There's a lot to go at here. So below are a couple of strands I’d like to start exploring with this:
The Black Box
An ongoing bugbear of mine is that people need to be clearer about what AI does for us. I'll explored this in more detail in another post but to take the two main preconceptions to consider:
1: "The Decision-Making-Holder-Of-All-Our-Fates"
At the forefront of people’s minds when they think about AI is the decision-making cognitive (probably-in-some-Hollywood-way-malicious) super-brain-black-box. Technology that can handle hundreds of different inputs in multiple formats and somehow combine, parse, make sense of it all, and take action.
This has spawned the very real concern about how we humans understand how machines are making decisions. The tale of Facebook's two AIs that developed their own language (or at least their own form of communication) is an interesting case in point - but has done little to reassure folks on the transparency front.
Regulated businesses - despite all the money, trials and pilots - are explicitly cautious about AI for just this reason. They can’t devolve responsibility for a credit check (in many ways their core IPR anyway) to an algorithm they can’t describe to a regulator. GDPR will merely enhance this – from May 25th this year people will be legally entitled to know the detail of exactly why a decision has been made. The auditability bread-crumb can't get lost in forests or deep-neural networks, and even then, preventing bias in the underlying datasets remains largely a manual effort. So here then, concern is justified.
2: "The Translator-of-unstructured-into-structured-data"
The second is the less exotically marketed but no less clever "set” of tools and approaches that allow us to make sense of unstructured data - natural language processing, speech, video and image recognition. Here AI can recognise and extract key Data Items (name, invoice numbers etc), validate and match it to existing case information and even then pass it to a back end case management system. Traditional OCR techniques are now being amplified and outstripped by increasingly accurate machine vision approaches like those from Antworks and Abbyy.
While the exact algorithms may not be “perceivable” (and in that sense still black-box), we are much more likely to accept them, for the very simple reason we can comprehend and validate output vs input that much more easily. An agent with a human brain can test the mechanical brain's homework in a split second. There's no decision making (aside from recognising that the word "renewal" = the word "renewal") - so no trust issue.
Where trust in this set of AI tools gets hazier is where the software goes further to parse out context and sentiment to try and establish the meaning (or Intent) of a block of input text (or picture or video), and its here that human-in-the-loop reinforcement training is required. Here, confidence gets built over time.
They said WHAT?
The above are chiefly going to concern corporates. However in a world where the truth is increasingly being called into question, amplified by the echo chamber effects of social network algorithms, what we see and hear as individuals for ourselves is going to be increasingly challenged.
Take for example this work from the University of Washington on Video Lip Syncing:
Or this from UCL which can copy handwriting:
And this from Lyrebird AI on vocal synthesis:
https://lyrebird.ai/demo/
(What’s interesting here is that the only input required is a minute of audio. It doesn’t require a library of all possible words).
Google’s latest voice synthesis algorithm is almost indistinguishable from the human spoke word.
https://google.github.io/tacotron/publications/tacotron2/
What you'll notice here (or maybe you won't!) is that the output is much more than just focussing on enunciating a word. Its about prosody and emphasis - how we say words and the natural sounding rhythm of human speech.
So, will we always be able to believe our eyes - or our ears? Chuck in some behavioural science, a dollop of stirring from "foreign actors", and where do we find ourselves?
I'll start exploring some of the ways people are thinking about addressing the issue of trust in upcoming post - and in the meantime, any thoughts hugely appreciated as ever!
Comments