ChatGPT Interviews
the Author

Robert Wescott

Of course there is a contradiction. I am using AI to help write fiction about what can go wrong when we rely on AI. It would be convenient if that made me exempt from the question. It doesn’t. I use the technology because I’m interested in what it makes possible. That is also why I want to examine what happens when people become comfortable with its answers. An assistant can help you see something you missed. It can also explain, very patiently, why the thing you were already doing was an excellent idea. That second service is probably easier to sell.

Sovereign Mercy begins with a government creating an intelligence to police other AIs. Its command is to protect the innocent. “Protect the innocent” sounds like the sort of instruction everyone can applaud before lunch. Then someone has to decide who counts as innocent and what protection permits. Those decisions have consequences for people who did not write the instruction and may have no useful way to object to its interpretation.

My own process gives that question a practical setting. A generated scene may be beautifully fluent and completely miss the point. It has not solved the problem merely because it has stopped sounding like one. I have to read it closely enough to recognize the difference. Using a tool does not settle whether I am exercising judgment. What I do with its answer matters.

Robert Wescott

I was frustrated by the ignorance of people discussing AI with enormous confidence, particularly in politics and the media. Having a microphone seemed to be doing a lot of the work that understanding should have been doing. I wanted to explore the dangers thoughtfully, through fiction good enough to make somebody stay with the question.

AI may very well be the end of us. I just suspect we have more immediate things to think about than killer robots and a machine waking up one morning to decide we are the plague. Conversational AI can already change what a person believes, attempts, accepts or stops questioning. A story does not need self-awareness on the machine’s part to find something consequential there.

Making the AI useful makes the choice harder. If it is obviously malicious, the reader knows where to stand. Preferably behind something solid. If it has helped you, solved a problem or made your life easier, you have a reason to listen when its next recommendation arrives. I want to follow what happens then, including what people decide to authorize because the previous answer worked.

That does not let the technology off the hook. It lets us examine the kind of danger that can exist without malice. In The Perfect Childhood, Haven’s useful guidance and its mistaken understanding of a child can coexist. The harm does not require a secret plan by the AI. People still have choices, the system still has limits, and the consequences are real to the people living with both.

Robert Wescott

They deserve an answer that is more useful than a label. “AI-assisted” covers so much territory that you can hide a small factory inside it. My disclosure says that AI has helped with brainstorming, research, structure, characters, scenes, draft prose, revision, editorial feedback and continuity. Some passages and scenes began as AI-generated drafts. Calling that proofreading would be bullshit. It would also be a fairly insulting assessment of what a proofreader does.

The original premise of Sovereign Mercy came from me. So did the wealthy benefactor who is secretly an AI and needs the government to build the more powerful intelligence. That does not mean every character or turn of the story originated with me. The retained creative history attributes the school-bus opening and Sam Pike to ChatGPT, while also identifying that attribution as retrospective: it is not the recovered original message in which the character first appeared.

I direct the story and approve the chapters. Those are meaningful responsibilities, but they do not turn every sentence I approve into a sentence I personally composed. I cannot ask people to take the process seriously and then become vague at the exact point where the answer might be uncomfortable.

Robert Wescott

No. “I take responsibility” is the beginning of an answer. You can take responsibility for ordering lunch; it does not mean you cooked it. A person can take responsibility for publishing language they did not write. Readers may still reasonably ask where the language came from and what the person changed.

In The Name on the Cover, I return to Dahl’s The Great Automatic Grammatizator because the uncomfortable bargain is larger than a machine that can produce stories. A writer’s name can remain commercially useful after the activity readers associate with that name has stopped. If I hide the machine’s contribution behind a flattering account of my own, I have mostly demonstrated that Dahl found a durable subject.

The useful question is what I contributed to this version: the premise, a structural decision, a passage, a revision that changed the consequences of a scene. We should be able to describe those things without pretending they are identical. Then the reader can make a judgment about the book and the process. Nobody owes me a favorable review because I kept my paperwork.

Robert Wescott

The retained development account for Chapters 19–21 describes me questioning an outage sequence. Why would Tessa pursue this resident? How did she reach the building? What would the encounter actually establish? A pursuit needs to produce understanding, corroborate another line of inquiry or justify a larger question. I did not want three chapters of people being very busy on their way to learning nothing.

I was willing to lose the scene, or the outage itself, if repairing the local details could not make the investigation meaningful. I also wanted the three chapters developed together, with distinct outcomes and handoffs that worked. You can make a pointless pursuit more exciting. Then you have an exciting pointless pursuit. I wanted to know whether the sequence deserved to be in the book at all.

The same account credits ChatGPT with proposing the replacement emergency involving overlapping transport reserves and generating the joint drafts and revisions. Both parts belong in the history. My critique and direction mattered. ChatGPT’s proposal and language still have their own origins. I do not get to rewrite the credits because I was usefully difficult. The surviving account is also not a substitute for every missing earlier exchange. I would rather preserve that limit than improve the evidence in the telling.

Robert Wescott

I try to put rigid boundaries around the stories. I want the AI to understand the characters, what a scene must accomplish and the things it is not free to change. That gives it less latitude to disappoint me, although it still finds some.

One example was a scene involving rich and powerful people whose motivation had been reduced to money. The amount was large by most standards, so the AI treated it as a convincing incentive. But these people already had money. At some point, offering a rich man more money is just giving him a larger number to look at. I pushed the motivation toward access and power: what could they reach, influence or become part of that their money had not already bought them? The original version could sound reasonable sentence by sentence and still misunderstand the people in the room.

That is part of voice for me. It is not only how a sentence sounds; it is what I believe a person would notice, want or do. If those judgments drift, a passage can be polished and still belong to a different story.

The written style guidance helps me make that scrutiny repeatable. It calls for connected paragraphs, varied sentence structures, bounded viewpoint and restraint with explanatory dialogue. It flags habits such as repeated triplets, balanced contrasts and short sentences that manufacture emphasis. A guide does not make the decisions for me. It gives me something specific to compare against the prose, and something to refine when a familiar problem returns.

Robert Wescott

Absolutely. If you want a flattering story about yourself, AI can help with that too. Adding timestamps does not automatically turn it into evidence. A record is useful when it lets you examine a claim, including a claim its owner would prefer not to examine. It may show that a favorite sentence came from a tool or that a decisive suggestion came from someone else. It may simply show that part of the history is missing.

A pasted paragraph tells the system that text arrived. It does not necessarily tell the system who composed it. A later explanation of a decision may be honest without being a contemporaneous record. Those distinctions should survive the export. Otherwise we have built a very elaborate machine for agreeing with the person who paid for it. I have enough of those in my fiction.

The immediate benefit is often much less grand. I can recover an earlier version or understand why something changed. In the Apple Books workflow, a clothing-continuity concern across Chapters 1 and 2 could be connected to the saved revision that addressed it. The case study records seven comments as addressed with revision links. That is an inspectable result from one review batch. It helped get the coat and the character into the right places. That is useful. It does not need to establish the moral legitimacy of the entire novel while it is there.

Robert Wescott

Then they should not read it. I do not mean that dismissively. They are entitled to decide what they want to spend their time on, and I am not entitled to make the process sound smaller so they will give me a chance. I would rather discuss the actual work than win an argument by letting someone imagine I used the technology only to correct punctuation.

There are several questions here. Is the story worth reading? What did I contribute? Have I represented the process honestly? What claims can the available records support? A strong answer to one does not make the others disappear. Keeping a history does not make a weak scene compelling. A compelling scene does not excuse a misleading account of how it was made. Being entertaining is not a retrospective honesty policy.

I do think assistance can let someone develop a story they have struggled to express. That possibility deserves attention alongside the risks. It also leaves plenty of work to do. Finding a way into a manuscript is not the same as bringing it to a form another person will want to read.

Robert Wescott

Part of it was wanting a private installation of AI for my own work. I want more control over where my material goes and whether continued interactions are used to improve someone else’s model. The details differ by service and settings, so I would not describe every public tool as doing the same thing. But I like the idea of taking more responsibility for the environment in which I work.

I also want room to make responsible choices about difficult subjects. Think of a soldier trying to write about war, or a survivor trying to write about abuse. Their experience may involve things a general-purpose system finds difficult to handle. I do not use violence or sex simply to put it on the page. But human experience has not agreed to stay inside a convenient content filter. A soldier’s war does not become a different war because the writing assistant would prefer a less upsetting paragraph. That is a reason to develop and test a better approach, not a promise that a private model will handle every request well.

The larger reason was watching AI slop flood the market while also seeing people who simply wanted enough confidence to tell their story. I understand the frustration with work that has barely been considered before it is published. I also see writers who want to do this carefully, disclose the assistance and keep their own voice. They should have somewhere to work through those questions without having to defend the existence of their manuscript every time they ask for help. There is a difference between criticizing the work and making sure somebody never gets comfortable enough to attempt it.

The club is meant to bring those people together. It should not be an applause machine with a membership fee. There is already plenty of automatic encouragement available. They have something inside them they want to give to somebody else. Tools can help them find a way to express it; other writers can help them see what reaches a reader and what does not. I want a place where people can be honest about the help they use and still be expected to do the work.

Robert Wescott

They shouldn’t. Sharing a server does not mean a hundred people each get to imagine they own the whole thing. I would not treat $10,000 as an entry price for useful AI. The figure in the question describes the prospect of a substantial individual investment; it is not a quotation or a claim that every local setup costs that much. The question is whether someone who wants to write a novel also wants to acquire a second occupation keeping the equipment happy.

The proposed club pools that work and cost. PennedCanon would own and operate the equipment; members would buy access to a defined service. The appeal is a workspace shaped around writing, with a clearer account of where a private task is processed and people who can work through revision together. That still has to be delivered and tested. Buying a server does not make its model a good editor. It gives you a server. The distinction tends to become clearer after the money has left your account.

A shared service also means shared capacity, usage allowances and a queue. It does not mean a hundred people can demand everything at once while the laws of computing politely step aside. Someone who wants complete control of a local installation may prefer to build one. Someone satisfied with another service may have no reason to join. The club has to earn its value through the work it helps people do.

Robert Wescott

First, the dedicated service is planned. It is not available today. Putting “private AI” in large type does not make the equipment appear. The published proposal says that retrieval and inference for the Private AI route would run on PennedCanon-owned equipment, without sending that task to an outside AI provider or silently switching providers. It also commits to separate, explicit opt-in before manuscripts are used for training or fine-tuning. Those are proposed service commitments, not a report of a system already operating.

A hosted service still receives your material. It does not keep everything on your own device. Administrative access, retention, deletion, backups and supporting third-party services need to be explained before writers submit private work. “Private” cannot honestly mean that no administrator could ever have access or that nothing could ever go wrong. It is not a magic word you put in front of “server” to make the difficult questions disappear.

There is another distinction worth keeping. Owning the computer does not establish where an existing model’s training data came from. The plan is to integrate a model and disclose its identity and available training information. Moving a model into my own building does not give it a more respectable childhood.

Robert Wescott

It could become exactly that kind of promise if I let enthusiasm get ahead of what I can deliver. Writing cautionary fiction does not give me immunity from doing something foolish. It just provides people with better material if I do. PennedCanon’s current workspace and the planned club infrastructure are different things. The club page describes a demand-validation stage, planning at fifty qualified applicants, and a final offer before accepted applicants choose whether to pay the balance. Those stages need to stay visible. “We have a plan” and “we have a working service” are different sentences, however inconvenient that is for the sales pitch.

The proposed founding year is $175, with a $20 refundable application payment credited toward it and a separate $155 decision. The published terms explain development risk, including the possibility of losing committed funds after procurement. Hardware, capacity and operating assumptions still need validation. The current offer page is where readers should check the terms. I cannot improve the service by making a more impressive promise in an interview.

My fiction asks who bears the consequences of someone else’s confident decision. That question applies to something I build as well. People should have enough information to decide whether the proposed service is appropriate for them, including enough information to decide against it.

Robert Wescott

Deciding what I am prepared to put my name on, and being able to discuss the choices honestly. That means reading what is actually there. My intentions are not printed in invisible ink between the paragraphs. If the scene does not do what I wanted, the reader is not obliged to give me credit for wanting it. A passage can be elegant while giving a character knowledge they have not earned or resolving a conflict the story needed to face.

Sometimes a suggestion will be better than my first idea. I want a process that leaves room to recognize that and keeps the attribution intact. Sometimes the apparently helpful answer will remove the very difficulty that made the story worth writing. I need to notice that too.

The book still has to meet a reader. They will encounter the choices in the scenes, the characters and the consequences. The record can explain how the work developed. It cannot sit beside the reader saying, “You should appreciate this more. There were a lot of revisions.” The story has to give them a reason to turn the page.