/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

A behind-the-scenes look at how OpenAI's GPT-2 predictive text algorithm works, which can be “fine-tuned” to write phony customer reviews or even news articles

\u003clink href="https://www.newyorker.com/ projects/interactive/2019/191014-seabrook/css/ header_override.css" …

New Yorker John Seabrook

Context & Ripple Effects

This 2019 New Yorker explainer lands mid-arc in OpenAI's text-generation story: GPT-2's ability to be fine-tuned into a phony-review or fake-news engine is exactly the capability that critics at The Gradient soon argued rests on superficial, unreliable knowledge — fluent output without grounded understanding.

The piece also reads as an early warning for what followed: GPT-3 reframed the same architecture as a step toward language understanding, and by the time OpenAI opened consumer creation tools, the abuse pattern this article demonstrated had reappeared at platform scale in the GPT Store's impersonation, copyright-infringing, and jailbreaking GPTs.

First-order effects

  • Review platforms and news publishers are directly exposed: anyone with the model and fine-tuning instructions can mass-produce customer reviews and articles indistinguishable in style from human writing.

Second-order effects

  • As generation gets easier, verification becomes the scarce input — pushing platforms toward provenance and authenticity tooling, and giving critics of these models' reliability (the superficial-knowledge critique) more evidence when outputs mislead.
  • OpenAI's own product decisions absorb the lesson: the guardrail questions around its later consumer surfaces, like the custom GPT builder Altman wants to keep simple, are the same open-access-versus-abuse tradeoff this article surfaced with fine-tuning.

Third-order effects

  • If each capability jump — GPT-2 to GPT-3 to GPT-4 — widens both legitimate use and misuse, the industry's structural response is trust infrastructure: provenance standards, detection, and platform moderation becoming as central to text AI as the models themselves.

The trend: Text generation is scaling from research demo to consumer platform faster than verification mechanisms, making synthetic-content abuse a recurring cost of every OpenAI capability release.

Discussion

  • @newyorker @newyorker on x
    Could a robot replace a New Yorker writer? We fed The New Yorker's archive to an artificial-intelligence writer, which predicts text based on preceding language. Then we asked it to write for us. https://www.newyorker.com/...
  • @slashml @slashml on x
    Can a Machine Learn to Write for the New Yorker? (OpenAI finetunes largest GPT-2 on New Yorker articles) https://www.reddit.com/...
  • @benwallacewells Ben Wallace-Wells on x
    The first snippet of computer-written predicted text is, “By that I mean, it seemed to want to distinguish my feelings from my thoughts. To put it another way, Smart Compose seemed to want to know me.” !!! https://twitter.com/...
  • @michaelluo Michael Luo on x
    Wow. Read this new piece in this week's @newyorker by John Seabrook on predictive text. At the end of every graf, read text that an artificial intelligence predicted would come next. https://www.newyorker.com/...