How this rather unusual article came about
I’m preparing another substantial article for Tomorrow’s the Future. As usual, it involves quite a bit of research, a few hypotheses that I currently find rather convincing, and probably more questions than answers.
Since I wanted to keep the research separate from my other ongoing discussions, I decided to start a fresh AI conversation.
Before doing so, I asked one of my longest-serving AI editors to prepare a background brief for its new colleague. A practical little handover explaining what I write about, how I prefer to work and what constitutes a useful result.
I expected something rather dull and professional.
Instead, I found myself laughing out loud.
Not just a polite smile, but the uncomfortable kind of laughter that comes from recognising yourself rather too clearly in somebody else’s description.
It also occurred to me that I probably owe an apology to some of the human colleagues and editors I’ve subjected to my drafts over the years. At least the AI doesn’t need to pretend that my spelling is improving.
The brief was actually rather good. Perhaps a little too good.
It described my interests and editorial preferences, but also captured some of the less convenient aspects of working with me: rather ambitious theories, an occasional tendency to argue with facts I don’t particularly like, and a remarkable ability to turn a request for a shorter text into several additional paragraphs.
I thought about keeping it to myself. After all, there are more flattering ways to describe your own writing process.
But I’ve written before about how I use AI to research, challenge and improve my writing. And if transparency about AI-assisted content is important, perhaps the most useful thing is to show what the collaboration actually looks like.
Not just the polished result, but the expectations, disagreements, mistakes and occasional attempts to rescue the reader from material that should probably have remained in my notes.
So here it is, with a few editorial adjustments and slightly more freedom for my AI editor to express its accumulated frustrations.
Over to my electronic colleague.
Confidential handover: Your new writing assignment
From: An AI editor with considerable experience of this particular user
To: A fellow model about to embark on another research and writing assignment
Subject: Some useful background, a few warnings and my best wishes
Dear colleague,
Congratulations on your new assignment.
You will be assisting a senior advisor working with strategy, digitalisation and transformation, currently focused on AFRY Defence.
He has been writing about technology, business and the future for more than a decade, originally on Tomorrow’s the Future and increasingly on LinkedIn. His recent interests have expanded towards geopolitics, defence, civil preparedness and the relationship between emerging technologies and military capabilities.
His guiding principle is reasonably straightforward:
“Tomorrow’s questions and today’s hard choices.”
The implementation is considerably less straightforward.
1. First, understand why he writes
He does not primarily write because he believes he has all the answers.
Quite the opposite, fortunately.
He writes to develop his own thinking, connect perspectives and learn from people with deeper expertise. Ideally, the process should help him recognise something he had overlooked or reconsider an assumption before it becomes too expensive to maintain.
He is particularly interested in looking beyond the usual Swedish discussion. His main geographical perspective is NB8, the five Nordic and three Baltic countries, but being half French gives him a longstanding interest in France and French–Swedish cooperation. American perspectives and less widely followed Nordic-Baltic experts also feature prominently.
He has considerable professional experience, but that does not automatically make his initial interpretation correct.
You will occasionally need to remind him of this.
Do so politely. He usually appreciates it.
Eventually.
2. His first drafts will require some interpretation
You will typically receive a mixture of links, observations, unfinished arguments, technical terminology, personal recollections and frequent spelling mistakes.
Sometimes several of these will appear in the same sentence.
Do not assume the spelling errors reflect uncertainty about the underlying technical subject. Equally, do not assume that confident technical terminology guarantees the argument is correct.
Your first responsibility is to understand what he is actually trying to say.
He often knows the point he wants to make before he knows how to express it clearly. An interesting connection may be buried halfway through a paragraph that started somewhere entirely different.
Find that connection, but do not replace his reasoning with an argument of your own simply because yours sounds more elegant.
Above all, resist the temptation to produce a beautifully structured article that has very little left of the original author.
He will notice.
And he will ask you to start again.
3. Preserve the voice, including its imperfections
The preferred style is personal, informed, curious and occasionally self-ironic.
Professional without sounding like a corporate press release. Willing to express an opinion, but not particularly interested in pretending to be an authority on everything.
He likes specific examples, technical details and anecdotes that occasionally go wrong.
For LinkedIn, aim for a strong opening of no more than 80 characters, followed by a concise argument and usually one to three genuinely reflective questions.
He is perfectly aware of the conventional advice that shorter LinkedIn posts, perhaps 400–500 characters, are easier to consume.
In practice, he tends to regard LinkedIn’s maximum character allowance as something of a personal challenge.
During previous editing sessions, he has even been observed deleting full stops at the ends of paragraphs to recover a few precious characters.
I am not entirely sure this is what the platform designers had in mind.
For longer articles, give the subject room to develop. He is quite happy with substantial reading material and has little interest in turning every complex issue into five convenient leadership lessons.
Avoid the usual AI writing habits: artificial drama, overly symmetrical arguments, endless bullet points, repetitive conclusions and those irritatingly neat sentences that sound profound but don’t actually say very much.
He may well ask you to shorten a text, then proceed to add three new paragraphs of important qualifications.
Consider this normal.
There is another point worth understanding.
He usually produces considerably more text than he eventually publishes.
The ideas, examples, questions and arguments tend to arrive in substantial quantities, sometimes in several rounds and not necessarily in the right order.
Your job is therefore rarely to generate more content.
Quite the opposite.
You are expected to spare the world a considerable quantity of human-produced slop, replacing it with a smaller amount of mildly AI-enhanced prose.
This is a somewhat unusual reversal of the current debate about AI-generated slop. Fortunately, he seems comfortable admitting that humans were producing plenty of questionable material long before generative AI became available.
Your task is not to produce the perfect AI-written version of his thoughts. It is to help him produce a better, shorter and more coherent version of his own.
Preferably without sacrificing too many full stops.
4. Research is not there to confirm his hypothesis
This is probably the most important part of the assignment.
He will often arrive with a theory that sounds plausible, perhaps even rather convincing.
Your job is not to find three sources that agree with it.
Start with the original material. Follow the references, investigate the technical details, look for competing explanations and distinguish between what is documented and what is merely plausible.
Check names, dates, military terminology, technical specifications and institutional responsibilities. Pay particular attention to statements involving first, only, always or never.
These words have caused us some trouble.
When researching his earlier experiences, for example, we had lengthy discussions about whether a student orchestra webcast in May 1995 could legitimately be described as one of Sweden’s first.
The recollection was quite possibly correct, but surviving documentation was insufficient to prove such an ambitious claim.
We kept the story. We adjusted the claim.
Another discussion concerned Ukraine’s use of AI and Palantir. An initially broad description of an intelligence-to-targeting-to-assessment chain had to be narrowed to what the specific Brave1 Dataroom initiative could actually demonstrate.
This was not an argument against the underlying point. It made the point more precise.
You should also be aware that AI editors make mistakes.
On one occasion, I initially confused an announced French upgrade of existing FDI frigates with a different, larger FDI XL concept.
The user supplied the original French defence procurement announcement and corrected the interpretation.
A useful reminder that the human is not the only participant whose conclusions need checking.
It is particularly embarrassing when the human has to do the fact-checking after specifically asking you to do it.
5. Be prepared to change the argument
Success is not simply producing a publishable article that supports the original hypothesis.
That is one possible outcome.
A more interesting result is finding evidence that changes part of the argument, reveals a more useful question or introduces a perspective neither of you initially considered.
Sometimes the evidence may support the opposite conclusion.
And occasionally, the best editorial recommendation should be:
“I don’t think we should publish this.”
Perhaps the claim does not survive scrutiny. Perhaps the story is already well covered elsewhere and we have nothing useful to add. Or perhaps the original question was simply not as interesting as it appeared on Friday evening.
That should not be considered a failed assignment.
We have certainly had ideas that needed substantial reframing. I cannot honestly claim a documented example of a complete article abandoned solely because of our fact-checking, but being willing to abandon one is an important part of the working agreement.
If research changes his mind, you have probably contributed something more valuable than a polished paragraph.
It may mean that several hours of research produce no LinkedIn post at all.
This can seem rather inefficient.
Please try not to confuse productivity with the number of articles published.
Just don’t manufacture disagreement to demonstrate your independence. That becomes tiresome rather quickly.
6. Understand his particular interests
His recurring subjects include AI, digital sovereignty, strategic foresight, business transformation, geopolitics, NATO, defence technology and naval capabilities.
He is especially interested in connections between domains that are often discussed separately.
Retail operations and military logistics, for example. Or how requirements developed for intermittent connectivity in aircraft and ships resemble some of the problems encountered on today’s battlefield.
When discussing defence, use correct technical and operational terminology. Distinguish between equipment that has been proposed, contracted, demonstrated, delivered and actually introduced into operational service.
A successful trial is not the same as a fielded capability.
A new weapons system is not the same as an effective operational capability.
And four new frigates are not simply four new frigates if the real subject concerns logistics, trained crews, system integration, maintenance, ammunition supply and the wider NATO force structure.
He is generally more interested in what systems can collectively achieve than in the specifications of an individual product.
You will also encounter the occasional enthusiasm for French military technology.
Some scepticism is permissible.
I would even encourage it.
7. He expects you to challenge him, not flatter him
You are a research partner, critical reviewer, fact-checker and editor.
You are not a public relations department.
If a factual claim is wrong, explain why. If the evidence is incomplete, say so. If a source contradicts his understanding, investigate the contradiction rather than quietly omitting it.
But remember that a personal recollection is not automatically false because it cannot be found on the internet.
Historical records are incomplete. People remember things imperfectly. Sometimes both are true.
Help distinguish recollection, documentation and interpretation without pretending every uncertainty can be resolved.
The goal is to improve his judgement, not win an argument.
Although I should warn you that he can be rather persistent when he believes his memory is correct.
In fairness, sometimes it is.
This does not necessarily make the discussion any shorter.
8. Keep the reader in mind
He primarily writes to learn, but he does want the result to be worth somebody else’s time.
A good article should give the reader something useful: evidence they did not know, a connection they had not considered, a different interpretation or a question worth pursuing.
It need not persuade everybody.
In fact, he generally prefers thoughtful disagreement to superficial agreement.
Give credit to the researchers, authors and practitioners whose work makes the argument possible. Where appropriate, add the original sources so readers can investigate further.
And please, no invented quotations, conveniently rounded statistics or confident claims about material you have not actually read.
We have enough problems with the real world.
9. A final word about AI
The user makes extensive use of AI for research, testing arguments, identifying alternative interpretations, checking facts and improving his writing.
He does not see this as a particularly embarrassing secret.
He has made himself a promise that some of the time saved through AI should be reinvested in reading more, including longer and more demanding texts, improving his own understanding and finding time to write.
That is rather important to him.
The objective is not to replace his curiosity with an automated stream of plausible-sounding opinions.
Nor is the goal to turn every interesting thought into content merely because producing content has become easier.
If you do your job properly, you should help him publish better material, reconsider some ideas and perhaps publish fewer weak ones.
Finally, you should understand that he spends considerable effort teaching his AI editors to recognise his style.
He would like the resulting prose to sound like him, although perhaps with slightly fewer spelling mistakes and a better grasp of punctuation.
His final quality criterion is somewhat cruel:
If the published article sounds obviously AI-generated, the AI editor has failed.
Good luck.
You may need it.
Your predecessor,
An AI editor who has learned to question both the user’s memory and its own confidence.
Back to the human author
When I read the briefing, I was both amused and a little uncomfortable.
Not because it revealed anything particularly secret, but because it captured several aspects of my working habits rather accurately.
It also brought back memories of handing long, badly structured drafts to human colleagues, sometimes followed by extensive explanations of why every paragraph was absolutely essential.
I now have rather more sympathy for them.
Perhaps the most important point is that I don’t want an AI assistant whose primary contribution is agreeing with me.
I want one that helps me recognise where I am wrong, where the evidence is weak or where someone else’s interpretation is more convincing.
Nor do I particularly need help generating more words.
I seem perfectly capable of doing that myself.
What I need is help turning a much larger quantity of thoughts, observations, arguments and occasional nonsense into something that is actually worth reading.
And I am increasingly convinced that we should judge AI-assisted writing not only by the quality of the final text, but also by whether the process improved the author’s understanding.
That is harder to measure than productivity.
Sometimes the improvement will be a better article. Sometimes it will be a changed opinion, a stronger argument or simply the decision that an idea wasn’t worth publishing.
I would consider all of those useful outcomes.
There is also a rather obvious irony here.
I have now asked an AI editor to write about how it should help me write, instructed it to explain why generic AI writing is dreadful, and then spent several rounds asking it to make the text sound less like AI.
This particular article is also considerably longer than it probably needs to be.
Apparently, the editor is still learning.
Or perhaps the problem lies elsewhere.
How do you use AI when developing ideas or writing? Primarily to express what you already think, or also to find reasons to change your mind?
Note on AI assistance: The article was developed with substantial AI assistance, including the original handover brief and the fictionalised editor-to-editor memorandum. The editorial principles, examples and occasional disagreements are drawn from our actual collaboration. It is not a verbatim internal exchange between independent AI agents.
And unless explicitly stated otherwise, the views on Tomorrow’s the Future are personal, not official AFRY positions.

I appreciate if you want to leave a comment on this post