On January 1st, I put most of my professional activities on indefinite hiatus. I am now semi-retired and I intend to spend most of my time, for the foreseeable future, writing fiction and (some) creative nonfiction. However, there is always a risk that a relaxed work schedule in which deadlines are few and far between might degenerate into an orgy of laziness. Therefore, I have decided to hold myself accountable by writing a report on my activities at the end of each quarter, starting now. These reports will cover my writing outputs, the associated marketing activities, and their results… or lack thereof. Read the following to help me maintain focus on my objectives for the next quarter.

Note: Exceptionally, this first report will also cover my activities during the transition period leading to semi-retirement. This period started in late October 2025.

Short Fiction

• Works completed — novelettes (8,000 to 20,000 words): 1
• Works completed — short stories (1,000 to 8,000 words): 4
• Works in progress — length TBD: 1
• Total number of words in completed works: 24,700
  

 

• Submissions to fiction magazines: 7
• Submissions to anthologies: 1
• Rejections received: 4
• Submissions under consideration: 4

The competition is ferocious. An author whom I admire greatly, Scott Edelman, once told me that he usually has between 9 and 15 pieces in circulation at any given time and that he has received as many as 82 rejections in one year. Looks like I’ll have to grit my teeth… 

Novel #1: Unconquered Magic

Writing

Completed in October 2025 (91,000 words)

Marketing: Literary Agencies

• Agents contacted: 36
• Rejections received: 19
• Proposals that have timed out: 4
• Proposal that bounced because the agent was closed to queries, although their web site said otherwise: 1 (twice)
• Proposals under consideration: 12
  

 

• Earliest proposal still under consideration: sent November 27th, 2025.
  
• Latest proposal still under consideration: sent March 1st, 2026.
  
• Rejections received in under a month: 12
• Rejections received in 1-3 months: 3
• Rejections received in 3+ months: 4
• Fastest rejection: 49 minutes, on a Friday during lunchtime. (Humiliating, yet impressive.)

It is usually considered a good sign when an agent takes time before responding—without exceeding the deadline by which all proposals are to be considered automatically rejected, of course—because that means that the proposal has not been dismissed out of hand by a first reader and that the agency is considering it seriously. That being said, unless and until someone actually requests a full manuscript, we are still far, far away from the goal.

Marketing: Small Presses (that accept unagented manuscripts)

• Small Presses contacted: 6
• Rejections received: 2
• Proposals under consideration: 4
  

Most small presses require 3 to 6 months, or even longer, to evaluate a manuscript. Patience.

Novel #2: The First Sails of Spring

Writing

• Theme and genre (sci-fi/thriller) selected.
• Main storyline selected and outlined (rough).
• Several potential secondary storylines selected and outlined (rough).

Others

• Essays written, translated and published (for free): 3

Second Quarter Objectives

• Unconquered Magic: Submit to 15 new agents and 5 new small presses, unless a publishing contract can be secured earlier.
• First Sails of Spring: Write synopsis; pick which secondary storylines to include in the novel and which to write as separate short works set in the same universe.
• Short Fiction: finish the story in progress; write at least one short story set in the First Sails universe; submit each new piece and each rejected piece to a new market within 30 days of it becoming available.
• Others: submit at least one proposal for a regular column to an established media outlet.

These days, the political and commenting classes are riveted by the cloak and dagger machinations involving the Carney government and a handful of vacillating opposition Members of Parliament. This is hardly surprising: the Liberals, elected with a minority mandate a year ago, seem on the verge of accomplishing a near-miracle: achieving majority status without a new election, simply by poaching Opposition members one by one.

Of course, not everyone is happy with what is going on. According to an opinion poll cited on The Numbers, a podcast hosted by poll analysts and aggregators Eric Grenier (The Writ) and Philippe J. Fournier (Qc125), many voters would like to see MPs who change their allegiances be forced to resign and win by-elections before they could be seated with their new affiliations. An opinion that, understandably enough, is more or less popular depending on whether one’s party of choice is currently benefiting from the trend or not…

The Context

Four MPs have crossed the floor to join the Liberals since November: three former Conservatives—one of them less than 24 hours after collecting his Secret Santa present at the Conservative Christmas party—and, just this week, Lori Idlout, the erstwhile New Democrat member for Nunavut.

Figuring out the former Conservatives’s motivations is fairly straightforward. The Tory leader, Pierre Poilievre, is a right-wing hardliner whose profound unpopularity among swing voters is the political equivalent of a ball and chain tied around the party’s ankle. Meanwhile, it seems that Mark Carney can do no wrong in the eyes of the electorate at the moment—and his program could very well, at another point in time, have been that of a Progressive-Conservative government. For the few Red Tories still hanging on to the PCC banner, joining the Grits is both the easiest way to achieve power and, possibly, the choice most in line with their values.

Ms. Idlout’s decision is harder to explain, because Mr. Carney’s government has very little in common with the left-wing New Democrats. However, Nunavut does not have strong partisan traditions (at the territorial level, political parties don’t exist at all) and the government’s intention to spend some 32 billions$ to defend the Canadian Arctic is a clear pull factor. Better, if you are one of only three MPs representing the Arctic territories, to join the governing party and help spend this money in ways that benefit the local communities, all of which are far, far away from Ottawa.

The Dilemma

This situation underlines one of the great paradoxes of Canada’s electoral system: we elect individuals, even as we vote for political parties.

Except for a handful of political superstars, the identities of local candidates barely influence electoral results. We pick the person whom we think would be the best leader, and we vote for the local candidate who represents that leader’s party, case closed. Not convinced? Think about this: do you remember the name of the last candidate you voted for who lost? I don’t. Of all the candidates I ever voted for, since the late 1980s, I remember exactly one: the one who was elected MP. But I do know that I voted for the same party every time.

Party loyalty, whether life-long or temporary, explains why the people who voted for the winning candidate are so frustrated when their MP crosses the floor. It’s like watching one’s favourite sports team’s best player sign with a rival as a free agent.

However, as far as the Canadian electoral system is concerned, none of this is relevant. Whether we intend to or not, we elect individuals, not parties. These are people who think, who evolve, and who watch the world evolve around them. It is hardly surprising, then, that they may change their minds from time to time, even in matters as politically charged as party affiliation. In fact, it’s the most normal thing in the world. And when we vote for these people, we also (again whether we intend to or not) choose to place our trust in their judgment. So trust them. No need to let the riding go unrepresented for several months and to waste money on a by-election. The time to punish a floor crosser, if we decide that they have failed us, is the next general election. Which will come soon enough.

I know, I know. “Trust the individual’s judgment” would be a more convincing line of argumentation if Canadian political parties did not whip every single vote to an extraordinarily absurd degree, thus reducing MPs to the level of barely sentient automata whenever they shamble into the House of Commons. But this is a reason to weaken party lines, maybe even to grant MPs the secret ballot when voting on particularly delicate issues, not to take away one of their few traces of autonomy.

As I conducted library research for my master’s degree in computer science, some 25 years ago, I came across a peculiar little book published by a Soviet engineer around 1984. The topic: chess programming.

At the time, researchers seeking to develop an algorithm strong enough to challenge grandmasters at their own game favored the so-called brute force approach. Very roughly speaking, this approach consisted of generating a list of every legal move in a given position, then every legal response to each of them, then every legal response to those, and so on, and so forth, until all of the sequences of moves of a desired “depth” had been examined. Each of the board positions that could be reached through these innumerable sequences would be evaluated, and the computer would play the move leading to the best possible outcome. (More precisely, it would play the move that left its opponent with the poorest outlook, if said opponent played flawlessly. A defensive but very effective strategy.)

The book’s author found this approach eminently distasteful. Chess grandmasters, he wrote, don’t come close to studying every possible move. Instead, they focus on a handful of promising threads (rarely more than two or three) and estimate their likely outcomes five, ten or twenty moves down the road. Thus, if we want the computer to fight with grandmasters on equal terms, shouldn’t we develop algorithms that mimic this much more efficient thought process?

The problem was that nobody could explain how grandmasters picked the promising moves to explore, least of all the grandmasters themselves. In some trivial cases, yes. Otherwise? Guided by a mixture of instinct and experience, they “saw” the good and bad moves, and that was that. Our Soviet engineer knew this better than anyone. After all, he was none other than Mikhail Botvinnik—the five-time world champion!

Nevertheless, Botvinnik spent long years trying to codify whatever formal or informal knowledge he could get his hands on, without much success. A posteriori, we can explain his failure through the following probabilistic argument:

• Suppose that there are, in a given chess position, about twenty legal moves. (The true number is closer to 35 in the midgame, and about 8 to 10 when only a few pieces remain on the board, but 20 is an acceptable average.)
  
• Further suppose that there is one, and only one, optimal move to make in any given position.
  
• In order to simulate a grandmaster’s thought process effectively, we would need to create an algorithm that, just by looking at a position, would select 2-3 promising moves including the optimal one at the expense of all of the others.
  
• A naive algorithm that picks two moves at random from the set of all legal moves would hit upon the optimal about 10% of the time. Its play would hardly be more sophisticated than that of a five year-old who had just been thought the rules of chess.
• However, let’s suppose that we could, somehow, develop a highly effective algorithm that picked the optimal move 95% of the time. Much better, isn’t it?
• Not so fast! As we said before, the secret to the grandmasters’s success is their ability to foresee the outcome of a move, many steps into the future. For an algorithm to match this ability, it would need not only to figure out the most promising moves in a given position, but also the most promising counters, and the most promising counters to the counters, etc.
• Let’s say that we wanted to predict the outcome of a move four steps down the road (or 4 plies, in computer chess parlance). Which, by the way, is a very modest objective by grandmaster standards. An algorithm that selects the optimal move 95% of the time would find all of the optimal moves in a four-ply sequence in only (0.95)*(0.95)*(0.95)*(0.95) of cases, or about 81.4% of the time. In other words, such an algorithm would commit a blunder in almost a fifth of cases, or between six and ten times per game. Any halfway competent club player would wipe the floor with such an opponent.
  
• And if we wanted to go eight-ply deep, which more accurately reflects the grandmasters’s thought processes? Then, the probability of examining the optimal move at each step would fall to 66%. Ouch.

Thus it is hardly surprising that Botvinnik’s approach, albeit smarter than brute force at first glance, never went anywhere—while brute force itself (which never misses a good move, since it looks at everything, all the time) ended up beating world champions.

What now?

Now, let’s transpose this probabilistic argument to the realm of generative artificial intelligence (genAI), i.e., the world of large language models such as ChatGPT.

GenAI relies on statistical knowledge about human language. We train it with huge textual corpora, which teach the AI the frequencies at which any given word X follows word Y or word Z in the written archive.

When we ask genAI to write something, it relies on this statistical knowledge to generate sequences of word tokens that are statistically plausible. For example, we are much more likely to obtain the phrase “Once upon a time” than “Once orange military drink,” because the word “Once” precedes the word “upon” many more times than it precedes the word “orange” in the archive of the English language, and so on.

Of course, the choice of the next word to add to a text that is being generated also depends upon the context. GenAI, however, has a rather limited grasp of that concept. To a language model, “context” is pretty much made up of the N previous words in the text being generated and of whatever happens to appear in the user’s prompt. Human beings, on the other hand, can also rely upon common sense. For example, we know that a car wash is designed to wash cars, which means that it is pointless to go there on foot. GenAI doesn’t, and is quite capable of recommending that the user walk to the nearest car wash to save on gasoline.

Now, let us suppose that we have developed a language model so effective that it picks the perfect word to add to a text in 99.9% of cases, no matter what the circumstances and no matter what the context. As you can well imagine, this is wildly optimistic, because the language model would need essentially infinite knowledge of the world around it (for example: the true purpose of a car wash) in every possible situation.

And yet, even in this wildly optimistic scenario, a genAI based on such a language model will barely manage to output an error-free 1,000-word text a third of the time. Indeed, (0.999 * 0.999 * … * 0.999), where 0.999 is multiplied by itself 1000 times, amounts to a little less than 37%.

Worse: a language model that picks the optimal word 99% of the time will make at least one mistake in about half of 69-word paragraphs.

Some of these errors will be inconsequential or easy to weed out, such as “Once orange”. But others may not be. Suppose that you ask a genAI for the name of the person that won the 2010 Oscar for best performance by an actor in a leading role. “Actor George Clooney” and “Actor Denzel Washington” are both perfectly valid word sequences that appear frequently in the English archive. In the context of Oscar queries, both will likely have high probabilities of being output. How will you know if the answer that the genAI is giving you is the truth? (Trick question: the 2010 winner was Jeff Bridges.)

And of course, when the training corpus is poor in relevant data points, genAI has a tendency to hallucinate an answer that sounds plausible, just for the sake of giving you an answer. Any answer. But that is a topic for another day.

The lesson

The human brain is wired to find patterns, even when they are not there. It’s a matter of survival. Popular books about evolution and neuroscience sometimes invoke the example of a snake in the grass: if our brain thinks it sees a snake when there isn’t one, we might run away screaming, but we’ll be safe. Whereas if we fail to see the snake that is actually in front of us, there is a decent chance that it will be the last mistake we ever make. The cost of a false negative (i.e., not seeing a pattern when there is one) is infinitely higher than that of a false positive (i.e., seeing a pattern when there is none).

Thus, we also tend to find signs of intelligence in the output of genAI algorithms, even though these outputs are no more than statistically plausible sequences of word tokens, or the visual equivalent for genAIs that produce images and videos. Nothing more. Even worse: while the results may seem convincing to us, the models themselves have absolutely no way of knowing whether what they are telling us is true or false—which, according to philosopher Harry G. Frankfurt, is the academic definition of bullshit.

The only way to guarantee that a text generated by AI isn’t full of nonsense is to double-check everything. In other words: to duplicate its work yourself, either before or after the fact.

How many people do you suppose will bother to do that?

The gods help us all…

I don’t hate AI. I would be a world-class hypocrite if I claimed otherwise.

My last sort-of-full-time job was with a company that developed AI software for speech recognition—in 1998-99. My Ph.D. in digital history was built upon a foundation of statistics and machine learning. Heck, if a certain professor hadn’t gone on sabbatical at the wrong time, I would have pursued AI as my primary field of research the first time I started a doctoral program, way back in the Dark Ages of 1991. And there is a nonzero chance that I would have stuck with it, becoming a life-long computer scientist instead of running away to join the video game industry circus.

But not all AI is created equal. The so-called “generative AI” that has been sucking the air out of public discourse for the past two years, and blowing it back into an unsustainable stock market bubble, is by no means representative of machine learning, much less of AI as a whole.

To put it bluntly: AI is fascinating, while generative AI is garbage.

Why I shouldn’t have to write this essay at all

There are a million reasons why genAI is ethically indefensible. The stolen training data. The criminal waste of energy and clean water. The enshittification of work, which is quickly devolving into nonstop post-processing of AI slop. The hallucinations (i.e., lies) that rot the brains of students and pollute everything from legal briefs to government policy reports. And we haven’t even talked about how chatbots are driving users to suicide; which, by itself, should be enough to have the technology banned and its proponents dragged to The Hague in handcuffs.

And yet, “writers” keep using genAI to “write” novels, short stories and nonfiction. Why?

Let us set aside the pathological cases: the terminally lazy, the impenitent fraudsters, etc. To Hell with them. I’m far more interested in the people who use genAI in (relatively) good faith. And I have identified two characteristics of the publishing market that might just drive them to do it.

Now, neither of these justifications is good, but they make intuitive sense. Here is how I think we can debunk them.

AI as a stepladder to nowhere

First reason to use genAI: because it makes writing easier.

Well, yeah, obviously. Crafting a story is hard work. Whittling every paragraph, sentence and word until it fits just so, even harder. And there is so much competition that even our best efforts may not be good enough.

Compared to real writing, prompt engineering is child’s play. That’s the whole point of genAI: to gain access to the fruits of hard labor without the hard labor.

The problem is that, for writers, the work is where the fun is.

Have you ever typed a line of dialogue and realized that the words that appeared on screen were not what you expected? That your subconscious had bifurcated from the story’s outline and given the character something far, far more interesting to say or do than you had planned? Is there a more exhilarating feeling? Or how about the moment when you discover the ideal turn of phrase, the perfect symbol or the right gesture that brings it all together?

No chatbot output, no matter how polished, will ever give you the same thrill, because chatbot output is not yours. At best, you might experience the kind of perverse satisfaction that a Neutral Evil boss feels when taking credit for an underling’s work. Is that enough for you? Aren’t you selling yourself short?

Of course, discovering something wonderful during the writing process may require retconning a previous scene, or maybe even rewiring the entire story until it slides back into its groove. But that, too, is fun. We humans are problem-solving creatures. Letting a chatbot figure out your story is like hiring someone else to solve your sudokus for you. What’s the point?

AI as a coping mechanism

The other reason why writers of good faith might resort to genAI is more insidious, because it is ingrained in human nature and because there is no straightforward counterargument for it, like the one I just made about the joy of writing being more important than saving time and effort.

I am talking about fear of failure.

In the sad, sad world of publishing, rejection is everywhere. Agent queries go unanswered. Magazines buy a minute fraction of the stories pitched to them. Books die on submission or fail to attract more than a shrug emoji in the marketplace. It’s a bloodbath out there.

In other words: for many of us, not only should the joy of writing be its own reward, it is also likely to be its only reward.

(Want a depressing anecdote? In all my years in academia, where publishing is famously competitive, I have never had a journal article rejected. Not once. My university’s press even contacted me to offer to publish my doctoral thesis in book form after I won an award. Meanwhile, I recently started submitting to fiction markets again, after a decades-long hiatus during which I wrote hundreds of pieces for TV, print and stage, and so far I have gone hitless for the season. Yeah.)

Now, I know that we’re supposed to grow a thick skin and learn to handle rejection.

Bullshit.

Name one writer who doesn’t feel the pain. Just one. I’ll wait.

That pain is the crack through which genAI whispers its poison. “Let me do the work for you,” it says as you squirm, “and you won’t have to invest so much of your blood, sweat and tears into it. Therefore, when rejection comes, as it inevitably will, you won’t hurt as much.” 

It’s a dreadful argument, but I don’t have a good answer for it. Indeed, when it isn’t your metaphorical firstborn being laid out for sacrifice, but rather some sort of weird third cousin twice removed whom you would barely recognize if you ran into each other at a wedding, it might be tempting to say, “who cares?”

Now what?

As an advocate against the use of genAI in writing, this leaves me in a bit of a pickle. When I invoke the “joy of writing” argument (and/or any of the other reasons listed in this essay’s introduction) in order to convince one of my peers not to use this technology, I am also asking them, at the same time, to withstand more pain in case of rejection. Which is… not great.

Especially since we humans are hardwired to avoid loss. Roughly speaking, most people would only take a 50-50 bet if they had a chance to win two dollars for every dollar they might lose. (This sounds weird, but it is one of the foundational findings of prospect theory, which earned Daniel Kahneman a share of the 2002 Nobel Prize for economics.) Thus, not only am I asking others to risk harm, but I am also asking them to go against human nature in the process.

I don’t have an objectively satisfying solution to this conundrum. However, I have come up with a temporary solution for myself. It is based on three maxims:

1. The joy of a single acceptance is greater than the pain of X rejections, for all values of X up to a large (but not infinite) parameter N.
   
2. It is possible to increase the probability of acceptance, and therefore decrease the probability of rejection, through hard work and effective market research.
   
3. If X ever exceeds N, it will still be possible to enjoy the fruits of writing without suffering the negative consequences of rejection by not submitting my writing for publication any more.

I believe that every writer has a personal N value. (Mine is probably lower than most.) The key is to give oneself enough time to figure it out.

We’ll see, a year from now, if this theory holds up. In the meantime, just don’t use genAI. It sucks.

As I transition away from traditional employment and towards full-time creative writing -- in two languages, as if one weren't enough trouble -- I welcome you to the brand-new English version of the web site that I've been maintaining in French, on and off, for a dozen years.

There isn't a whole lot to see here as of yet, but watch out for musings about the publishing industry and (hopefully) some good news about my work.