Jumping in to assuming the existence of a Generally Accepted AI is a useful exercise for seeing if we would even want to go there.
Getting there will, of course, be Hard, since it is a characteristic of AI broadly considered (machine learning and perhaps anything more statistically complex than multiple regression) that - with certain notable exceptions - it is very difficult to interrogate how the model you've trained is doing what it's doing. And there is plenty of research showing how it is very feasible to smuggle in easter eggs and back doors at model building time in ways that are very hard to detect. You might think you could perhaps get round this by having a government and/or university project build it, rather than the private sector. But that's just replacing an obvious problem (the incentive to make it say good things about their accounts and bad ones about their rivals) with a less obvious one (the people building it will be much more cheaply bought than if they worked at a commercial outfit).
And! It will have the same problem that quant analysis has. It's one thing to build a model that backtests really well. It's quite another to deploy it successfully in a competitive, adversarial context.
not so sure of this because I think the context can be made non-adversarial, in the same way in which the accountancy profession imperfectly protects itself against conflicts of interest today
Ok yeah quite possibly, but the new and likely rapidly-changing opportunities for opaque shenanigans it would create mean we will need nimble adjustments to the regulatory and social context. Probably not quite as exciting an environment as crypto trading though, I'd guess. Although at least in that context, the blockchain records the transactions in the open so it is possible to work out what just happened after the fact.
yep - getting people to cuddle the computer (sorry, "embrace technology") is a real hard problem and I agree with all your points about the *social* problems which everyone is hoping will be dissolved by the sheer attractiveness of the tech. (which tbf did happen with the internet and Microsoft Excel).
In the Higher Education sector we used to have a JACS code that we returned to government to describe the subject of every single module in the curriculum. Someone had to assess each module to decide the right code. A decade or so ago I tried to persuade colleagues that it would be far easier to just upload the description we give to students in the module catalogue which would not only save us the bother of doing the coding, but also be a far richer resource for research and analysis. Everyone hated that idea too, and I think it is a smaller and more tractable version of this idea.
My problem was that I was trying to persuade the people who did the coding that I could save them effort, when in fact knowing how to do that work to best advantage was their professional expertise. I might have gotten further trying to persuade the people who want to consume the data. I don't really know anything about finance but it seems to me that you will never persuade people to give up the power to tell their own stories about their own businesses. You need to persuade someone else to take that power away from them.
I very much like the idea of creating multiple abstracted summaries of the same underlying data, for different audiences and purposes. But I am really uneasy about there being much role for artificial neural networks in this.
Look at what people have been able to do with adversarial-example manipulations of LLMs (e.g. [https://arxiv.org/abs/2307.15043]). Now imagine somebody's bonus riding on getting the machine to say good things about the company's performance... I get that you're not necessarily advocating LLMs, specifically, but big neural networks seem pretty generically subject to adversarial examples. Worse, I don't think we really have much more of a grasp on the reasons why than we did in 2014.
I mean ... if you consider a human brain to be a neural network, we have plenty adversarial examples of accounts being falsified because somebody wanted a bonus! I'm going to write another post on this because I am obviously way into it, but the issue here is exactly that of "professional standards" needing to be reproduced at the level of the machine, and another illustration that many of the problems of AI are replicating existing problems of business ethics and of politics ...
Specifically with respect to adversarial datasets - I guess we would be thinking here about someone putting transactions in the ledger to encourage the neural net to catch irrelevant features - I am not so sure that's the problem, because adversarial datasets, as far as I read the (abstracts of the) literature. tend to be like other kinds of fraud - they only work when tested from one particular perspective.
If the new paradigm for accounting was that you would have a trusted API and trusted set of algorithms assembling the transactions for different purposes, I think it would be that much more difficult to generate a general-purpose adversarial dataset. Although I suppose that if you did solve that problem, you'd be able to cause significantly more harm.
This is like something out of a capitalist version of an Iain M. Banks novel. As in, if you imagine friendly superhuman Minds that still allow humans a bunch of agency, but for whatever reason there are still corporations running around, then yes, this is the kind of thing the Minds might do to help the human corporate auditors out. I have a harder time imagining it working if the AIs are still programs that humans are designing for particular purposes. It just seems like so much could go so wrong in such horrible ways. Could be a failure of imagination on my part, though.
shared set of accounts == communists blockchain == decisions outsourced to virtual market of pretending agents = arbitrage destroyed == narcissism vanquished or accelerated
Coming from a career on the buy-side of private equity and funds management I was always struck by the difference between those who see accounts as a record and those (such as myself) that primarily see accounts as a narrative.
Accounts as a record is the biggest mental framework issue, the record should be the log of changes, the active working coalface is the cash register, and its efforts to sell, (the bookkeeper is the middleware) for some reason reporting to the emperor with the dead letter past is more important than the interface with customers. This informs the design too much.
(The same is true in the industry where I work in Museums, where the legal work of the registrar and the museum registry is designed to produce reports but doing actual work in the software is absolute hell (eponymously named Vernon CMS as designed by an alcoholic Accountant from NZ)
Accounting is one part technical moving of numbers in a coherent and transparent fashion to keep track of transactions, two parts applied philosophy on what constitutes value and worth, and one part exploitation of gray areas of legality. Predictive summarisers can help with the first part, have little to no value on the second and…. well, you can probably train a model on borderline criminality, but it would be a borderline criminal model, along the lines of criminal lawyers in Breaking Bad.
Always interesting, Dan. I am an accountant and ex-auditor and was a much better general ledger auditor than I was a financial statements auditor, so leaving auditors to concentrate on the actual numbers sounds delightful!
One of the challenges, as I see it, is that lots of the stuff that a business does to generate value isn’t really captured by a number, and then text based disclosures are almost always completely boiler plate. I’m not sure how AI can help explain what those judgements driving the numbers actually ARE but that would help comparability more!!
If you had a Generally Accepted AI, the brightest minds in the business would be devoted to reverse-engineering it--getting from open accounts to confidential ledger entries.
Firstly my understanding of AI is a little dated - I flirted with genuine LLM's back in early '00's with the aim of conquering Finnish to Hungarian translation (failed, but the company still exists having escaped Russia sometime after Crimea (most recent not 1856)
I think the idea is fine for companies in the 1880's which produced widgets - widget gets sold, you get paid with which you pay your costs and, if applicable taxes, net profit and cash flow are largely the same thing
And today - revenue is an assumption and doesn't equal cash. Costs are fairly real, taxes were devised by fairies on speed and then the real fun starts; leases, acquisition accounting, anything other than plain vanilla debt.
Can you imagine letting an AI bot loose on a set of assumptions for a monte carlo simulation of embedded equity in a debt instrument.
Notwithstanding, for AI to work each number needs to be tagged or associated so that the machine can understand it - we do that already, it's called a chart of accounts - not a brilliant system but without it, or another tagging system, you don't just have unstructured data, you just have numbers.
I can see a great role in parsing all my peers ESG reports and reproducing them as if they were mine. Cynical moi - never
I may not know enough to ask this question intelligently, but when you say "important decisions would be delegated to the ai” aren't you really saying "important decisions are being delegated to the people that train the AI". Those people must have the requisite knowledge of accounting, the purposes to which the data will be put (the needs of the users of the data), and also be alert to the biases implicit in the data used in the training. Or am I wrong?
in what I've been reading in non-technical accounts of the promise and perils of ai, the importance of training seems to be all too often mentioned in passing or not at all.
Excellent and thought provoking. I am an old accountant at heart but I don't hate this. A few thought it triggers:
You are talking about published accounts. In the real world these are not used to run businesses. Companies rely on management accountants made to different rules to suit the focus of management. Different classifications but rarely different time periods.
Virtually all accounting within companies includes some comparison to budgets and forecasts these time periods can vary a bit. Seems like a strength of LLMs or similar might be improving the value of this element.
Getting to both management and statutory accounts is a huge spreadsheet miasma. Just making this process cleaner and more efficient would make a decent business case (and massively improve the auditability of the final numbers).
Measuring intangibles like brand, customer loyalty (not the same thing) and the real value of investment in software is the big hole in current numbers as you say. This probably means finding a way to how people and assets are actually used. Some proxies exist but would be genuinely interesting to see how AI interpreted data to answer this kind of question. Lots of the reality is never recorded I suspect.
“The problem is that it would all involve actually handing over responsibility and actually delegating important decisions to the AI. My view is that if we’re not going to do that, what’s the point of it?” --> this is a Stafford Beer level provocation. What are you doing to promote the book and how can I help?
Some people think the future of dating will involve "have your avatar chat to my avatar to see if it's worth meeting". As you note it is much more likely to happen in corporate courtships where "have your M&A banker talk to my M&A banker" is less esoteric. And the success of the match depends less on actual chemistry and more on information compatibility.
You mention accountants and analysts, but investment bankers and management consultants might have AI models of companies they follow, to simulate mergers or restructurings. As you say companies are information processing units, and this idea applies the new information technology to their management.
A little while ago I wanted to know how many sheets of A4 paper I could get from a roll of paper I was considering buying so I Googled the question. The AI generated solution was too big by a factor of 10. It seems large language models are a poor way to learn arithmetic let alone accounting.
Jumping in to assuming the existence of a Generally Accepted AI is a useful exercise for seeing if we would even want to go there.
Getting there will, of course, be Hard, since it is a characteristic of AI broadly considered (machine learning and perhaps anything more statistically complex than multiple regression) that - with certain notable exceptions - it is very difficult to interrogate how the model you've trained is doing what it's doing. And there is plenty of research showing how it is very feasible to smuggle in easter eggs and back doors at model building time in ways that are very hard to detect. You might think you could perhaps get round this by having a government and/or university project build it, rather than the private sector. But that's just replacing an obvious problem (the incentive to make it say good things about their accounts and bad ones about their rivals) with a less obvious one (the people building it will be much more cheaply bought than if they worked at a commercial outfit).
And! It will have the same problem that quant analysis has. It's one thing to build a model that backtests really well. It's quite another to deploy it successfully in a competitive, adversarial context.
not so sure of this because I think the context can be made non-adversarial, in the same way in which the accountancy profession imperfectly protects itself against conflicts of interest today
Ok yeah quite possibly, but the new and likely rapidly-changing opportunities for opaque shenanigans it would create mean we will need nimble adjustments to the regulatory and social context. Probably not quite as exciting an environment as crypto trading though, I'd guess. Although at least in that context, the blockchain records the transactions in the open so it is possible to work out what just happened after the fact.
yep - getting people to cuddle the computer (sorry, "embrace technology") is a real hard problem and I agree with all your points about the *social* problems which everyone is hoping will be dissolved by the sheer attractiveness of the tech. (which tbf did happen with the internet and Microsoft Excel).
In the Higher Education sector we used to have a JACS code that we returned to government to describe the subject of every single module in the curriculum. Someone had to assess each module to decide the right code. A decade or so ago I tried to persuade colleagues that it would be far easier to just upload the description we give to students in the module catalogue which would not only save us the bother of doing the coding, but also be a far richer resource for research and analysis. Everyone hated that idea too, and I think it is a smaller and more tractable version of this idea.
My problem was that I was trying to persuade the people who did the coding that I could save them effort, when in fact knowing how to do that work to best advantage was their professional expertise. I might have gotten further trying to persuade the people who want to consume the data. I don't really know anything about finance but it seems to me that you will never persuade people to give up the power to tell their own stories about their own businesses. You need to persuade someone else to take that power away from them.
I very much like the idea of creating multiple abstracted summaries of the same underlying data, for different audiences and purposes. But I am really uneasy about there being much role for artificial neural networks in this.
Look at what people have been able to do with adversarial-example manipulations of LLMs (e.g. [https://arxiv.org/abs/2307.15043]). Now imagine somebody's bonus riding on getting the machine to say good things about the company's performance... I get that you're not necessarily advocating LLMs, specifically, but big neural networks seem pretty generically subject to adversarial examples. Worse, I don't think we really have much more of a grasp on the reasons why than we did in 2014.
I mean ... if you consider a human brain to be a neural network, we have plenty adversarial examples of accounts being falsified because somebody wanted a bonus! I'm going to write another post on this because I am obviously way into it, but the issue here is exactly that of "professional standards" needing to be reproduced at the level of the machine, and another illustration that many of the problems of AI are replicating existing problems of business ethics and of politics ...
Specifically with respect to adversarial datasets - I guess we would be thinking here about someone putting transactions in the ledger to encourage the neural net to catch irrelevant features - I am not so sure that's the problem, because adversarial datasets, as far as I read the (abstracts of the) literature. tend to be like other kinds of fraud - they only work when tested from one particular perspective.
If the new paradigm for accounting was that you would have a trusted API and trusted set of algorithms assembling the transactions for different purposes, I think it would be that much more difficult to generate a general-purpose adversarial dataset. Although I suppose that if you did solve that problem, you'd be able to cause significantly more harm.
This is like something out of a capitalist version of an Iain M. Banks novel. As in, if you imagine friendly superhuman Minds that still allow humans a bunch of agency, but for whatever reason there are still corporations running around, then yes, this is the kind of thing the Minds might do to help the human corporate auditors out. I have a harder time imagining it working if the AIs are still programs that humans are designing for particular purposes. It just seems like so much could go so wrong in such horrible ways. Could be a failure of imagination on my part, though.
shared set of accounts == communists blockchain == decisions outsourced to virtual market of pretending agents = arbitrage destroyed == narcissism vanquished or accelerated
Coming from a career on the buy-side of private equity and funds management I was always struck by the difference between those who see accounts as a record and those (such as myself) that primarily see accounts as a narrative.
Accounts as a record is the biggest mental framework issue, the record should be the log of changes, the active working coalface is the cash register, and its efforts to sell, (the bookkeeper is the middleware) for some reason reporting to the emperor with the dead letter past is more important than the interface with customers. This informs the design too much.
(The same is true in the industry where I work in Museums, where the legal work of the registrar and the museum registry is designed to produce reports but doing actual work in the software is absolute hell (eponymously named Vernon CMS as designed by an alcoholic Accountant from NZ)
Accounting is one part technical moving of numbers in a coherent and transparent fashion to keep track of transactions, two parts applied philosophy on what constitutes value and worth, and one part exploitation of gray areas of legality. Predictive summarisers can help with the first part, have little to no value on the second and…. well, you can probably train a model on borderline criminality, but it would be a borderline criminal model, along the lines of criminal lawyers in Breaking Bad.
Always interesting, Dan. I am an accountant and ex-auditor and was a much better general ledger auditor than I was a financial statements auditor, so leaving auditors to concentrate on the actual numbers sounds delightful!
One of the challenges, as I see it, is that lots of the stuff that a business does to generate value isn’t really captured by a number, and then text based disclosures are almost always completely boiler plate. I’m not sure how AI can help explain what those judgements driving the numbers actually ARE but that would help comparability more!!
If you had a Generally Accepted AI, the brightest minds in the business would be devoted to reverse-engineering it--getting from open accounts to confidential ledger entries.
Some thoughts.
Firstly my understanding of AI is a little dated - I flirted with genuine LLM's back in early '00's with the aim of conquering Finnish to Hungarian translation (failed, but the company still exists having escaped Russia sometime after Crimea (most recent not 1856)
I think the idea is fine for companies in the 1880's which produced widgets - widget gets sold, you get paid with which you pay your costs and, if applicable taxes, net profit and cash flow are largely the same thing
And today - revenue is an assumption and doesn't equal cash. Costs are fairly real, taxes were devised by fairies on speed and then the real fun starts; leases, acquisition accounting, anything other than plain vanilla debt.
Can you imagine letting an AI bot loose on a set of assumptions for a monte carlo simulation of embedded equity in a debt instrument.
Notwithstanding, for AI to work each number needs to be tagged or associated so that the machine can understand it - we do that already, it's called a chart of accounts - not a brilliant system but without it, or another tagging system, you don't just have unstructured data, you just have numbers.
I can see a great role in parsing all my peers ESG reports and reproducing them as if they were mine. Cynical moi - never
I may not know enough to ask this question intelligently, but when you say "important decisions would be delegated to the ai” aren't you really saying "important decisions are being delegated to the people that train the AI". Those people must have the requisite knowledge of accounting, the purposes to which the data will be put (the needs of the users of the data), and also be alert to the biases implicit in the data used in the training. Or am I wrong?
in what I've been reading in non-technical accounts of the promise and perils of ai, the importance of training seems to be all too often mentioned in passing or not at all.
Excellent and thought provoking. I am an old accountant at heart but I don't hate this. A few thought it triggers:
You are talking about published accounts. In the real world these are not used to run businesses. Companies rely on management accountants made to different rules to suit the focus of management. Different classifications but rarely different time periods.
Virtually all accounting within companies includes some comparison to budgets and forecasts these time periods can vary a bit. Seems like a strength of LLMs or similar might be improving the value of this element.
Getting to both management and statutory accounts is a huge spreadsheet miasma. Just making this process cleaner and more efficient would make a decent business case (and massively improve the auditability of the final numbers).
Measuring intangibles like brand, customer loyalty (not the same thing) and the real value of investment in software is the big hole in current numbers as you say. This probably means finding a way to how people and assets are actually used. Some proxies exist but would be genuinely interesting to see how AI interpreted data to answer this kind of question. Lots of the reality is never recorded I suspect.
National statistical accounts could also do with a refresh, there are working groups as we speak
“The problem is that it would all involve actually handing over responsibility and actually delegating important decisions to the AI. My view is that if we’re not going to do that, what’s the point of it?” --> this is a Stafford Beer level provocation. What are you doing to promote the book and how can I help?
Some people think the future of dating will involve "have your avatar chat to my avatar to see if it's worth meeting". As you note it is much more likely to happen in corporate courtships where "have your M&A banker talk to my M&A banker" is less esoteric. And the success of the match depends less on actual chemistry and more on information compatibility.
You mention accountants and analysts, but investment bankers and management consultants might have AI models of companies they follow, to simulate mergers or restructurings. As you say companies are information processing units, and this idea applies the new information technology to their management.
A little while ago I wanted to know how many sheets of A4 paper I could get from a roll of paper I was considering buying so I Googled the question. The AI generated solution was too big by a factor of 10. It seems large language models are a poor way to learn arithmetic let alone accounting.