When one AI writes the message and another decides whether you see it
A few days ago, I saw someone explain how they reviewed AI-written text before publishing it. They did not simply ask the same chatbot to “make it sound more human,” which is usually how you end up with the same text wearing a slightly different tie. Instead, they had built a serious process involving six separate AI reviewers. One checked the facts, another the logic, while the others examined the intended audience, concreteness, editorial quality and voice.
I thought this was clever. It was certainly more thoughtful than generating a post, removing the em dashes and hoping nobody noticed.
Then a less generous thought occurred to me.
This was an extraordinary amount of work to make an AI-written text feel human when the person receiving it might immediately hand it to another AI and ask:
Summarise this and tell me whether there is anything relevant to me.
The first machine may have spent several minutes restoring warmth, rhythm and personality. The second may remove all of it in seconds and return:
Generic argument about authentic leadership. No new evidence. No action required.
At that point, I began wondering whether we were building AI for the wrong side of the conversation.
Most of the excitement has been about production: more emails, applications, articles, presentations and videos, personalised until each appears to have been created for one person. The attention may be simulated, but it can now be simulated at industrial scale.
The cost of plausible communication is collapsing. Human attention is not.
Once production becomes abundant, selection becomes power.
Humans will receive the exceptions
Customer support already provides a model for what comes next.
The older arrangement placed people in front of a queue. They answered the easy requests and escalated the difficult ones. The emerging arrangement reverses this. AI handles the routine cases, resolves what it can and passes the uncertain or sensitive ones to a person. Systems such as Intercom Fin are already built around that principle.
Humans receive the exceptions.
The same architecture can spread elsewhere. An executive receives the three proposals that appear relevant to current priorities. An investor examines the pitch decks that survive an initial assessment. A manager sees the risks, contradictions and decisions extracted from a stack of reports rather than reading every document behind them.
The important AI may therefore not be the one speaking for us. It may be the one standing in front of us, deciding what we never need to see.
I understand the appeal. I receive messages from people who have apparently found my background fascinating and would love to connect, exchange ideas or explore cooperation. I am curious, so this works on me more often than it should. Quite frequently, the interesting conversation turns out to be a sales call from someone who has not established that I need the product.
After enough of these meetings, a receiver-side AI starts to sound less like dystopian machinery and more like self-defence. I would happily use a system that reads the message, considers my current work and tells me:
This appears to be a sales approach. No specific reason has been given for why it is relevant to you. No response recommended.
I have already asked AI to surface messages involving clients, paid work, tax, legal matters or family logistics, while routine receipts can remain out of the way. That sounds sensible, and mostly it is, but it also delegates authority. The system is beginning to decide which parts of the outside world may interrupt me.
Earlier filters mostly classified, ranked or blocked information. The new gatekeeper can interpret intention, compare a message with private context and act on the recipient’s behalf. It may answer, reject, request evidence or negotiate the next step before a person becomes involved.
That is the difference between a spam filter and a personal representative.
Once this becomes normal, sending a message will no longer mean placing it in front of another person. It will mean submitting it for admission.
The assistant and the institution
There is an important difference between the AI protecting my attention and the AI deciding whether I deserve someone else’s.
When I choose the criteria, the gatekeeper feels like an assistant. The relationship changes when a company, platform or institution controls the gate, because I may not know the criteria, see how my message was interpreted or have any meaningful way to challenge the result.
The same technology feels like an assistant when I control it and like an unseen institution when somebody else does.
Recruitment makes this particularly visible. Candidates use AI to study job descriptions, adapt résumés and draft applications. Recruiters are gaining tools that score and rank applicants before most résumés receive sustained human attention. Gem, for example, describes a system that ranks candidates against criteria chosen by the recruiter, while leaving the formal decision to advance or reject them to a person.
The machine may not make the final decision, but it can influence who is seen first, who receives careful consideration and who remains near the bottom of a very large queue.
The resulting loop is slightly comic. A candidate’s AI helps write the application, a recruiter’s AI evaluates and ranks it, and the candidate’s AI later writes the follow-up. The humans may not meet until several machines have already formed opinions about them.
I have submitted applications where I spent hours trying to understand what the company actually needed, looked beyond the title and explained why experience from several industries and countries might be relevant. Frequently, nothing came back.
Perhaps a person read the application and remained entirely unconvinced. That is normal. The more unsettling possibility is that the explanation never reached a person capable of being unconvinced.
Recruiters have always screened applicants, editors have always rejected submissions and executives have always relied on assistants. What changes is how much interpretation can now happen invisibly before a human becomes involved. Thousands of small judgments can be converted into a reasonable-looking order of attention, and being ranked low enough may produce exactly the same outcome as rejection.
What survives compression?
A sender may add a story, improve the rhythm and remove anything generic. The result may genuinely be better. Then the recipient’s AI compresses it into four lines.
A large preprint published in July 2026 suggests that this is already changing the character of what people receive. Researchers examined tens of thousands of AI-generated summaries of news articles and found that, while the summaries were broadly accurate, they also reduced ideological language, partisan tone, negativity, anger, fear and personal voice while increasing clarity. The machine was not simply making the original text shorter. It was editing its character.
Personality will still matter once communication reaches a person. The problem is that it may have less influence over whether it reaches them at all.
After compression, what remains may be less flattering. Is there a distinctive idea, and is there evidence behind it? Is it relevant to something the recipient is trying to decide? Has the sender established a credible reason to make the claim, and is there a clear request?
A beautifully written pitch with no customer evidence may be reduced to exactly that. A personalised email may become “sales message with unclear relevance,” while a long strategy document becomes “several activities, but no decision.”
There is something healthy in this. It may force us to replace superficial personalisation with actual relevance and rhetorical confidence with evidence.
It will not, however, remain an innocent standard for long.
As soon as receiver AIs reward credibility, specificity and relevance, sender AIs will learn to produce the signals those systems expect. We have seen simpler versions whenever people write for search engines, applicant systems, spam filters or social feeds. The measure becomes a target, and the language evolves to satisfy the gate rather than inform the person behind it.
The result may be machine-readable relevance theatre: messages engineered to look concrete, well-evidenced and personally significant to the receiving system, whether or not they deserve anyone’s time.
At that point, communication becomes a negotiation between agents before it becomes an encounter between people.
My principal would like thirty minutes to discuss a partnership.
What evidence suggests this is relevant?
Three comparable companies improved conversion by 18 per cent.
Send the case study, pricing and the decision you require. No meeting recommended yet.
This sounds faintly absurd, although it may be less absurd than two people spending an hour discovering that neither knows why the meeting was booked.
The filter that protects us from surprise
The danger is not primarily that these systems will be incompetent. Bad filters are irritating but visible. A highly competent filter may be more influential because we stop noticing its decisions.
My AI could learn that I rarely respond to generic sales approaches, distrust unsupported claims and care about a handful of current problems. This would be useful. It would protect hours from people who want my attention without having earned it.
It might also learn that vague propositions from unfamiliar senders have a low probability of producing value.
Statistically, it would probably be right. The hit rate is terrible.
Unfortunately, the badly written message from the wrong person, concerning something I did not know I cared about, is occasionally the one that changes what I do next.
I have written before about the golden retriever that seems to live somewhere inside me. It becomes interested in unusual people, unclear opportunities and unfamiliar problems long before the more sensible part of me has established whether any of them are worth pursuing. It has led me into pointless meetings, but also into friendships, work and possibilities that could not have demonstrated their relevance in advance.
The same tension appears in how we encounter information. Another recent study, based on 24,000 searches, found that AI search surfaced fewer long-tail sources and produced less varied results than traditional search. It does not prove that every AI gatekeeper will narrow our world, but it shows how easily decisions about relevance become decisions about what remains visible.
The purpose of the system is to reduce noise, yet much of what changes us initially resembles noise. New ideas arrive badly packaged. Important people do not always introduce themselves convincingly, and a surprising opportunity may be impossible to distinguish from a hundred irrelevant ones until someone spends time with it.
A humane gatekeeper may therefore need a serendipity budget: some attention reserved for messages it cannot justify recommending but cannot confidently dismiss. It may need to show us what it excluded and occasionally permit the poorly explained possibility through.
That sounds inefficient because it is. Curiosity has always involved accepting some waste.
The gatekeeper we need may otherwise create the enclosure we should fear.
The machines between us
A good gatekeeper could protect time, block manipulative outreach and prevent urgent information from being buried. I am not opposed to such a system. I increasingly need one.
But the power deserves more attention than the convenience.
When the gatekeeper belongs to me, I need to decide how much authority to give it and how much surprise I am willing to lose. When it belongs to an employer, platform or institution, we need to ask whose priorities it serves, what it treats as credible and whether the people it excludes can even see that a judgment was made.
The AI will not merely help us understand the world. It will participate in deciding which parts of the world we encounter, and which people are permitted to encounter us.
We have spent enormous effort teaching machines to sound more like humans, perhaps because we assumed that a human would remain at the other end of the conversation. Increasingly, there will be another machine standing in front of that person, removing our tone, compressing our argument and deciding whether we are worth the interruption.
We will have to say something that survives being understood by a machine.
We will also have to decide how much of the world we are willing to let that machine discard before we have had the chance to understand it ourselves.
Jörn Green 
Lämna en kommentar