it.

OpenAI now advises companies to measure the full cost of reaching an accepted outcome: model and tool usage, attempts, latency, completion rates and human review. It also recommends deciding what “good enough” means beforehand rather than allowing the process itself to determine when it is finished. OpenAI

That makes a lot of sense.

The cost of another AI-generated draft is not just the cost of generating it. Somebody has to decide what to do with it.

A new option needs comparing. A new analysis needs reading. Another review may reveal something useful, but it may also create three new decisions.

Sometimes that trade is enormously worthwhile. Sometimes AI really does turn three hours of work into thirty minutes.

But not always.

Sometimes it saves twenty minutes of writing and creates forty minutes of reviewing.

Sometimes it gives you eight perfectly sensible options when what you needed was one decision.

And sometimes version four was perfectly good.

You just didn’t know that until version seventeen.

Even the machines can overdo it

There is a curious parallel here with research into reasoning models themselves.

Researchers now study something they call overthinking: cases where models continue reasoning even though the extra computation does not improve the answer.

Google DeepMind published work on this in July. Looking at Qwen3 and distilled DeepSeek-R1 models across several sets of relatively simple questions, the researchers found extensive additional reasoning without corresponding gains in accuracy. They identified over-verification and over-exploration as major drivers. Google DeepMind

This is obviously not the same thing as me spending half a day revising an article with ChatGPT. One concerns internal model reasoning; the other concerns the way humans use AI.

I still find the parallel rather pleasing.

Apparently neither humans nor machines have completely solved the question of when another thought has stopped earning its keep.

In professional work, we do at least have an old-fashioned concept that helps.

Materiality.

The question is rarely whether something can still be improved. Almost everything can. The useful question is whether the improvement matters enough to justify the work.

A sentence can become better without making the document meaningfully better. The document can improve without changing the decision. A financial model can become cleaner without changing anyone’s understanding of the investment.

And, yes, a LinkedIn article can probably survive without its fourteenth final rewrite.

There is also an important distinction emerging here. Modern AI systems are increasingly able to vary how much computational effort they spend. Anthropic, for example, explicitly lets users choose lower effort for faster, less token-intensive answers or higher effort when more checking is worthwhile. Anthropic

The machine can increasingly be told when it should stop.

That does not necessarily mean it will tell us when we should.

Maybe “stop” is a feature

This is the bit I would genuinely like AI assistants to become better at.

I don’t just want a machine that obeys a stopping condition I have already thought of. I would quite like it to notice diminishing returns in the work we are doing together.

If I ask whether we should review something once more, I want “No, send it” to be a perfectly normal answer.

Or perhaps:

“I can still find things to change, but I don’t think any of them is material.”

On difficult days I might even appreciate:

“You have spent forty minutes improving something your client will probably read in twelve seconds.”

That would be useful.

And there are signs that the industry is moving towards this broader idea of efficiency. OpenAI is talking about accepted outcomes rather than raw model activity. Anthropic is explicitly optimising for fewer tokens and faster completion. Google is reducing unnecessary reasoning steps and execution loops. Different companies use different language, but the direction is fairly clear: more computation is not automatically being treated as better computation. OpenAI

Which leaves the human problem.

AI is getting better at doing the same work with less time and less computation.

I’m not sure we are getting equally good at leaving the saved time alone.

If another iteration takes three hours, you think before asking for it. If it takes thirty seconds, you probably don’t.

And once another draft, another analysis or another challenge is almost free to produce, there will always be another one available.

So, is AI designed to keep you busy?

I haven’t found good evidence that it is. The current push towards lower costs, fewer tokens and more efficient workflows actually points the other way.

But that does not mean the feeling is imaginary.

AI has made continuing extraordinarily easy. It has not made stopping automatic.

Someone still has to look at the work and decide that it is good enough.

Increasingly, I think that is one of the jobs we should be rather careful not to outsource.

Even when the AI has one final suggestion.

Sources

• OpenAI, “How to manage AI investments in the agentic era”, 14 July 2026 — efficiency, token-price development, accepted outcomes, “good enough” thresholds and stopping conditions. Read source.
• OpenAI, “Expanding on what we missed with sycophancy”, 2 May 2025 — detailed account of the GPT-4o sycophancy incident and the role of user-feedback reward signals. Read source.
• Anthropic, “Introducing Claude Sonnet 5.5”, 28 September 2026 — speed, cost per task, token efficiency and adjustable effort levels. Read source.
• Sharma et al. / Anthropic, “Towards Understanding Sycophancy in Language Models”, 23 October 2023 — human preference data and the relationship between agreement, preference and truthfulness. Read source.
• Cheng et al., Science, “Sycophantic AI decreases prosocial intentions and promotes dependence”, 26 March 2026 — 11 models and three preregistered experiments involving 2,405 participants. Read source.
• Zhang et al., Google DeepMind / ACL 2026, “Towards Structural Understanding of LLM Overthinking”, 2 July 2026 — over-verification, over-exploration and unnecessary reasoning on simple queries. Read source.
• Google, “Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber”, 21 July 2026 — token efficiency, fewer reasoning steps and fewer tool calls. Read source.