OpenAI today announced GPT-2, a language model that can write news articles, translate text, and answer questions — all from a single system trained only to predict the next word in a sentence. But the nonprofit research lab is declining to release the full model or its training data, citing the risk of malicious use.

‘We need to perform experimentation to find out what they can and can’t do,’ said Jack Clark, OpenAI’s policy director. ‘If you can’t anticipate all the abilities of a model, you have to prod it to see what it can do.’ The model, trained on 8 million web links scraped from Reddit, can produce convincing fake news, product reviews, and even anti-Semitic rants when given a toxic prompt. The Guardian reports that the system is being called ‘deepfakes for text.’

OpenAI is giving temporary access to the algorithm to a number of media publications, including The Verge, but the full model and its training data remain under wraps. The decision is drawing debate. ‘I have a term for this. The escalator from hell,’ Clark said. ‘It’s always bringing the technology down in cost and down in price.’ The debate is just beginning.