Tuesday, September 26, 2023

Creativity and Recursivity

Prompted by @jjn1's article on AI and creative thinking, I've been reading a paper by some researchers comparing the creativity of ChatGPT against their students (at an elite university, no less).

What is interesting about this paper is not that ChatGPT is capable of producing large quantities of ideas much more quickly than human students, but that the evaluation method used by the researchers rated the AI-generated ideas as being of higher quality. From 200 human-generated ideas and 200 algorithm-generated ideas, 35 of the top-scoring 40 were algo-generated.

So what was this evaluation method? They used a standard market research survey, conducted with college-age individuals in the United States, mediated via mTurk. Two dimensions of quality were considered: purchase intent (would you be likely to buy one) and novelty. The paper explains the difficulty of evaluating economic value directly, and argues that purchase intent provides a reasonable indicator of relative value.

The paper discusses the production cost of ideas, but this doesn't tell us anything about what the ideas might be worth. If ideas were really a dime a dozen, as the paper title suggests, then neither the impressive productivity of ChatGPT nor the effort of the design students would be economically justified. But the production of the initial idea is only a tiny fraction of the overall creative process, and (with the exception of speculative bubbles) raw ideas have very little market value (hence dime a dozen). So this research is not telling us much about creativity as a whole.

A footnote to the paper considers and dismisses the concern that some of these mTurk responses might have been generated by an algorithm rather than a human. But does that algo/human distinction even hold up these days? Most of us nowadays inhabit a socio-technical world that is co-created by people and algorithms, and perhaps this is particularly true of the Venn diagram intersection between college-age individuals in the United States and mTurk users. If humans and algorithms increasingly have access to the same information, and are increasingly judging things in similar ways, it is perhaps not surprising that their evaluations converge. And we should not be too surprised if it turns out that algorithms have some advantages over humans in achieving high scores in this constructed simulation.

(Note: Atari et al recommend caution in interpreting comparisons between humans and algorithms, as they argue that those from Western, Educated, Industrialized, Rich and Democratic societies - which they call WEIRD - are not representative of humanity as a whole.)

A number of writers on algorithms have explored the entanglement between humans and technical systems, often invoking the concept of recursivity. This concept has been variously defined in terms of co-production (Hayles), second-order cybernetics and autopoiesis (Clarke), and being outside of itself (ekstasis), which recursively extends to the indefinite (Yuk Hui). Louise Amoore argues that, in every singular action of an apparently autonomous system, then, resides a multiplicity of human and algorithmic judgements, assumptions, thresholds, and probabilities.

(Note: I haven't read Yuk Hui's book yet, so his quote is taken from a 2021 paper)

Of course, the entanglement doesn't only include the participants in the market research survey, but also students and teachers of product design, yes even those at an elite university. This is not to say that any of these human subjects were directly influenced by ChatGPT itself, since much of the content under investigation predated this particular system. What is relevant here is algorithmic culture in general, which as Ted Striphas's new book makes clear has long historical roots. (Or should I say rhizome?)

What does algorithmic culture entail for product design practice? For one thing, if a new product is to appeal to a market of potential consumers, it generally has to achieve this via digital media - recommended by algorithms and liked by people (and bots) on social media. Thus successful products have to submit to the discipline of digital platforms: being sorted, classified and prioritized by a complex sociotechnical ecosystem. So we might expect some anticipation of this (conscious or otherwise) to be built into the design heuristics (or what Peter Rowe, following Gadamer, calls enabling prejudices) taught in the product design programme at an elite university.

So we need to be careful not to interpret this research finding as indicating a successful invasion of the algorithm into a previously entirely human activity. Instead, it merely represents a further recalibration of algorithmic culture in relation to an existing sociotechnical ecosystem. 

Louise Amoore, Cloud Ethics: Algorithms and the Attributes of Ourselves and Others (Durham and London: Duke University Press 2020)

Mohammad Atari, Mona J. Xue, Peter S. Park, Damián E. Blasi and Joseph Henrich, Which Humans? (PsyArXiv, September 2023) HT @MCoeckelbergh

David Beer, The problem of researching a recursivesociety: Algorithms, data coils and thelooping of the social (Big Data and Society, 2022)

Bruce Clarke, Rethinking Gaia: Stengers, Latour, Margulis (Theory Culture and Society 2017)

Karan Girotra, Lennart Meincke, Christian Terwiesch, and Karl T. Ulrich, Ideas are Dimes a Dozen: Large Language Models for Idea Generation in Innovation (10 July 2023)

N Katherine Hayles, The Illusion of Autonomy and the Fact of Recursivity: Virtual Ecologies, Entertainment, and "Infinite Jest" New Literary History , Summer, 1999, Vol. 30, No. 3, Ecocriticism (Summer, 1999), pp. 675-697

Yuk Hui, Problems of Temporality in the Digital Epoch, in Axel Volmar and Kyle Stine (eds) Media Infrastructures and the Politics of Digital Time (Amsterdam University Press 2021)

John Naughton, When it comes to creative thinking, it’s clear that AI systems mean business (Guardian, 23 September 2023) 

Peter Rowe, Design Thinking (MIT Press 1987)

Ted Striphas, Algorithmic culture before the internet (New York: Columbia University Press, 2023)

See also:  From Enabling Prejudices to Sedimented Principles (March 2013)

Thursday, March 9, 2023

Technology in use

In many blogposts I have mentioned the distinction between technology as designed/built and technology in use.

I am not sure when I first used these exact terms. I presented a paper to an IFIP conference in 1995 where I used the terms technology-as-device and technology-in-its-usage. By 2002, I was using the terms "technology as built" and "technology in use" in my lecture notes for an Org Behaviour module I taught (together with Aidan Ward) at City University. With an explicit link to espoused theory and theory-in-use (Argyris).

Among other things, this distinction is important for questions of technology adoption and maturity. See the following posts 

I have also talked about system-as-designed versus system-in-use - for example in my post on Ecosystem SOA 2 (June 2010). See also Trusting the Schema (March 2023).

Related concepts include Inscription (Akrich) and Enacted Technology (Fountain). Discussion of these and further links can be found in the following posts:

And returning to the distinction between espoused theory and theory-in-use. In my post on the National Decision Model (May 2014) I also introduced the concept of theory-in-view, which (as I discovered more recently) is similar to Lolle Nauta's concept of exemplary situation.

Richard Veryard, IT Implementation or Delivery? Thoughts on Assimilation, Accommodation and Maturity. Paper presented to the first IFIP WG 8.6 Working Conference, on the Diffusion and Adoption of Information Technology, Oslo, October 1995. 

Richard Veryard and Aidan Ward, Technology and Change (City University 2002) 

Saturday, February 18, 2023

Hedgehog Innovation

According to Archilochus, the fox knows many things, but a hedgehog knows one big thing.

In his article on AI and the threat to middle class jobs, Larry Elliot focuses on machine learning and robotics.

AI stands to be to the fourth industrial revolution what the spinning jenny and the steam engine were to the first in the 18th century: a transformative technology that will fundamentally reshape economies.

When people write about earlier waves of technological innovation, they often focus on one technology in particular - for example a cluster of innovations associated with the adoption of electrification in a wide range of industrial contexts.

While AI may be an important component of the fourth industrial revolution, it is usually framed as an enabler rather than the primary source of transformation. Furthermore, much of the Industry 4.0 agenda is directed at physical processes in agriculture, manufacturing and logistics, rather than clerical and knowledge work. It tends to be framed as many intersecting innovations rather than one big thing.

There is also a question about the pace of technological change. Elliott notes a large increase in the number of AI patents, but as I've noted previously I don't regard patent activity as a reliable indicator of innovation. The primary purpose of a patent is not to enable the inventor to exploit something, it is to prevent anyone else freely exploiting it. And Ezrachi and Stucke provide evidence of other ways in which tech companies stifle innovation.

However the AI Index Report does contain other measures of AI innovation that are more convincing.

 AI Index Report (Stanford University, March 2022)

Larry Elliott, The AI industrial revolution puts middle-class workers under threat this time (Guardian, 18 February 2023)

Ariel Ezrachi and Maurice Stucke, How Big-Tech Barons Smash Innovation and how to strike back (New York: Harper, 2022)

Wikipedia: Fourth Industrial Revolution, The Hedgehog and the Fox

Related Posts: Evolution or Revolution (May 2006), It's Not All About (July 2008), Hedgehog Politics (October 2008), The New Economics of Manufacturing (November 2015), What does a patent say? (February 2023)

Sunday, January 22, 2023

Reasoning with the majority - chatGPT


#chatGPT has attracted considerable attention since its launch in November 2022, prompting concerns about the quality of its output as well as the potential consequences of widespread use and misuse of this and similar tools.

Virginia Dignum has discovered that it has a fundamental misunderstanding of basic propositional logic. In answer to her question, chatGPT claims that the statement "if the moon is made of cheese then the sun is made of milk" is false, and goes on to argue that "if the premise is false then any implication or conclusion drawn from that premise is also false". In her test, the algorithm persists in what she calls "wrong reasoning".

I can't exactly recall at what point in my education I was introduced to propositional calculus, but I suspect that most people are unfamiliar with it. If Professor Dignum were to ask a hundred people the same question, it is possible that the majority would agree with chatGPT.

In which case, chatGPT counts as what A.A. Milne once classified as a third-rate mind - "thinking with the majority". I have previously placed Google and other Internet services into this category.

Other researchers have tested chatGPT against known logical paradoxes. In one experiment (reported via LinkedIn) it recognizes the Liar Paradox when Epimenides is explicitly mentioned in the question, but apparently not otherwise. No doubt someone will be asking it about the baldness of the present King of France.

One of the concerns expressed about AI-generated text is that it might be used by students to generate coursework assignments. At the present state of the art, although AI-generated text may look plausible it typically lacks coherence and would be unlikely to be awarded a high grade, but it could easily be awarded a pass mark. In any case, I suspect many students produce their essays by following a similar process, grabbing random ideas from the Internet and assembling them into a semi-coherent narrative but not actually doing much real thinking.

There are two issues here for universities and business schools. Firstly whether the use of these services counts as academic dishonesty, similar to using an essay mill, and how this might be detected, given that standard plagiarism detection software won't help much. And secondly whether the possibility of passing a course without demonstrating correct and joined-up reasoning (aka "thinking") represents a systemic failure in the way students are taught and evaluated.

See also

Andrew Jack, AI chatbot’s MBA exam pass poses test for business schools (FT, 21 January 2023) HT @mireillemoret

Gary Marcus, AI's Jurassic Park Moment (CACM, 12 December 2022)

Christian Terwiesch, Would Chat GPT3 Get a Wharton MBA? (Wharton White Paper, 17 January 2023)

Related posts: Thinking with the Majority (March 2009), Thinking with the Majority - a New Twist (May 2021), Satanic Essay Mills (October 2021)

Wikipedia: ChatGPT, Entailment, Liar Paradox, Plagiarism, Propositional calculus 

Wednesday, August 17, 2022

Discipline as a Service

In my post on Ghetto Wifi (June 2010), I mentioned a cafe in East London that provided free coffee, free biscuits and free wifi, and charged customers for the length of time they occupied the table.

A cafe has just opened in Tokyo for writers, which charges people for procrastination. You can't leave until you have completed the writing task you declared when you arrived.

Justin McCurry, No excuses: testing Tokyo’s anti-procrastination cafe (Guardian, 29 April 2022)

Related posts: The Value of Getting Things Done (January 2010), The Value of Time Management (January 2010)

Wednesday, April 20, 2022

Constructing POSIWID

I've just been reading Harish Jose's latest post A Constructivist's View of POSIWID. POSIWID stands for the maxim (THE) Purpose Of (A) System Is What It Does, which was coined by Stafford Beer.

Harish points out that there are many different systems with many different purposes, and the choice depends on the observer. His version of constructivism therefore goes from the observer to the system, and from the system to its purpose. The observer is king or queen, the system is a mental construct of the observer, and the purpose depends on what the observer perceives the system to be doing. This could be called Second-Order Cybernetics.

There is a more radical version of constructivism in which the observer (or perhaps the observation process) is also constructed. This could be called Third-Order Cybernetics.

When a thinker offers a critique of conventional thinking together with an alternative framework, I often find the critique more convincing than the framework. For me, POSIWID works really well as a way of challenging the espoused purpose of an official system. So I use POSIWID in reverse: If the system isn't doing this, then it's probably not its real purpose.

Another way of using POSIWID in reverse is to start from what is observed, and try to work out what system might have that as its purpose. If this seems to be the purpose of something, what is the system whose purpose it is?

This then also leads to insights on leverage points. If we can identify a system whose purpose is to maintain a given state, what are the options for changing this state?

As I've said before, POSIWID principle is a good heuristic for finding alternative ways of understanding what is going on as well as seeing why certain classes of intervention are likely to fail. However, the moment you start to think of POSIWID as providing some kind of Truth about systems, you are on a slippery slope to producing conspiracy theories and all sorts of other rubbish.

Philip Boxer and Vincent Kenny, The Economy of Discourses: A Third-Order Cybernetics (Human Systems Management, 1990)

Harish Jose, A Constructivist's View of POSIWID (17 April 2022)

Related posts: Geese (December 2005), Methodological Syncretism (December 2010)

Related blog: POSIWID: Exploring the Purpose of Things

Tuesday, January 4, 2022

On Organizations and Machines

My previous post Where does learning take place? was prompted by a Twitter discussion in which some of the participants denied that organizational learning was possible or meaningful. Some argued that any organizational behaviour or intention could be reduced to the behaviours and intentions of individual humans. Others argued that organizations and other systems were merely social constructions, and therefore didn't really exist at all.

In a comment below my previous post, Sally Bean presented an example of collective learning being greater than the sum of individual learning. Although she came away from the reported experience having learnt some things, the organization as a whole appears to have learnt some larger things that no single individual may be fully aware of.

And the Kihbernetics Institute (I don't know if this is a person or an organization) offered a general definition of learning that would include collective as well as individual learning.

I think that's fairly close to my own notion of learning. However, some of the participants in the Twitter thread appear to prefer a much narrower definition of learning, in some cases specifying that it could only happen inside an individual human brain. Such a narrow definition of learning would not only exclude organizational learning, but also animals and plants, as well as AI and machine learning.

As it happens, there are differing views among botanists about how to talk about plant intelligence. Some argue that the concept of plant neurobiology is based on superficial analogies and questionable extrapolations.

But in this post, I want to look specifically at machines and organizations, because there are some common questions in terms of how we should talk about both of them, and some common ideas about how they may be governed. Norbert Wiener, the father of cybernetics, saw strong parallels between machines and human organizations, and this is also the first of Gareth Morgan's eight Images of Organization.

Margaret Heffernan talks about the view that organisations are like machines that will run well with the right components – so you design job descriptions and golden targets and KPIs, manage it by measurement, tweak it and run it with extrinsic rewards to keep the engines running. She calls this old-fashioned management theory.

Meanwhile, Jonnie Penn notes how artificial intelligence follows Herbert Simon's notion of (corporate) decision-making. Many contemporary AI systems do not so much mimic human thinking as they do the less imaginative minds of bureaucratic institutions; our machine-learning techniques are often programmed to achieve superhuman scale, speed and accuracy at the expense of human-level originality, ambition or morals.

The philosopher Gilbert Simondon observed two contrasting attitudes to machines.

First, a reduction of machines to the status of simple devices or assemblages of matter that are constantly used but granted neither significance nor sense; second, and as a kind of response to the first attitude, there emerges an almost unlimited admiration for machines. Schmidgen

On the one hand, machines are merely instruments, ready-to-hand as Heidegger puts it, entirely at the disposal of their users. On the other hand, they may appear to have a life of their own. Is this not like organizations or other human systems?

Amedeo Alpi et al, Plant neurobiology: no brain, no gain? (Trends in Plant Science Volume 12, ISSUE 4, P135-136, April 01, 2007)

Eric D. Brenner et al, Response to Alpi et al.: Plant neurobiology: the gain is more than the pain (Trends in Plant Science Volume 12, ISSUE 7, P285-286, July 01, 2007)  

Anthea Lipsett, Interview with Margaret Heffernan: 'The more academics compete, the fewer ideas they share' (Guardian, 29 November 2018)

Gareth Morgan, Images of Organization (3rd edition, Sage 2006)

Jonnie Penn, AI thinks like a corporation—and that’s worrying (Economist, 26 November 2018)

Henning Schmidgen, Inside the Black Box: Simondon's Politics of Technology (SubStance, 2012, Vol. 41, No. 3, Issue 129 pp 16-31)

Geoffrey Vickers, Human Systems are Different (Harper and Row, 1983)

Related post: Where does learning take place? (January 2022)