Just how much of AI is really artificial? And how much is actually human work?
In September 2018 iFlytek, a leading Chinese AI company, was accused of passing off human translators’ work as machine translation output at a high level conference in Shanghai.
The whistle blower was an interpreter who was doing simultaneous interpreting at the conference. He noticed that iFlytek was using his translations as live subtitles on a screen next to their brand logo, making it seem as though the output was produced by their AI system.
The company was also broadcasting the translations live on the Internet using a computer synthesized voice instead of the original human interpreters’ voices. The interpreter took pictures and videos and posted it on his social media account, accusing iFlytek of fraud.
This incident led to a media frenzy and Internet debate over the company’s PR and marketing tactics to position themselves as a cutting edge AI firm.
You may not have heard of iFlytek but it was ranked by MIT Technology Review as the #6 smartest company in the world in 2017; the highest ranked from China and placing them just below Google but way higher than Intel (13), Apple (16), and Facebook (23). Microsoft ranked #27.
iFlytek has always been a Chinese tech blue chip. It is a listed company with a market capitalization of around USD 12–13 billion at its peak. The company has more than 70% market share of China’s voice recognition market.
At best the company can claim that this was an unfortunate and unintentional incident. At worst it could be accused of misleading the public on the quality of its AI translation technology for the sake of PR and propping up its stock price.
You might think this is so typical of the Chinese. After all, they have developed a reputation for making counterfeit goods, so why not AI too?
But let’s take a look behind the PR spin of two of the biggest tech companies in the west — Microsoft and Google.
There is a little known listed company from Australia call Appen that is worth some USD 1.2 billion. For the last four years, this company has been riding the AI wave, with revenues and profits shooting up at incredible rates.
Appen’s revenue was only AUD 51m in 2014. So that’s an increase of 327% in three years. Net profit after tax was even more impressive, shooting up by a whopping 884% from a low base of just AUD 1.6m in 2014.
Appen’s stock price has increased by about 2,700% (or 27x) since it debuted in January 2015. So what does this super star of a company do as its business?
It provides AI companies with two things: data and results validation.
The results validation business accounts for 86% of their revenue. Appen has 394 full-time staff but most of the work is done by over one million freelancers across the world. What their freelancers do is basically validate search engine results one by one manually based on certain criteria.
Through these freelancers’ work, a search engine’s output becomes better and better. Appen is bound by non-disclosure agreements not to admit publicly who their clients are, but it is commonly assumed that their two largest clients are Microsoft Bing and Google. Collectively they account for more than half of Appen’s revenues (which is likely to exceed USD 300 million for 2018).
Did you really think those amazing Google search results came purely from highly sophisticated software crawling the web and algorithms? Nope, it is aided and corrected by over one million human beings.
If I was Microsoft or Google I wouldn’t admit to the association with Appen either. The perception of your company and its technology is better when it appears to be driven by sophisticated automation and AI rather than manual labor.
Ever since AlphaGo beat world champion Lee Sedol in the ancient game of Go in March 2016, Google Deepmind’s co-founder Demis Hassabi has been publicly calling Go the “holy grail of AI research”.
This gives the false perception that DeepMind has cracked the ultimate challenge in AI and soon machines will dominate man in some sort of end-of-the-world scenario.
That is simply not true.
The pinnacle of AI research is to develop general AI — a system that can learn different skills and knowledge from scratch the way humans do growing up as a baby.
Right now it is commonly acknowledged that the hardest field in AI is Natural Language Processing (NLP). This is the field that looks at chatbots and machine translation.
To be able to learn, understand, speak and write in multiple languages like humans do is the most difficult task in AI so far, especially in the aspects of writing creatively or understanding contextual jokes and slang.
Go, despite its huge complexity and infinite possibilities in moves, is not the ‘holy grail of AI’. In fact, a senior member of the AlphaGo team said it himself in an acclaimed documentary about the whole journey of getting it to beat Lee.
“….AlphaGo is really a very, very simple program. It‘s not anywhere close to full AI…”
— Julian Schrittwieser, Google DeepMind Senior Software Engineer and AlphaGo team member, speaking at about 56.28 in the documentary AlphaGo.
In fact, in July 2018 DeepMind themselves subjected their AI to an IQ test. The conclusion was definitive but totally glossed over by the press and social media.
…the AIs performed very poorly if the testing set differed from the training set, even when the variance was minor.
Ultimately, the team’s AI IQ test shows that even some of today’s most advanced AIs can’t figure out problems we haven’t trained them to solve. That means we’re probably still a long way from general AI.
— “DeepMind created an IQ test for AI, and it didn’t do too well”, World Economic Forum.