Recycling is NOT good for the world. It is bad for the environment, it is bad for our health, and it is bad for our economy. I鈥檓 not kidding.
These are the words of GPT-2, an artificially intelligent super-troll. It needs just a few words to prompt a hundreds of words long on almost any topic and its creators say it may be too dangerous to release to the public because of potential misuse.
However, these fears are overblown. GPT-2 is no more dangerous than fakery that has gone on聽for years.
Advertisement
GPT-2 was built by OpenAI, a San Francisco-based AI research lab co-founded by Elon Musk. The firm says GPT-2 is able to understand and write English text better than previous systems. The snippet above forms part of a full-page rant and came as a response to this short prompt:
Recycling is good for the world.
NO! YOU COULD NOT BE MORE WRONG!!
GPT-2 can also read a document and answer questions about it. For example, given a couple of paragraphs about the torch relay at the 2008 Olympic Games, it can answer questions such as 鈥淲hat was the length of the race?鈥�, 鈥淲as it longer than previous ones?鈥� and 鈥淒id they visit any notable landmarks?鈥�
The whole system is based on a simple idea: predict the next word in a body of text. When faced with a writing prompt, word by word it guesses what it believes is most likely to follow. Eventually this results in an essay backing up a statement or a response to a question.
To do this the system was trained on 8 million web pages, amounting to 40 gigabytes of data. GPT-2 isn鈥檛 the first AI to be trained this way, but the size and quality of the data set appears to be key to its success.
鈥淚t鈥檚 good but not a major step change,鈥� says Marcus du Sautoy, a mathematician at the University of Oxford who has .
GPT-2 is better at generating text on topics that came up a lot in the training data, including Miley Cyrus, Lord of the Rings and Brexit. It is less convincing when riffing on technical or specialist subjects.
Cherry picking
Another important caveat is that the samples of generated text released by OpenAI were each cherry-picked from several attempts on the same prompt. The recycling rant is the best of 25 attempts.
This is more common with generative systems than people think, says du Sautoy. We tend to focus on the results and overlook the human involvement.
For example, in 2017 the entertainment company Botnik Studios used an AI to automatically generate a short piece of fan-fiction called 鈥溾�� written in J. K. Rowling鈥檚 style. 鈥淚t was hilarious and quite convincing,鈥� says du Sautoy. 鈥淏ut look under the bonnet and there are lots of choices still made by humans.鈥�
All of which makes the claim that the AI was too dangerous to let loose on the world seem hyperbolic. The OpenAI team published a research paper on the work but didn鈥檛 release the full system or the data set that they trained it on.
鈥淪aying it was too dangerous hypes it up a bit and spikes people鈥檚 appetite,鈥� says du Sautoy.
One reason OpenAI gave for its decision is that the system could be used to mass-produce fake news articles or fake product reviews.
Everything the system learned it read online and it is impossible to know how many lies and toxic opinions it has absorbed, says at Queen Mary University of London. The AI has no way to tell fact from fiction.
Yet fake messages aren鈥檛 new. We have been living with spam, bot accounts on social media and malicious copies of genuine websites for years.
We are also now used to fake political messages intended to influence opinions or incite dissent. The quality of such messages often doesn鈥檛 even have to be that good. As Mark Riedl, who has worked on similar technology at the Georgia Institute of Technology in Atlanta : 鈥淚t doesn鈥檛 take a lot to fool people already primed to be fooled.鈥�
