The Guardian GPT-3 Op-Ed: What It Wrote and Editors Changed

The Guardian’s 2020 GPT-3 opinion article was generated by OpenAI’s language model from prompts and then edited by Guardian journalists before publication. GPT-3 produced several possible drafts and arguments; the editors selected and reshaped material, checked the claims, and made clear that the published piece was an edited collaboration rather than an untouched machine output.

How the Guardian experiment began

The article appeared on 8 September 2020 under the headline “A robot wrote this entire article. Are you scared yet, human?” It was an experiment in whether a large language model could produce readable opinion writing, not a test of whether a machine had beliefs or independent intentions.

OpenAI had released GPT-3 earlier that year. The Guardian asked the model to write an article explaining why humans should not fear artificial intelligence. Editors supplied instructions, including the desired argument and tone, and generated multiple passages. The project then involved choosing and revising the model’s output to make a coherent newspaper article.

The headline’s “entire article” phrasing was attention-grabbing, but the process was more complicated than asking a chatbot one question and printing its first answer. The Guardian’s accompanying explanation described editorial work and the generation of several versions. The published text was the result of both model output and human judgment.

What GPT-3 wrote

The chosen article argued that artificial intelligence should not be treated as an imminent threat to humanity. It presented the system as a tool built by people, said that it lacked independent desires, and suggested that fears about AI often reflect human concerns rather than the technology’s own intentions.

Its language was fluent and rhetorically organized. It used familiar opinion-writing moves: it acknowledged public anxiety, offered reassurance, and urged readers to focus on human choices. Yet those qualities do not mean GPT-3 had formed a considered view. The model generated likely sequences of words in response to its prompt, drawing on patterns learned from text. Its apparent position was shaped by the request it received.

That distinction matters. The article’s argument was not evidence that GPT-3 understood the future of AI or could independently assess risks. It was an example of a model producing persuasive prose that fitted a requested viewpoint. Different instructions could have led to very different passages.

What the editors did

Guardian editors did not simply publish a single raw response. They generated several versions and worked with the resulting material, selecting passages and editing them into a readable article. Editorial involvement included shaping structure and flow, removing or revising material, and ensuring the final piece met the publication’s standards.

The editors also had to contend with a practical problem: a language model can produce plausible-sounding claims that are inaccurate, repetitive, vague, or contradictory. Human review is therefore important even when the goal is to showcase machine-generated writing. A smooth sentence is not the same as a verified statement.

The Guardian’s publication included an account of the process, helping readers understand that the article had been generated with GPT-3 and edited by people. That disclosure was part of the experiment: readers were invited to consider both the writing and the boundary between automated drafting and human editorial responsibility.

• Prompting: Editors set the subject and the argument the model was asked to make.
• Generation: GPT-3 produced multiple possible drafts or sections rather than one definitive article.
• Selection and revision: People chose and edited material to create the published version.
• Disclosure: The newspaper explained that the text came from GPT-3 and had undergone editorial work.

Why GPT-3 could sound convincing

GPT-3 was a large language model trained to predict likely next words from context. Given a prompt, it could continue in a style that resembled essays, news commentary, or other forms of writing found in its training material. This made it capable of producing paragraphs that read naturally, even without human-like understanding.

Fluency can create a misleading impression of knowledge. A model may state a claim confidently without knowing whether it is true, and it may imitate the form of reasoning without independently checking its premises. In the Guardian experiment, the reassuring tone followed from the assignment: the model was specifically asked to make the case that people should not fear AI.

It is also useful to separate the model from the wider technology debate. The article addressed fear of AI in broad terms, but it was not a technical risk assessment. Questions about bias, misinformation, privacy, labor, accountability, or the uses of powerful systems require evidence and analysis beyond a generated opinion piece.

What the article did—and did not—prove

The experiment showed that GPT-3 could generate substantial, readable prose that editors could turn into a publishable opinion article. It also demonstrated that humans remained central to the result: they defined the task, evaluated outputs, selected text, and took responsibility for publication.

It did not prove that GPT-3 was conscious, had an opinion, or could reliably write journalism without oversight. Nor did it establish that AI was harmless. The model was asked to argue for reassurance, so its output should be read as a response to a framing choice, not as an impartial verdict on the risks of AI.

A common misconception is that “AI wrote it” means every word appeared exactly as published, untouched by humans. In this case, that is not an accurate description. Another is that editing turns generated material into fact-checked reporting automatically. Editing can improve clarity and structure, but claims still need appropriate verification, and responsibility rests with the publisher.

Why the experiment mattered to journalism

The article arrived as newsrooms were beginning to explore how language models might assist with drafting, summarizing, or other text-heavy tasks. Its value was partly practical and partly theatrical: it put a visible example in front of readers and exposed the questions that come with using generated text in a newspaper.

Those questions include who is accountable for errors, whether readers should be told when a model contributed, and how much editing is enough. The Guardian’s disclosure and process note addressed some of these concerns, though the experiment itself was not a universal policy for journalism. Different uses—such as a clearly labeled creative exercise versus a factual news report—carry different expectations.

The lasting lesson is not that a language model can replace an editor. It is that generated text may be persuasive enough to require careful review. The model supplied drafts; humans supplied the assignment, judgment, and publication decision.

Frequently Asked Questions

When did the Guardian publish the GPT-3 article?
The Guardian published “A robot wrote this entire article. Are you scared yet, human?” on 8 September 2020.

Did GPT-3 write the article without human help?
No. GPT-3 generated multiple pieces of text in response to prompts, and Guardian editors selected and edited material for the published article.

What was the article’s main argument?
It argued that people should not fear AI as an independent force with its own desires, emphasizing that humans build and direct the technology.

Did the experiment show that GPT-3 understood what it wrote?
No. It showed that GPT-3 could produce fluent text suited to a prompt, not that it had consciousness, beliefs, or human-like understanding.

Was the article proof that AI is safe?
No. It was an opinion-writing experiment framed around reassurance, not a technical safety study or comprehensive evaluation of AI risks.

Related

• Raspberry Pi 5: Reasons to Upgrade?
• Sony’s Forgotten Fitness Gadget That Predicted Wearables
• Apple TV 4K Benefits: Is It Worth Switching?