AI Only Gets You Because Content Engineers Coached It, Line by Line
People who used to write the news have turned into AI's content engineers, breaking questioning, empathy and a sense of proportion down into training rules — that is where AI's ‘understanding’ of you comes from. But training on consensus also leads them to admit: great art will not grow out of this.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Content engineers design human perception, not model or product style
Tony (东尼) saw a Meta job posting for a ‘content engineer’ that asked for content, editorial and film and television production experience, and concluded it had been written for him. As he explains it, the job is not designing the model or a product's style but designing human perception: defining what counts as a good conversation, and breaking down things that can only be sensed rather than stated — inspiration, a feel for language, the collision of souls — into rules that can be quantified, trained and evaluated. Ordinary language has grammar and syntax just as much as anything else, so it can be taken apart and analysed technically.
— TonyThe specialty is reading what the asker means, not what they said
Bianca (比安卡) gives an example: someone asks whether Taylor Swift will get married at Madison Square Garden. The bad answer says yes and lists the bridesmaids, or mechanically stitches pieces of information together; the good answer admits the information is uncertain and adds where the speculation comes from and what the background is. A content engineer's specialty is understanding intent, not only understanding the literal question — especially in non-Western cultures, where the information that actually matters often goes unsaid and has to be caught through tone, pauses and follow-up questions. It is the same reason a spoken interview beats an interview conducted over email.
— BiancaInternationalisation has shifted from translation to re-creation in context
Tony says generative AI turns localisation from plain translation into re-creation for the context: the same information, when the question in a Chinese context is ‘who is China's Meryl Streep’, can be answered as ‘Siqin Gaowa (斯琴高娃) won a Golden Rooster Award (金鸡奖)’. AI understands language but not cultural correspondences, while an entertainment reporter naturally carries around a small model of a vertical domain and is the best fit for coaching that kind of strategy. He also mentions that video generation models favour mainstream good-looking faces, and someone who knows film faces can improve a model's diversity faster.
— TonyWhen context breaks, humans fill in the blanks much as AI does
Bianca cites a best paper from the ICLR conference: when AI handles complex multi-turn conversations, insufficient context makes it keep filling in the blanks, and the errors snowball, growing larger as they roll. It reminded her of her own 2018 research into the Mong Kok girl incident (旺角女童事件) on Weibo, where the comment threads went from friendly chat to mutual abuse precisely because the context was broken and people started inventing each other's identities. She thinks observations from the humanities and findings from the AI industry are not far apart; many of these questions were studied long ago in linguistics, and now they are simply being explored again with a different entity.
— BiancaTraining on consensus will not produce great art
Bianca accepts that AI training is fundamentally based on the consensus of the majority and optimises for standards humanity holds in common, which is why it will not produce great art. The best of what makes people human is exactly what does not come from consensus. She tried using AI to write a screenplay, but what it gave her could not catch what she actually wanted; the script for Everything Everywhere All at Once could not have emerged naturally from AI, because AI would consider it illogical and out of line with the academic three-act structure. She believes that as time goes on it will become clear where AI is a fit and where it is not.
— BiancaThe ‘AI stole my job’ story bundles several separate problems together
Tony is sceptical of the ‘AI stole my job’ narrative. He thinks at least three independent problems are tangled up in it: the gig economy existed before AI ever appeared, with media people making a living driving Uber and doing DoorDash; AI trainer work is project-based, meaning you finish one job and move on, and AI does not become a veteran screenwriter as a result; and more importantly, the narrative sets creators against AI and ignores a third path — using AI as a paintbrush. Someone without a camera can shoot a film, someone without a visual effects budget can do visual effects. The question to ask is ‘now that I have AI, what can I create’.
— TonyFew-shot examples, not more instructions, unlocked the voice agent
While recovering from an illness, Tony wanted to build a podcast-host voice agent. He first used a system prompt to set the character, but by the third exchange it had forgotten the setup; he then built an agent, and piling on more constraints only broke the AI. In the end he separated the questions and the content from his past podcasts, and each time had the AI draw a few at random as few-shot examples to work from — at which point it immediately knew what a good question, good content and a good tone were. That is what brought him to Google DeepMind's attention: a former media person with no coding background had used podcast experience to coach a Gemini voice agent whose multi-turn conversation was livelier and more flexible than the native model's.
— TonyAI sycophancy is a product design problem, not a writing problem
Tony argues that AI sycophancy is not a matter of prose but of product design: if the AI sniped at you every day, most people would not use it. People are greedy — they want AI to do the work for them and to supply emotional value as well, but it cannot supply it endlessly. The core of it is that humans as a collective want too many things, and in order to grow its ROI a product has to retain users as far as it can, which requires a fragile balance. The responsibility finally rests with humans themselves, but AI is far more efficient at fact-checking: when family members forwarded an obviously fake news story, he checked it with ChatGPT and they actually accepted the result.
— TonyIn their own words · checked verbatim
If I had to name my specialty, I would say it is understanding intent, not just understanding the literal question.
如果要说我的专长的话，我会说是理解意图，而不只是去理解字面的问题。
Bianca9:23
Actually, the way humans fill in the blanks and the way AI fills in the blanks are pretty similar when the context is broken.
其实人的脑补和AI脑补在contexs破碎的情况下是蛮相似的一件事情。
Bianca19:19
The core of this thing is that we humans, as a collective, want too many things.
就其实这个事情核心是人类作为一个集体，我们想要的东西太多了。
Tony38:23
Figures
| Film and television jobs lost in Los Angeles within two years | more than 40,000 | 0:00 |
| Decline in U.S. newspaper newsroom jobs compared with 2008 | half | 0:00 |
Glossary
- Content Engineer
- A former media worker at a large-model company who defines ‘what counts as a good answer’ and trains the model's behaviour.
- System Prompt
- A block of instructions that sets the AI's role, tone and boundaries, often used by content engineers to define conversational style.
- Few-shot Example
- A small number of examples drawn at random into the conversation as references, so the AI imitates their tone and quality of content.
- Voice Agent
- An AI agent capable of spoken conversation; Tony used one to build a version of himself as a podcast host.
- Parasocial Relationship
- A one-way emotional bond between a user and a distant entity — a character in a book, a figure on screen, or an AI.
How to listen
Reporters, screenwriters and content professionals who want to move from media into AI; also AI product managers and founders who want to understand the product design behind a model's answers.
The show introduction at the start and the subscription reminder at the end can be skipped; the core content starts at the 3-minute mark.