Why your AI sounds like everyone else's

Hey, I’m Isaac 👋 I founded Pistachio, a growth agency working with B2B brands like Atono and Clay to build trust, relationships and loyalty with their current and future customers.

If you want to chat about your own growth strategy, book a free call with me here.

Read time: 5 mins 33 secs

Interested in sponsoring these emails? See our partnership options here.

Through the 1840s and 50s, more than 300,000 people flooded into California to mine gold. Most didn't strike it rich. The real winners were selling picks, shovels, and jeans (Levi Strauss didn't become a legend by digging). That same San Francisco that the gold rush built would, 175 years later, see an almost identical flood of capital into a new and very familiar kind of gold rush.

Whether you think AI is the beginning of something genuinely useful or a very expensive distraction, it's not stopping. The models are already embedded in your workflow, and the more interesting question for marketers isn't whether to use them but whether you understand what's actually happening inside them. Because what's happening inside them has direct implications for your brand, your content, and your ability to stand out.

I'm Bill Kerr, founder of Athyna and the Open Source CEO newsletter. Isaac invited me to guest post here because there's a story happening in the engine room of AI that most marketers haven't heard, and probably should.

My brother just had his first child. His name is Sidney. And Sidney, who is not yet a year old, is the best analogy I've found for how a language model actually develops. There are three distinct stages a model goes through, and they map surprisingly well to how a person makes their way through the world.

The first is pre-training. This is where the model reads the internet, basically all of it. Wikipedia, Reddit, textbooks, news articles, GitHub repos, brand blogs, ad copy, agency decks, obscure forums. Trillions of words, consumed in weeks by a data center the size of a small town. Just as Sidney learns that 'dada' is usually followed by something warm, the model is mapping patterns in language. By the end of this stage, it's essentially someone who has read every book in the library but has never had a conversation. Technically knowledgeable. Contextually limited.

The second stage is fine-tuning. Here, the model gets shaped into something that actually feels like a product. It shows examples of good behavior. When a human asks X, a good response looks like Y. It learns to follow instructions, format its outputs, and aim to be genuinely useful rather than just technically correct. Think of it as the difference between someone who knows every accounting theory and someone who can actually file your taxes.

The third stage is post-training, and this is where things get interesting.

This is where real, living, breathing humans sit down and evaluate the model's outputs. They compare responses. They flag what's off. They push the model toward something better.

'This response was accurate, but the tone was completely wrong.'
'The grammar is perfect. The meaning is lost.'
'This copy follows the brief. It just sounds like it was written by a committee.'

The model learns from this feedback in a loop (generate, evaluate, improve, repeat), a process called RLHF or Reinforcement Learning from Human Feedback. What it really means is humans in the loop, making the machine less generic, one correction at a time. Which humans, and how good they are at the task, turns out to matter enormously.

Here's the insight that should make every marketer pay attention. At the first two stages, every major AI model trains on largely the same data. ChatGPT, Claude, Gemini and Grok have all consumed the same public internet, optimised toward the same benchmarks, and passed through similar processes. The gap between them has mostly closed. They perform more similarly than the marketing around each would have you believe.

For your brand, this means something equally important. The generic outputs you're getting from AI tools are a feature of the system, not a bug. The model was trained on everything, which means it learned to sound like everything, including your competitors.

The AI failure stories that get the most attention tend toward the dramatic. The hallucinated legal citation. The medical misdiagnosis. The maths error that cascades through a financial model. But the failures that matter most for marketing are quieter and more insidious. They're the ones that don't get flagged as wrong, because technically, they aren't.

A global brand runs an AI-translated campaign from English to Japanese. The grammar is flawless. The idioms, however, have been taken literally. The humor has become humorless. The emotional register reads like a legal disclaimer. Research shows AI translation can lose over 47% of contextual meaning and more than half of emotional nuance. The words were right. The meaning evaporated.

A B2B SaaS brand uses AI to scale its content output. The articles are accurate, well-structured, and correctly optimized for search. They also sound identical to every other B2B SaaS brand doing the same thing. The distinctive voice that took years to develop gets averaged out by a model trained to produce content that performs, rather than content that's distinctively theirs.

A growth team runs AI-generated ad variations. The copy follows every best practice in the model's training data. The CTRs are acceptable. But the ads don't build anything. No memory. No recognition. No compounding brand equity.

This is what model convergence looks like at the marketing level. Outputs that are technically correct but impossible to tell apart from your competitor's.

Post-training and data labeling is an industry growing faster than almost anything else right now, and most marketers have never heard of it. The companies helping AI labs make their models smarter, through expert human evaluation at scale, are in short supply and growing at a remarkable pace.

Mercor, a company that pivoted into AI post-training, grew recurring revenue from $1M to $500M in 17 months. Benchmark's Mitch Lasky called it the fastest-growing company in history. They're not alone. With Meta's 2025 acquisition of Scale AI prompting OpenAI and Google to immediately pull their contracts, a vacuum opened up, and a wave of companies moved in to fill it.

The insight driving all of it is the same one that drives the whole industry. The most valuable commodity in AI right now is trained humans who can train machines, not compute, code, or data.

According to Deel's 2026 Global Hiring Report, AI trainer roles are growing at 283% year on year, cross-border. It's one of the fastest-growing roles in tech, with more than 40% of the talent coming from outside the US. The picks and shovels of this gold rush are people, specifically domain experts with the judgment to evaluate outputs that generalists can't.

Two practical conclusions. The first is about your content. AI tools are not a neutral surface. They have a gravitational pull toward the average, and the average is, by definition, undifferentiated. The brands that use AI well are the ones that understand this and treat their voice, tone, and positioning as proprietary inputs, not optional style guides. Your brand's accumulated knowledge is an asset. Feed it in deliberately. Audit what comes out. Don't let the model flatten years of brand-building into something that could have been produced by anyone.

The real winners of this gold rush won't be the companies building the models but the ones solving the human layer, those with access to the deepest pool of domain expertise who can deploy it fastest. Only this time, the picks and shovels are people.

And that's it! You can subscribe to my newsletter here, follow me on Twitter and LinkedIn, and also don’t forget to check out Athyna while you’re at it.

If you enjoyed this post or know someone who may find it useful, please share it with them and encourage them to subscribe: brandchemistry.co/p/ai-post-training