New AI fake text generator may be too dangerous to release, say creators (theguardian.com).
OpenAI, an nonprofit research company backed by Elon Musk, says its new AI model, called GPT2 is so good and the risk of malicious use so high that it is breaking from its normal practice of releasing the full research to the public in order to allow more time to discuss the ramifications of the technological breakthrough.
Dubbed “deepfakes for text” (Blog 19th Feb. 2018) and perhaps actually being the “summoning the demon” the somewhat melodramatic Musk has suggested humanity’s “biggest existential threat”, but not because it has any Skynet aspirations to take over, rather the danger being the misinformation it may be made to speak on behalf of humans with agenda in a post-truth world:
At its core, GPT2 is a text generator. The AI system is fed text, anything from a few words to a whole page, and asked to write the next few sentences based on its predictions of what should come next.
So if fed a few paragraphs of a news story about Brexit or Orange Don refusing to be budge on his wall it would output plausible prose drawing details replete with quotes presumably drawn from other news stories circulating:
It was trained on a dataset containing about 10m articles, selected by trawling the social news site Reddit for links with more than three votes.
And there’s the rub, being how to educate a machine scanning stories on Web 2.0 to distinguish between real and fake news with the latter getting as many likes and up votes.
To show what that means, OpenAI made one version of GPT2 with a few modest tweaks that can be used to generate infinite positive—or negative—reviews of products. Spam and fake news are two other obvious potential downsides, as is the AI’s unfiltered nature. As it is trained on the internet, it is not hard to encourage it to generate bigoted text, conspiracy theories and so on.
And on the subject of AI potentially up to no good:
Google’s ‘deceitful’ AI assistant to identify itself as a robot during calls (theguardian.com).
The feature, called Google Duplex, was demonstrated at the company’s I/O developers’ conference on Tuesday. It is not yet a finished product, but in the two demos played for the assembled crowd, it still managed to be eerily lifelike as it made bookings at a hair salon and a restaurant.
Although doubtful as to its ability to pass a serious Turing Test (Wikipedia), still sparking concerns that it could mislead those being briefly called into thinking such bookings etc. are being made by a real person rather than a machine.
The generated voice not only sounds extremely natural, but also inserts lifelike pauses, um-ing and ah-ing, and even responding with a wordless “mmm-hmm” when asked by the salon worker to “give me one second”.
But what about the human who before would have made that call?
The issue not being addressed of course being that it will no longer require a real person to make that booking with, although economical for companies, in the UK another poor soul forced into zero-hour employment contract fragility or forced to food bank with the DWP’s pitiful bureaucratically bungled Universal Credit rollout. Wasn’t the Jetson-esque robot worker future supposed to benefit us all rather than just further increase inequality with ever richer robot worker companies?
Despite Finland ending its experiment (theguardian.com) is there really a solution for the future that does not involve some kind of non-stigmatised universal basic income? Indeed, with 40% of British public in support and Labour to include a pilot in next manifesto (independent.co.uk, Aug. 2018).
Recent/related stories
- Facebook shuts down robots after they invent their own language (Latest Picks 1st August 2017)
- Salvation Army raises alarm over doll ‘sexbots’ (Blog 7th July 2017)
- Robots will satisfy everyday sex needs in the future (Blog 20th December 2016)
Share this story