AI research firm holds back on release of text generator over fears it could create fake news

AI research firm holds back on release of text generator over fears it could create fake news

A Silicon Valley research company will not release a new tool that can generate coherent text using artificial intelligence, including news articles, over fears it could be used to create fake news.

But, non-profit firm Open AI has released a smaller version of its GPT-2 text generator for researchers along with a technical paper on its tool.

The tool, which the firm says was trained on a dataset of 8m web pages (equal to 40GB) to predict the next word in a sentence can produce “realistic and coherent” text following a prompt written by a human.

Open AI said that, while tools like its own could be used for unsupervised translation and other societal benefits, AI language tools could also be used to create fake news as well as automated spam and abusive content.

The firm provided samples of articles its bot had written after being fed short prompts written by humans.

These included articles about the election of a resurrected John F Kennedy, scientists finding a herd of unicorns in the Andes and a thief stealing a train carriage with nuclear material on board.

In a blog post about the text generator, Open AI said: “Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT-2 along with sampling code.

“We are not releasing the dataset, training code, or GPT-2 model weights.”

It added: “We are aware that some researchers have the technical capacity to reproduce and open source our results.

“We believe our release strategy limits the initial set of organisations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems.”

Open AI said it would review its limited release strategy in six months time.

The AI research firm was co-founded by Tesla chief executive Elon Musk three years ago and announced it was conducting “large-scale experiments” on Microsoft’s Azure platform in November 2016.

Musk left Open AI last year, saying in a tweet: “I had to focus on solving a painfully large number of engineering & manufacturing problems at Tesla (especially) & SpaceX.

“Also, Tesla was competing for some of same people as OpenAI & I didn’t agree with some of what Open AI team wanted to do. Add that all up and it was just better to part ways on good terms.”

Open AI said that its text generator was able to produce decent samples “about 50 per cent of the time” when prompted with popular subjects, such as Brexit, but was not as effective when fed “highly technical” content.

Other faults in the tool included repetitive text, “unnatural topic switching” and making errors such as writing about fires underwater.

Picture: Pexels



Press Gazette's must-read weekly newsletter featuring interviews, data, insight and investigations.