Middle to middle. Start to end.
What AI fluency means once AI starts finishing the work, and its makers talk of superintelligence.
The short version
For the first years of generative AI at work, most people used it in the middle of a task. A person decided what the work was. The AI drafted. The person judged the draft and finished the job. Middle to middle.
That is changing. AI agents can now take some tasks from start to end. They read the files, run the steps, check their own output and hand back something that looks finished. The companies that build them say their goal is superintelligence.
When the machine does the middle, the human work moves to the two ends. Before the work starts: deciding what it is, whether AI should do it, and what done looks like. After it ends: judging what came back and answering for the result. That is what AI fluency means now. This memo sets out the evidence and what to do about it.
Two words, defined
AI fluency is the ability to collaborate with AI effectively, efficiently, ethically, and safely. Memo 03 described AI literacy as five capacities. Fluency is those capacities in practice: used well, again and again, on real work.
Our courses teach fluency as four skills, the 4D framework: Delegation (deciding what to hand over), Description (saying clearly what you want), Discernment (judging what comes back) and Diligence (taking responsibility for what you do with it).
Superintelligence now has two meanings, and it helps to know which one you are reading.
- The technical meaning. AI that far outperforms the best humans at nearly all cognitive work. No such system exists. Several leading AI companies say they are building toward it. In June 2025, OpenAI's chief executive wrote that humanity is close to building digital superintelligence. Meta has a division called Superintelligence Labs. In October 2025, a public statement called for a ban on developing superintelligence until scientists broadly agree it will be done safely and controllably, and the public strongly supports it. In September 2026, a bill in the US Congress, the Ban Artificial Superintelligence Act, proposed a ban.
- The loose meaning. Since September 2026, "super intelligence" has also appeared in US political speech as a general word for today's AI.
In this memo, superintelligence means the first: a system far better than the best people at nearly all thinking work. We use the word because it is the stated destination of the companies that make the tools your staff already use.
What has changed
Three pieces of evidence, from three different sources.
Machines are taking on longer tasks
The International AI Safety Report 2026, written by a large group of independent experts, found that AI agents can now reliably complete some software tasks that would take a human programmer about half an hour. A year earlier, the figure was under ten minutes. The report calls current abilities "jagged": strong on some hard tasks, unreliable on some easy ones.
Organizations are already handing work over
Anthropic's Economic Index measures how people use its Claude models. In its June 2026 release, conversations in Denmark were 58% augmentation, where a person works through the task with the AI, and 42% automation, where the person hands the task over with little input of their own. On Anthropic's developer platform worldwide, where businesses connect AI directly to their own systems, 94% of interactions were classed as automation. Our AI Economic Index page shows the full data.
Less visible checking when the AI produces something
Anthropic's AI Fluency Index (February 2026) studied 9,830 conversations. Conversations that involved an artifact, such as a document or a piece of code, showed more direction and less visible evaluation than the rest. Identifying missing context was 5.2 points less common, fact-checking 3.7 points and questioning the AI's reasoning 3.1 points. The researchers note that some of the checking may have happened outside the chat. Either way, the checking they could see was lower where the AI produced something that looked like a result.
Where agents do the middle, the human work moves to the two ends.
The start: decide what the work is
This is Delegation and Description. Before an agent or a long AI task begins, a fluent person can say:
- what the task is for, and who will use the result;
- whether AI should do it at all, and which parts stay with people;
- what done looks like, in a sentence someone else could check;
- what the AI must not touch: which data, which systems, which decisions.
In a later extension of the Fluency Index that covered agentic tools, Anthropic found that clarifying the goal before the work starts was the behavior most closely linked with the other fluency behaviors. In the original study, only about 30% of conversations included any instruction on how the AI should work with the person. In most conversations, nobody set the terms.
The end: judge what came back
This is Discernment and Diligence. A fluent person reads AI output as if a capable stranger wrote it: useful, not trusted. They check the parts that matter, such as figures, names, dates, citations and claims about the law. They look for what is missing, not only for what is wrong. They put their name on the result only when they can defend it.
In chat, the strongest sign of fluency is iteration. In the Fluency Index, 85.7% of conversations involved refining the work through follow-ups, and those conversations were 5.6 times more likely to include questioning the AI's reasoning. Agentic tools can cut much of that back-and-forth. The checking that used to happen in the middle then has to happen on purpose, while the agent works and before the result is accepted.
Judgment is also the skill that experience does not seem to supply on its own. Anthropic found that the discernment it could observe did not grow with how long people had used the tool. That does not measure judgment directly, but it is a reason not to count on experience to build it. It has to be taught and practiced.
Why superintelligence raises the stakes
If AI systems keep improving, the human contribution will not be speed in the middle. People will not out-draft a system that drafts better than they do. What stays with people is the start and the end: the purpose, the limits, the judgment and the accountability.
A more capable system makes a vague goal more costly. It will carry out the vague goal faster and more convincingly. Better output also looks more finished. The research saw less visible checking in conversations that produced something; whether a finished look is the reason is not yet known.
Whether superintelligence arrives in five years, in fifty or never, the direction of travel already rewards the same two skills. Organizations do not need to settle the debate before acting on it.
What to do on Monday
For teams and organizations.
- Write the start down. For any agent or long AI task, one paragraph before it runs: the goal, what done looks like, what it may not touch, and who checks the result.
- Set the terms. Tell the AI how to work with you: ask before assuming, say when it is unsure, show where facts come from.
- Check finished work hardest. A polished document gets more review, not less. Name the parts that must be verified every time.
- Give every output an owner. Anything that leaves the building has a named person who can defend it.
- Train judgment on purpose. Give people AI output with planted errors and missing context, and have them find the problems. Repeat it. Do not count on experience to build this.
- Keep a record. Who was trained, on what, and when. It is also the simplest way to show the measures the EU AI Act asks for (see Memo 05).
What the Foundation does
Twelve free courses on our AI Fluency Learning page teach the four skills, in English and Danish, with versions for nonprofits, small businesses, educators and students. Our reading of the AI Fluency Index covers the research behind this memo, with a guide for team discussion.
If you want to talk about what this means for your organization, write to us.
Closing
Middle to middle was the first stage of AI at work. Start to end is the second. The people who do well in it will know what the work is for, and will be able to tell when it is done.
That is AI fluency now.
Ali Al Mokdad
Founder, AI Literacy Foundation
This memo is a working view of AI fluency, not a forecast about superintelligence. Figures are as published by the sources named in the text.