The thesis
You felt it before you could put it into words. Content produced with a language model looks alike. Not only yours, your competitors’ too. Smoothed syntax, identical transitions, a generic flavour that turns an opinion piece into a LinkedIn post and a white paper into a brochure. You improved the prompts. You fed the model your editorial charter. You even paid for brand voice tools that promised singularity. The result stays, to varying degrees, the same: recognisable at a glance as machine-made.
The common answer circulates, and it is wrong. People say the prompt was badly written. People say the output needs humanising, more friction, a broken rhythm, a tighter brief. All of that treats the symptom, and ignores the cause.
The cause is mathematical. Until you understand it, you keep spending energy on the wrong variable.
Here is the thesis of this article: the blandness of language models is not a bug to fix downstream, it is the direct expression of their objective function. A language model statistically optimises the probability of the next token. That optimisation is, by construction, a regression to the mean. Singularity cannot emerge from it spontaneously; it can only be obtained by constraining the model from the outside, through editorial infrastructure defined upstream. Three proofs will support this.
Proof one: what a language model actually optimises
Start with what nobody says clearly enough. A large language model is not trained to write well. It is not trained to understand. It is trained to minimise a loss function, cross-entropy, on the prediction of the next token.
Here is what happens during training. The model receives the start of a sentence, say “the mouse ate the”. It has to predict the next word. It generates a probability distribution across its entire vocabulary, perhaps 50,000 tokens, assigning each one a probability of occurring. The word that actually follows in the training corpus, “cheese”, receives a probability. The loss function measures the distance between that predicted probability and the truth. The model adjusts its parameters to raise the probability of “cheese” in this context. It repeats the operation billions of times, across trillions of tokens.
Yann Dubois, now a researcher at OpenAI where he co-leads the Post-training Frontiers team and contributed to o1, o3 and GPT-5 Thinking, put it unambiguously in 2024, while finishing his doctorate at Stanford under Percy Liang and Tatsunori Hashimoto. In the lecture he gave that year on building large language models, he states a fundamental mathematical equivalence: minimising that loss amounts to maximising the likelihood of the text. In other words, the model is trained to produce whatever is statistically most probable.
Consider what that implies.
“Statistically most probable” is, by definition, whatever most resembles the average of the training corpus. If the model has seen a billion paragraphs opening with “In a world of constant change”, it assigns that opening a high probability. If the phrase “it is important to note” appears in hundreds of millions of corporate texts, the model reproduces it. Not because it is accurate, or beautiful, or distinctive. Because it is probable.
This mechanism has a name: regression to the mean. It is a statistical phenomenon, not a design flaw. Asking a raw language model to produce distinctive content means asking it to work against its own objective function. It is as absurd as asking a thermometer to heat the room.
Note a counter-intuitive point Dubois raises with his students: what machine learning courses teach about overfitting, the idea that a model trained too long ends up memorising instead of generalising, does not apply to language models. The more you train them, the larger you build them, the more data you give them, the better they get. No plateau has been observed empirically at current scale. The consequence: regression to the mean is not a side effect you could remove by adjusting the training. It is constitutive of the model, and it strengthens as the model grows.
Singularity, by definition, is a deviation from the mean. It is the mathematical opposite of what the model optimises.
Proof two: why RLHF amplifies the smoothing instead of correcting it
You might object: ChatGPT, Claude and Gemini are not raw models. They have been aligned. They went through RLHF, reinforcement learning from human feedback. Surely that step corrects the blandness.
It makes it worse. Here is why.
RLHF works as follows. After pre-training, the model generates two answers to the same question. Human annotators pick the one they prefer. A reward model learns to imitate those preferences. The language model is then adjusted to maximise that preference score. This is what turns a model that completes text into an assistant that answers questions. It is the step that took GPT-3, known to researchers, and made ChatGPT, known to everyone.
The problem, documented by the researchers themselves, is that human annotators carry biases. One of the most powerful is a length bias: they prefer long, structured answers that look considered. So the model learns to produce length. Dubois points this out explicitly: if you find that ChatGPT systematically answers with too many words, bullet lists, disclaimers and “it is important to note that”, this is a direct consequence of RLHF, not a bug.
Worse still: annotation noise is massive. In experiments run in his own laboratory, Dubois reports that trained human annotators agree with each other in only around two thirds of cases on which of two answers is “better”. In other words, a third of RLHF’s training signal is human disagreement. To stabilise learning despite that noise, the model converges mechanically on what commands consensus, which is to say on surface characteristics annotators do agree about: length, visible structure, measured tone, absence of edge. Not on substance, which measures poorly and divides opinion.
And then it gets worse. Once the bias was identified in humans, it was transferred to language models themselves, because paying humans to annotate is expensive, so the human annotator is replaced by another model. This is called preference distillation. Where a human might correct their bias when confronted with an obvious excess, the model corrects nothing. It optimises. Mechanically. Without end.
The consequence: with each generation of models, annotation biases compound. Stylistic smoothing amplifies. Answers grow longer, structured the same way, more predictable. Singularity erodes with every training cycle. This is not an opinion, it is an observed and published dynamic.
RLHF does not solve the blandness set by the initial objective function. It inherits it, and compounds it.
A more current objection deserves treatment. Since Dubois’s 2024 lecture, the methodological landscape has moved: RLHF is no longer the dominant alignment method. More efficient approaches have taken over. Direct preference optimisation, DPO, removes the intermediate reward model. GRPO and DAPO handle optimisation at scale. Reinforcement learning from verifiable rewards, RLVR, covers tasks where truth is measurable. Reasoning RL trains models to work step by step on verifiable problems. The leading 2026 models — o3, GPT-5 Thinking, DeepSeek-R1, Claude 4 — use combinations of these rather than classic RLHF.
The technical detail matters less than the conclusion: regression to the mean persists across these developments. DPO removes some of RLHF’s costs but remains grounded in annotator preferences, so it inherits their surface biases. RLVR works well on verifiable tasks, mathematics and code, but does not apply to editorial production, where no single measurable truth exists. Reasoning RL improves depth of reasoning, not stylistic singularity. None of these methods addresses the initial objective function of pre-training, which remains the optimisation of statistical likelihood.
Put differently: singularity is not an alignment problem. It is a problem that precedes alignment. No improvement in post-training can, on its own, solve it. The thesis Dubois set out in 2024 about the structuring role of the objective function comes out of two years of methodological progress strengthened, not weakened.
Proof three: what the engineers who build these models actually say
There is a striking gap between the public conversation about generative AI and what its builders say internally. The public conversation is about models: their size, their architecture, their benchmarks. The internal conversation is about data.
In the same Stanford lecture, Dubois gives a figure that should reframe the discussion. On Meta’s Llama team, around 70 people work on building the model. Of those 70, around 15 — more than a fifth — work exclusively on training data: collection, cleaning, filtering, deduplication, domain weighting, source qualification.
Why that ratio? Because the engineers building these models know what the marketing industry has been slow to absorb: a language model’s performance does not come from its architecture, it comes from the quality of what it is given to learn. Dubois puts it plainly:
“What matters in practice is mostly data, evaluation and systems.”
Yann Dubois, Stanford CS229, 2024
Academia over-invests in architecture and algorithms because that is what gets taught and published. Industry knows where real performance is decided.
That lesson transfers word for word to your situation. You are not going to modify the architecture of GPT-5 or Claude. You are not going to change their objective function. You are not going to rewrite their RLHF. But you can act on the one variable that decides the real editorial quality of your output: what you give the model to work with. And more than that: the frame you make it work inside.
This is exactly the shift the best research laboratories made five years ago, and that most marketing departments have not yet made. They stopped looking for performance in the model. They looked for it, and found it, in the infrastructure that feeds the model.
You have the same shift to make.
The implication: singularity is not prompted, it is engineered
If a language model’s objective function is regression to the mean, and if that regression is amplified at every alignment cycle, then editorial singularity cannot be obtained by means internal to the model. It can only be obtained externally: through constraints set upstream that force the model to deviate from its default behaviour.
This is what we call editorial engineering. The term is not cosmetic. It names precisely what we are describing: the layers of infrastructure that turn a generic language model into a producer of content aligned with a brand, a voice, a thesis, a body of work.
That infrastructure rests on four concrete pillars.
A brand voice codified rather than described
The difference is essential. A described voice — “we are warm and professional” — is useless to a model. A codified voice, with its brand statement, values, archetypes, registers, accepted and rejected traits, target emotions, is a grid of constraints the model can interpret. The Brand Voice Framework we use rests on nine pillars precisely because nine is the minimum number of dimensions that lets a voice be reproducible without being reductive.
An editorial system, not merely a calendar
A calendar lists publications. An editorial system defines the rules of production: what gets published, what gets refused, what gets archived, how pieces connect to one another, how they are measured. This is what we call the Content Operating System. Without a system, every piece produced with a language model is an isolated drift. With one, every piece joins a coherent body of work.
Source governance before after-the-fact checking
Language models hallucinate. This is not an occasional defect, it is a predictable consequence of supervised fine-tuning: the model is trained to imitate plausible answers, without knowing whether the underlying information is true. The only industrial defence is upstream source governance: systematic validation of figures, explicit marking of unverified data, categorical refusal to publish without traceability, attention to the EEAT signals of experience and expertise. It is slow, it is costly, it is non-negotiable.
Measuring singularity, not only performance
Classic indicators — impressions, clicks, leads — measure distribution, not singularity. Yet if your content resembles your competitors’, it can perform in visibility while eroding your differentiation. So singularity has to be measured explicitly: diversity of reasoning, rhythmic variability, lexical dispersion, contextual anchoring, originality of the narrative arc. Without that measure built into the editorial value chain, you are steering blind.
These four pillars are not an agency luxury. They are the material conditions that make a language model usable in the service of a brand. Without them, you are not using AI. You are dissolving into it.
What this changes for you, concretely
If the thesis of this article holds, three operational consequences follow.
Stop optimising the wrong link. If you invest in prompt engineering tools to fix a problem that occurs upstream of the prompt, you are spending energy on the wrong variable. The return on that investment will be, by construction, low.
Accept that the lever is editorial, not technological. The performance of your content produced with a language model depends on the quality of the editorial infrastructure that frames that production. This infrastructure is not coded. It is written, governed, maintained. It belongs to editorial engineering, not software engineering.
Measure what actually matters. Distribution is easy to measure. Singularity is less so, and it is what decides your buyability across long B2B sales cycles. Most brands work on their saleability: how hard they can push. Few work on their buyability: how easily a committee can actually say yes. A brand that language models cannot replicate is a brand that survives the commoditisation of generic content.
The question is no longer how to use AI better. It has become: how far is your editorial infrastructure able to impose your singularity on a system that regresses, by construction, towards the mean?
If you cannot answer that precisely, you are losing your editorial sovereignty — your capacity to decide what your brand says, how it says it, and why it deserves to be read.
Main source: Yann Dubois, Stanford CS229, Lecture 12: Building Large Language Models, Stanford University, summer 2024. Yann Dubois is now a researcher at OpenAI, where he co-leads the Post-training Frontiers team and contributed to o1, o3 and GPT-5 Thinking. He completed his doctorate at Stanford in 2025, supervised by Percy Liang and Tatsunori Hashimoto, and is a co-author of the Alpaca project.
Frequently asked questions
Why does content produced with a language model all look alike?
Because that is what the model optimises. A large language model is not trained to write well: it minimises a loss function on the prediction of the next token, the unit of text it manipulates. Minimising that loss amounts to maximising the likelihood of the text, that is, producing whatever is statistically most probable. And the most probable resembles the average of the training corpus. This mechanism has a name in statistics: regression to the mean. The blandness comes from the objective function itself, not from how the model executes it.
Can a better prompt fix the blandness?
No. Rewriting the instruction, asking for something “more human”, breaking the rhythm: these treat the symptom. Singularity is, by definition, a deviation from the mean, and therefore the mathematical opposite of what the model optimises. Asking a raw model for distinctive content means asking it to work against its own objective function. This is the market’s central misreading: investment goes into prompt engineering to address a problem that occurs upstream of the prompt.
Doesn’t RLHF correct this?
It amplifies it. Reinforcement learning from human feedback works like this: the model produces two answers, human annotators pick the better one, a reward model learns to imitate those preferences. Those annotators carry a length bias, so the model learns to produce length. Annotation noise compounds it: trained annotators agree in only around two thirds of cases on which answer is better, so a third of the training signal is human disagreement. The model then converges on the surface characteristics that command consensus: length, visible structure, measured tone, absence of edge. Preference distillation, where a model annotates in place of a human, amplifies the movement further: the model does not correct its bias, it optimises it.
Do the more recent methods solve the problem?
No. RLHF is no longer the dominant alignment method, but those that replaced it move the mechanics without changing the result. Direct preference optimisation, DPO, removes the intermediate reward model while remaining grounded in annotator preferences, so it inherits their surface biases. Reinforcement learning from verifiable rewards, RLVR, trains models on tasks where truth is measurable, mathematics and code, which excludes editorial production. Reasoning RL improves depth of reasoning, not stylistic singularity. None of these touches the objective function of pre-training. The problem precedes alignment.
What can an organisation actually do?
Act on the one variable it controls: the frame the model works inside. This is the shift the best research laboratories made five years ago, when they stopped looking for performance in the model and started looking for it in what feeds the model. On Meta’s Llama team, around 15 of the roughly 70 people building the model work exclusively on training data. Transposed to a brand, that gives four pillars: a voice codified rather than described, an editorial system rather than a calendar, source governance upstream rather than checking after the fact, and a measure of singularity rather than of performance alone.