Preventing Robo Debt Happening Again

Preventing Robo Debt Happening Again

The evidence given by a former member of the AAT is pretty damning

We can’t claim that mechanising the Social Services legislation as a way of avoiding the Four Pieces Limit will prevent the behaviour described. What we can claim is that creating an authoritative source, where all the meanings of all the words have been clarified and agreed to when the legislation is written, will reduce the possibility of such behaviour occurring in the future. The legislation becomes activatable, so it is no longer just words on a page, the words become nodes and operators in an undirected network – an Active Structure. It can make calculations or deduce conclusions exactly according to the legislation – there is no human involvement by someone insufficiently au fait interpreting the words or trying to turn the words into a program. 

The machine reads the legislation, makes all the necessary connections, and turns it into an active network. The approach is not suitable for handling millions of accounts, but it is suitable for testing theories, handling contested accounts, or debt calculations.

This is not Machine Learning – there is one document to be read, not thousands. The machine has to know the meanings of the words before it started. It is using definitions from a well-known dictionary, extensively curated where the dictionary is inadequate or suffers from circularity.

The techniques of ML, DL, LLM are suited for regurgitation, without any understanding of what the words mean, only that words follow other words on a statistical basis in a particular sort of text. In an Active Structure approach, each word has all the meanings that English gives it – 62 for “set”, 77 for “on”, 82 for “run”. A little tedious to set up, but then visible to anyone who needs help in understanding the legislation – particularly useful where the line staff have little or no legal training.

Robo-debt Problems

Robo-debt Royal Commission

(based on a report in The Saturday Paper, by Rick Morton, published December 10)

The RRC heard evidence that the person in charge of Robo-debt had never read the Social Services Act, which is quite clear that payments should be based on a particular fortnight, and not averaged over a longer time, as Robo-debt sought to do. Several people strongly suggested this was so, making the robo-debt scheme unlawful, but each time they were overridden. The opinion of a consultant was sought, by colouring the meanings of words, as one means of discrediting internal dissent. The Ombudsman’s efforts in response to concerned politicians and Social Service groups were thwarted, by allowing the people being reviewed to alter the report on their activities. It took three years for the question to reach the Solicitor-General, who immediately declared it unlawful, with several suicides and much distress among impoverished people having occurred in the interim.

What to do? A person keeping up to date as several Acts evolve over time is a large commitment. The commitment can be considerably reduced by having a machine read and “understand” each of the Acts, including their glossaries, their inline definitions of terms, and references to each other, potentially having different meanings for some words across the connection. It would mean that everything is linked together as it should be. If there is dissent over the wording, it can either be resolved, or left open for review. Someone coming cold to the Act in its Active Structure form can then be sure that all connections have been made, and they can confidently rely on what they read in a local area of the Act, and don’t need to read the whole thing carefully, together with its history of revision, to make sure they haven’t missed anything. Such a tool would be a godsend to tyro public servants or politicians.

This sounds easy, with the rise of technology such as Machine Learning, Deep Learning, and Large Language Models. Unfortunately, the goal of having the machine “understand” the meanings of words has largely been abandoned as too hard, and most current methods use statistics – this word is associated with that word or group of words – a method completely unsuited to the complex text found in legislation or specifications.

A Few Colours

Are a Few Colours Really Going to Help?

Humans have a very powerful visual channel why not emphasise that? We can colour the words wordgroups are one obvious candidate we can indicate where there are complex objects in the text objects like

same-institution person-to-person electronic funds transfer instruction

We can do more with a wordgroup that has a definition we can show the complex objects in the definition – like same-institutionor bank account or carry out.

That is perhaps 5% of the words what do we do with the rest?

We could go through the document, work out what is the most common use of the word, then have a colour range for meanings ranging from the most common (white) to the least common (say purple). But this would look like a mess. It would be better only to colour words having an unusual meaning, or words where there is dissension over the meaning, or words in the glossary, or words that are unusual for the persons specialty (so for a lawyer, all the legalese would stay black on white, whereas technology terms would stand out), or where the surrounding text does not give a clear demarcation between a meaning in the readers specialty and other meanings.

As the intention is to deal with large, complex documents, the ability to flit through the document, looking at things that may be hard for the person to understand, should be welcome.

This facility is not a game-changer by itself, but its completion does allow turning overa piece of machinery while it is still made only of words.

What Does “Turning over a piece of machinery” Mean?

It means that words, when connected together, can be pieces of machinery, and when activated, can perform similar (although abstract) functions to the real objects they represent. So a funds transfer instruction can transfer funds from one bank account to another. The text of legislation allows a programmer to write a program, but it also allows a person to simulate the operation in their head, if it is not too complicated, or learn a complex operation, like piloting a plane, at an unconscious level. Having the machine read complex text allows it to simulate what the text says, without being troubled by the Four Pieces Limit. Getting the text cleanis an important step.

Future of Work

Deloitte’s Cameron Pitt recently (November 2022) gave a talk to an IPAA ACT forum about the “future of Work”.

He made several important points (taken from the Mandarin report on the forum):

  • Government leaders needed to support more interdependent ecosystems that include how departments worked together, shared data, moved resources and delivered services to the community.
  • Pitt said there was also a concerted push to build a so-called “collective intelligence” capability.
  • “Ours is all about making our consultants smarter with the use of AI,” he added.
  • He said project-focused work that saw consultants come and go after a few months, with no cultural investment in outcomes and delivery, would need changing to a more meaningful interaction.
  • “Our clients […] have said to us: ‘It’s an ecosystem, it’s an orchestration, we actually don’t want you to deliver services. We want you to own outcomes for the organisation, and work deeply with us to deliver outcomes not deliver a document or an engagement’.”

He didn’t mention the Four Pieces Limit, but it is an important effect, underlying most of what he said.

Departments can’t work well together when their documents use different meanings for words. A “collective intelligence” will require either a considerable dumbing down, or something which can translate a meaning for one department into a meaning for another.

“… smarter with the use of AI”
Nowadays, the original intention of AI, that of embedding intelligence in a machine, has been largely abandoned. Machine Learning (ML, DL, LLM) is about using a pre-programmed device (an ANN or Artificial Neural Network) to connect inputs to outputs and hope that the weightings will allow the correct response. This has been done because it is much easier to do than trying to build intelligence into a machine – you just use a big block of text as the data. This effectively lives in the past, and cannot respond to new events (or even a newly defined term in a piece of new legislation or a specification). An alternative approach, with a new name, is Artificial General Intelligence (AGI), where the structure much more closely emulates the activity in a real neural network, and by doing so, the path from words to outputs is much more easily understood and believable. Needless to say, if an AI application has no understanding of what the words mean, the Four Pieces Limit is irrelevant.

“He said project-focused work that saw consultants come and go after a few months …”
We see value in a project which introduces a means of handling the Four Pieces Limit into a department. It requires someone who is familiar with the complexities of language, and who can embed a culture of having everyone on the same page, ensuring that people with objections because they don’t understand are respected and coached, and people who are diffident about putting forward their valid objections are coaxed into contributing. One or two projects, and the organisation should be able to carry on unassisted.

‘… work deeply with us to deliver outcomes, not deliver a document or an engagement’
Getting down into the details of how words work really is “working deeply”. It won’t be everyone’s cup of tea, but if it can fend off billions wasted because a specification was misunderstood, or legislation that doesn’t do what was intended, it will be worth it. If all it can do is make for an easier working environment, knowing a machine “has your back” in keeping the document “alive” so confusion is minimised, it will be worth it.

Using English

Using English

If the Machine Is Using English, Won’t it Be Too Slow?

It depends what it is doing. A person working with one document on one machine should be fine. But it is not as simple as that – the document may reference other documents, which may use different definitions for some of their words, and this has to be handled when any connection takes place. In other words, a document linking to other documents will need to wake up and communicate with another machine with the other document loaded, the connection taking care of any differences in meaning. This is also the way we see of integrating inputs from different specialties, where the vocabulary of one specialty is foreign to the other – as example, economists and epidemiologists.

Where the intention is to develop strategy in English for detecting fraud or money laundering, this would be done on one machine, and then a fleet of machines deployed using the same instructions to look for instances in a large database of transactions. As an aside, Google “reads” every new document that appears on the Internet, a process that is very slow. To get around this, Google deploys millions of computers on the task. I am not suggesting millions, but a thousand computers come fairly cheap, and can easily switch from one task which requires subtlety and self-awareness of what it is doing, to another.

pastedGraphic.png

Where the machine is tasked with integrating many different and complex models, Climate Change as an example, this would be done with a fleet of machines, each managing an area – say meteorology, oceanography, agriculture, emissions, resilience – and communicating with an integration model, which communicates with each of them in their specific technical jargon, and rolls up the result into a more easily understood description.

What to do about the four pieces limit

What to Do about the Four Pieces Limit

This is probably the most pressing problem in software – how to improve the ability of a person to handle complex tasks.

For some tasks, we can take them away from a person – dangerous manoeuvres in military and rescue helicopters, pilot override in commercial jets. It is a blow to the pilot’s ego, but reliability in life-critical situations …

For other tasks, we can’t take the task away, so we have to augment the person’s ability. A person is very good when focusing on a facet of a problem, not so good at keeping a large, complex problem “live” in their head.

Having a machine read the text and hold it “live” has two major advantages.

Everybody can see which definition is being used for every word or wordgroup, and can argue the case that a different definition should be used. A good way to get a tyro up to speed. 

The machine will have built a structure from the text that can be activated, so questions like “What happens if do this” can be easily answered. The machinery the machine has built can be run, and results observed – particularly results where only imaginary data exists.

By playing with a working model of the text, a great deal of insight can be gained before “cutting steel”.