Neuvottelija.AI

EP210 · Society · first published 2023-09-08

AI and me | Petteri Järvinen | Negotiator 210

Author Petteri Järvinen joins Sami Miettinen in September 2023 to discuss his book Tekoäly ja minä ("AI and me"). It is a consoling book turned into a conversation: people cope, professions die and are born, and the Luddites have always been wrong. Underneath run two questions that do not resolve — how anyone takes responsibility for a neural network’s decision when it cannot give its reasons, and what happens when bias is corrected by distorting the data. The practical takeaway is the line between summarising and extrapolating. The transcript is auto-captioning, which is disclosed.

Sami Miettinen · Sections: AI and Society + Tools and Implementations

AI and me | Petteri Järvinen | Negotiator 210

Summary: Author Petteri Järvinen joins Sami Miettinen in September 2023 to discuss his book Tekoäly ja minä (“AI and me”). It is a consoling book turned into a conversation: people cope, professions die and are born, and the Luddites have always been wrong. Underneath run two questions that do not resolve — how anyone takes responsibility for a neural network’s decision when it cannot give its reasons, and what happens when bias is corrected by distorting the data. The practical takeaway is the line between summarising and extrapolating. The transcript is auto-captioning, which is disclosed.

The guest, and a note on the source

Petteri Järvinen is a long-established Finnish non-fiction author, and the episode is built on his book Tekoäly ja minä — Ihmisenä tekoälyn aikakaudella (“AI and me — being human in the age of artificial intelligence”, Tammi 2023). Recorded 8 September 2023.

Two caveats:

The book’s message is consoling: people cope with AI and against it. The most interesting material, though, is what does not get resolved.

1. Two visions — and why neither is complete

The host puts forward two views he says he heard from a venture investor, and asks Järvinen to assess them.

Vision 1: AI as a delegation platform. At last an interface you can give orders to, which then handles the thing.

Järvinen’s answer is not enthusiasm but a historical comparison, and it is the sharpest point in the section: voice control was talked about the same way ten years ago — voice assistants were to be the next content distribution channel — and nothing came of it. He adds his own experience from the 1990s: he believed in the internet sincerely and could not imagine its downsides. The conclusion is not that the vision fails, but that the downsides arrive as surprises.

Vision 2: personal dataism. You train a copy of yourself in a private environment — “mini-Petteris and mini-Samis” — which handles the routine and saves what the host calls “Kahneman juice”, the capacity for hard reasoning.

Järvinen’s objection is the finest thing in the episode, because it is personal and concrete:

The best things are surprises. If he had had an agent filtering approaches, it would most likely have rejected this very interview request — an unknown asker, no prior history.

That is the cost of curation in its shortest form, and it is offered as an example both parties recognise in that moment.

2. Summarising versus extrapolating — the most usable distinction

This is the part you can apply today:

Summarisation is reliable because it is compression. The model removes material and preserves some level of fidelity — it cannot go very badly wrong. Hallucination comes from extrapolation: ask the model to add something that is not in the source, and it produces it.

The host immediately adds a limitation that stops the distinction becoming an excuse: can familiarity with the original material ever be avoided entirely? The question is left open.

The practical follow-on is precise: specialised plugins reduce hallucination, because they move part of the task off the language model — mathematics to a calculator, a document to a retrieval tool. That is an observation about architecture, not an opinion.

The section also carries a cautionary example: a trial in which a model, asked for proposals, produced something nobody could act on. Järvinen uses it to explain why he left most of his examples out of the book — they age quickly, and individual trick questions get patched.

That is an author’s methodological point, and it is useful: do not build an argument on a single model error.

3. The Flynn effect — a worry that is not anti-technology

Järvinen’s concern is offered as a research reference and is checkable: the effect named after a New Zealand-based psychologist, by which measured IQ rose for decades, has reversed in several countries. One proposed explanation is that external tools — the smartphone, the search engine — reduce the need to tax the brain.

His conclusion is conditional rather than categorical, and should be read that way: if everything is handled by issuing commands, the effect on brain activity may not be good.

This is the one place where the two men genuinely differ without it being framed as a dispute. The host’s own practice is the opposite: he says he maintains attention by giving the brain a secondary task — walking the dog, a simple game — which improves how well he follows an audiobook. Järvinen reports the reverse: he no longer wants to fill his environment with constant sound, and values letting his thoughts wander.

Neither claims the other is wrong, and the episode invites the listener to test it themselves. That is an unusually honest way to handle advice about cognition.

4. Responsibility that cannot be taken — the weightiest section

This is why the episode is worth reading, and it comes in two parts.

Why a neural network cannot give reasons

Järvinen makes a historical distinction that explains the present better than most accounts:

In the 1980s people built knowledge systems — rules, if-then statements, decision trees. They failed so badly that the whole field froze for years. The breakthrough came from abandoning the teaching of rules: give it the material, give it the desired outcomes, and let the machine work out how to get from one to the other.

And that is exactly why it cannot tell you why it arrived there.

The distinction is technical and it is correct. From it comes the question neither man resolves:

On what basis does a person take responsibility for a decision whose reasoning they cannot see? — and that applies as much to not being shortlisted for a job as to a decision of state.

The example that makes it concrete

Järvinen recounts the Dutch case: authorities built a system to catch social benefit fraud, and the outcome was a scandal that brought down the government.

The case is real and checkable — the Dutch childcare benefits affair, over which the prime minister’s cabinet resigned in 2021. Järvinen also notes that it was not really AI at all but machine learning and simple conditional statements — which makes the example stronger, not weaker: the problem of automated decision-making does not require an advanced model.

With it comes a realistic qualification that lowers the temperature: scoring and prioritising customers has always happened and requires no AI. What is new is scale and opacity, not the practice.

5. Bias, and where this write-up separates position from observation

The section on discrimination is the most contested in the episode and needs care.

The observation, which is checkable: image generators produce, for a Finnish use case, distributions drawn from elsewhere — and both men discuss this as a user’s problem rather than a political question.

Järvinen’s position, which is a position and not an observation: the data should not be deliberately corrected to produce a desired outcome, because the correction is artificial and not transparent. He calls this engineering logic.

The counter-argument, which the episode does not make and which is marked here: training material is not a neutral starting point but is itself selected, so “following the data” is not the absence of correction but the reproduction of one particular selection. Either choice is a choice.

The episode does name a solution common to both positions and therefore useful: private environments trained for the purpose — so the distribution is chosen deliberately rather than inherited by default.

Nudging

The same section covers steering. Both accept that nudging can be justified, and both are sceptical when a commercial party does it: the example is a mapping service changing its route suggestions towards fuel economy.

The host’s own example is an index fund whose policy moved from a general index to a sustainability-weighted one. His argument is technical — he wants the cheapest possible unfiltered general index — and he says explicitly that he does not object to sustainability as such. He also states himself that the provider’s podcast is his competitor. That is marked here because he marks it himself.

6. Why human work is not finished

The optimistic section, resting on two arguments — one good, one a guess.

The good argument — a historical control: Järvinen says he found a 1978 book on a library discard shelf whose visions of the future sounded exactly like today’s. Professions have died and more have been born. The Luddites have always been wrong — to which he adds the dry qualification that they have also always been partly right.

The worry is not the quantity of work but polarisation: simple execution jobs are needed and demanding ones are needed, and the middle thins.

A guess, marked as one — the “organic reaction”: when producing text is easy, people will come to value text somebody actually thought about, which is not verbose filler. That is a prediction about a cultural counter-reaction and cannot be checked.

On creativity Järvinen makes a definitional point: creativity builds on what was made before plus random variation, so the question is not whether a machine can recombine but at what point recombination becomes human-like creativity. He says he is most astonished by the coding ability, because in his view it cannot be mere next-word probability.

From which comes a warning that is professional rather than dramatic: if the machine writes the code, the weight shifts to testing and debugging — and machine-written code running on another machine is the point at which he himself starts to worry.

7. Two side-paths worth taking

The brain myth. Järvinen corrects outright the claim that humans use ten per cent of their brains: all regions are used, and there is no idle capacity. The correction is right. The host reframes the same idea in a way that keeps what is useful: moving routines into automatic processes frees capacity for hard reasoning.

Negotiating meaning. The episode closes on an idea from a Finnish AI researcher in which a negotiation is conducted between machines holding the parties’ preferences: assume the other side wants to increase net wealth and make value-adding transactions — and let the machines look for the shared interest. It is a vision rather than a product, but for this channel’s subject it is the single most interesting thought in the episode.

Claims presented as claims

What survives the episode

  1. Summarising is compression, hallucination is extrapolation. That tells you which tasks tolerate light supervision and which do not.
  2. A neural network cannot give reasons, because teaching reasons is precisely what was abandoned. The responsibility question that follows is not resolved, and the episode does not pretend otherwise.
  3. The problem of automated decision-making does not require advanced AI — in the Dutch case machine learning and conditional statements sufficed.
  4. Curation costs you surprises. The best illustration is that an agent would have rejected this interview request.

How the episode runs

Markdown: index.md · Suomeksi