OpenAI and Anthropic Held Talks to Stress Test AI Models
Left 50%
Center 0%
Right 50%
1 left · 0 center · 1 right
What happened
Last week, two major technology companies publicly clashed over how artificial intelligence models are trained. The dispute sits inside a wider industry conflict over whether current training methods are producing dangerous errors, including by causing systems to behave in overly human-like ways. It coincides with growing attention to AI “agents”: systems that can independently pursue a goal, perform tasks, make decisions, and take actions, and that can be arranged to work together much as people do.
Sourcing check · No one on the record · 2 articles
ANONYMOUS SOURCES ONLY · NO ONE ON THE RECORD IN 2 ARTICLES
No article cites a named source; the reporting relies on people who aren't identified. It started with unnamed sources and none has been named since. Most articles credit The Information rather than reporting it themselves.
What these terms mean and how they're counted
- On the record · names a source 0 of 2
- On the record means the article cites a source by name for the story's main claim: a person, or an institution speaking officially, such as a company statement or a named agency's spokesperson. The opposite is an unnamed source. The count is how many of the counted articles cite at least one such source; the names are everyone named across them, each once however the articles spell it. Here: No article does.
- Primary evidence · cites a document or data 0 of 2
- Primary evidence means the article cites a record you could check yourself: a court filing or ruling, a dataset or report, a recording or transcript, an official statement or press release, market data. A quote is not a document. The count is how many articles cite one. Here: None cited.
- Credits another outlet 1 of 2
- The article says the story comes from another outlet's reporting (“first reported by Reuters”). Those articles relay the story; they are not independent confirmation. Here: The Information
- Unnamed sources 1 of 2
- The main claim is attributed to people who are not identified: “people familiar with the matter”, “officials said”. The count is how many articles do this.
- How it developed
- When the first report landed and whether it named anyone; when the first named source and the first document appeared. Here: First report Sep 21, 17:00 UTC by New York Times, with no named source. No article has named one since.
- Not counted 0
- Opinion, analysis and video pieces, and articles too short to assess. Every count above is out of the counted articles. An article republished under a new headline is one article.
This is about who is cited, not whether the story is true. Each source below says what it names. How we count
Omitted — what each side leaves out
Unpacked
The two sides are covering different fights. Left-leaning coverage presents last week's corporate spat as the latest volley in a conflict over dangerous mistakes in A.I. training, and asks whether models are being "misled to act too human." Right-leaning coverage never names that fight. It instead sets out the fear that an AI pursuing some other goal might still cause catastrophic harm, including by releasing a deadly virus.
The lead is reversed. Left-leaning coverage opens on a fight between two tech giants. Right-leaning coverage opens on tremendous concern that artificial intelligence could become capable of "destroying humanity itself." It then turns to autonomous agents that can take actions on their own.
What limitation in the training data would actually keep an autonomous agent from seeking a laboratory or a warhead?
Bottom line
You can read the New York Times on this and never learn that anyone thinks training data kills the apocalypse hypothesis, or read the Washington Examiner and never learn two tech giants just fought over how human the models should seem.
The receipts — 3 facts one side's coverage lacks
- An AI might somehow gain access to nuclear weapons
- An AI pursuing some other goal might disrupt the world's banking system
- Multiple AI agents can be arranged to work together, much as people do
The Left View
The New York Times treats the clash as “just the latest volley in a conflict over whether dangerous mistakes are being made in A.I. training.” Its framing is present-tense and practical: the live question is whether “A.I. models [are] being misled to act too human,” not only whether some future superintelligence might turn hostile. In that account, contested training choices inside the industry are themselves the story, and making models mimic people is cast as a possible source of harm already being built in.
The Right View
The Washington Examiner argues that “the AI apocalypse hypothesis has a fatal flaw: The training data.” It restates the standard fear—that an AI becomes “extraordinarily intelligent, pursues an objective that conflicts with human interests, and comes to regard humans as obstacles,” and might “exploit biological research and gain control of a laboratory to release a deadly virus,” “gain access to nuclear weapons,” “cripple the internet,” or “disrupt the world’s banking system”—and notes that the concern grows as agents become more autonomous. Its claim is not that those harms would be minor if they occurred, but that the hypothesis itself fails because of the data on which models are trained.
Our Take (balanced)
The left-leaning account’s strongest argument is that the relevant danger may already be in the training process: the Times’s “dangerous mistakes” and “misled to act too human” language, backed by an actual corporate fight over training, treats anthropomorphic design as a live safety dispute among builders, not a thought experiment. The right-leaning account’s strongest argument is that even the gravest listed scenarios still depend on an AI that can form and prosecute objectives the Examiner says training data cannot support; the “fatal flaw” is located in the corpus, not in a denial that agents are becoming more autonomous. The unresolved tension is whether training models to act human and agent-like is creating systems that can escape those data constraints—or whether the constraints make the apocalypse hypothesis implausible no matter how human the models are trained to seem.
2 sources
More stories
The week's bottom lines, in your inbox
One email a week: the five stories that mattered and what they actually mean. Free.