Feeds

DARPA seeks 'Machine Reading' AI auto-analysis bot

Human analysts' unanimous view - it'll never work

Choosing a cloud hosting partner with confidence

The US military is seeking revolutionary new AI software which would be able to read text - and so effectively do research - in the same way that humans do. The so-called "Machine Reading" ware would initially be used for such tasks as automated military-intelligence analysis, but it would have wide consequences in civilian life as well.

As regular readers would expect, the auto-reader notion comes from DARPA. DARPA has the usual Pentagon revolving doors connecting it to the military-industrial complex and the ivory tower; it also has an express lift to the castle in the sky, a secret tunnel to the secure wing of the scientists' insane asylum and a special shooting range situated on both sides of the Grand Canyon - where only long shots can be taken.

DARPA explain the problem thus:

The U.S. military frequently faces impediments to stability and reconstruction operations in a new location due to the lack of understanding of the local situation. Similarly, strategic assessment of a foreign nation’s science and technology base involves the continuous assessment of technical articles, bibliographies, conference agendas, etc. This information is often available on the World Wide Web, and some tools to assist this analysis are available, but the process would be significantly enhanced by a system that could directly analyze the information found in these text sources. The same reasoning could be equally valuable if applied to other types of open-source intelligence analysis, including assessing military readiness and posturing; political speeches, actions, and more obscure “messages”; economic trends and sentiments; and propaganda from terrorist groups and even their hidden web-based communications.

In other words, rather than having thousands of expensive human intel analysts surfing the web and putting together reports, why not just have some computers?

DARPA graphic explaining how you evaluate AI auto-reader software

It's as simple as a funnel stuck into your head

Well, it seems that it's all a matter of reasoning. The DARPA people say that computer software, even software dubbed "AI", doesn't do well at pulling together information in different kinds of document the way a human can - and indeed struggles to tie together information in single documents.

The Pentagon brainiacs say that this is because computers are best at "formal" language. While they can work with text, if the text is written in "natural" human languages - for instance standard English prose, as opposed to data in fields - the machines don't do very well.

That's where "Machine Reading" AI-ware would come in. According to DARPA, what's needed is simple to describe in natural language:

A successful Machine Reading Program will be enabling all knowledge encoded as natural text to be combined with the power of AI reasoning methodologies, which will unleash a wide variety of new AI applications ranging from intelligent bots to personal tutors. For example, all of the text in the World Wide Web will become available for automating the monitoring and analysis of technological and political activities of nations; plans, rhetoric, and activities of transnational organizations; and scientific discovery within various disciplines.

However, there's another thing which doesn't lend itself to natural language apart from computers - that is, federal contract awards. There has to be a formal-language way to decide which competing Machine Reader systems are better than others, and indeed much of the DARPA solicitation (pdf) is devoted to accomplishing that very feat (see pic, which makes the methods clear even to the meanest intelligence).

Normally, of course, one looks at AI systems from DARPA and it's clear that if they work the consequences would be awful: humanity's few survivors scuttling like rats in the ruins of civilisation, etc.

In this case, the casualties might be more tightly focused. Essentially, anyone whose job involves primarily reading text on the internet and producing more text from it would be out of work. Graduate students in the humanities subjects, intelligence analysts, think-tank people, quangocrats ... and about 90 per cent of journalists.

Plainly a mad notion; it'll never happen. ®

Internet Security Threat Report 2014

More from The Register

next story
Bladerunner sequel might actually be good. Harrison Ford is in it
Go ahead, you're all clear, kid... Sorry, wrong film
Musicians sue UK.gov over 'zero pay' copyright fix
Everyone else in Europe compensates us - why can't you?
I'll be back (and forward): Hollywood's time travel tribulations
Quick, call the Time Cops to sort out this paradox!
Megaupload overlord Kim Dotcom: The US HAS RADICALISED ME!
Now my lawyers have bailed 'cos I'm 'OFFICIALLY' BROKE
Euro Parliament VOTES to BREAK UP GOOGLE. Er, OK then
It CANNA do it, captain.They DON'T have the POWER!
Forget Hillary, HP's ex CARLY FIORINA 'wants to be next US Prez'
Former CEO has political ambitions again, according to Washington DC sources
prev story

Whitepapers

Go beyond APM with real-time IT operations analytics
How IT operations teams can harness the wealth of wire data already flowing through their environment for real-time operational intelligence.
A strategic approach to identity relationship management
ForgeRock commissioned Forrester to evaluate companies’ IAM practices and requirements when it comes to customer-facing scenarios versus employee-facing ones.
Why CIOs should rethink endpoint data protection in the age of mobility
Assessing trends in data protection, specifically with respect to mobile devices, BYOD, and remote employees.
Reg Reader Research: SaaS based Email and Office Productivity Tools
Read this Reg reader report which provides advice and guidance for SMBs towards the use of SaaS based email and Office productivity tools.
Seattle children’s accelerates Citrix login times by 500% with cross-tier insight
Seattle Children’s is a leading research hospital with a large and growing Citrix XenDesktop deployment. See how they used ExtraHop to accelerate launch times.