Baldur Bjarnason

My position on AI (for future reference)

– Baldur Bjarnason

I’m writing this mostly for myself.

It’s light on references. If you want more referenced and more solidly argued overviews of some of the issues with “AI” I invite you to go explore the “AI” category on this site and the “AI” tag on my newsletter archive.

(This week’s links are at the end of this post.)

The short version is that I think generative models and ML-implemented automated decision-making systems are extremely harmful.

Their systemic flaws are a Y2K-style economy-wide disaster in the making.

Many individuals are able to use generative models safely. The problem is that a small percentage will suffer through catastrophic errors, bias, or misinformation and many of the biggest systemic downsides will only appear at a later date, long after the fact.

This creates a dynamic that’s similar to what we faced with COVID or Y2K. The extreme harm is there, but is at such a low percentage and at such a delay that it doesn’t meaningfully affect broad societal behaviour until it’s too late.

If the technology gets adopted widely then the “low” percentage becomes a humongous number in real terms.

One percent of all society is a lot of people.

This is why my current position is that we should do anything we legally can to slow or limit the adoption of these models in our systems, organisations, and institutions.

The harms #

“AI” systems are not functional #

These systems are not nearly as functional as vendors make them out to be.

The core issue is that these systems are non-deterministic and prone to unpredictable errors. They are prone to making major factual errors, severe “reasoning” failures, suffer from systemic biases of pretty much every kind, and their output – at best – is of a pretty low quality, forcing organisations to hire low paid workers to do the unsavoury job of making unacceptable machine-generated pap acceptable.

In pragmatic business terms, using and integrating these systems presents enough risks to fill an entire book.

These systems do not present a long-term competitive advantage. The harms come from their use.

Statements such as “if we don’t make them then the Chinese will” are not a counter-argument because the biggest problems aren’t caused by the existence of these systems but from their use in critical or vulnerable contexts.

It’s systemic legitimate use at scale that is likely to cause the most harm.

(This list is not even remotely exhaustive.)

Many of these harms do not directly impact the organisation implementing the “AI” system, which means that they cannot be trusted to correctly assess the potential harm.

Most of the harms happen at such a rate that their effect won’t become noticeable in small-scale or casual use.

This means that efforts to replace human labour with generative model automation is likely to be disastrous for both our economies and our society.

These systems are unpredictable and untestable #

Their very design means that we cannot test them properly to discover and prevent their failure conditions. Generative models cannot be integrated safely at scale. They will always have unpredictable errors.

This would be manageable if it weren’t for the fact that…

Many of their failures are invisible to the user #

Because these systems are used to automate, very few of their users will have the expertise to be able to correctly identify issues with the output.

Most flaws will go unnoticed.

Even when the system’s user has the skills to identify problems, they are likely to suffer from automation bias: tools for cognitive automation are there to help you think less, and because you’re thinking less, you’re less likely to be critical of the output than you would normally.

Automation bias is a major issue for aviation and the “AI” industry has it in spades.

That, for many use cases, the error rates might be relatively low (single digit percentages) makes these systems more harmful not less. It means that their severe dysfunction is less likely to be spotted until they have been integrated or used at scale.

We can’t trust the AI industry #

“AI” has a long, long history of snake oil, false promises, and even outright fraud. Most of the studies and papers I’ve reviewed over the past few months (and I’ve reviewed hundreds) are flawed, misleading, or even outright deceptive.

The industry also makes independent replication of research essentially impossible.

The poor quality of much existing research and the roadblocks preventing replication means that the capabilities of these systems are vastly overstated, but we won’t be able to conclusively demonstrate their limitations to the media or policy-makers in the near or medium term.

“AI” systems are vulnerable #

Generative models are prone to training data poisoning and their prompt inputs can’t be secured.

Widespread use is, effectively, baking systemic security flaws into our society and infrastructure.

Language models are prone to memorisation and copying #

Many of the major benchmarks used by OpenAI and other vendors to guide language model development are human exams such as medical exams or bar exams. These exams primarily test rote memorisation.

Companies like Microsoft and Google are hell-bent on using language models as information systems to help people research and discover facts. This is only possible through even more memorisation and copying of strings of text from the training data sets.

The use of language models for programming requires the memorisation of platform APIs, control syntax, globals, and various other textual elements that are necessary for functioning code.

Together, these requirements mean that current large language models are optimised to memorise and copy strings of text.

As they are purely language models, they exclusively operate in the world of text, not rules, facts, or judgement. To be able to correctly answer the question of who was the first person to walk on the moon, they have to memorise much of that answer as textual fragments.

This bears out in research as performance of these systems degrades when they memorise less.

This is an issue because there is no way to control exactly how these memorised texts get used or even what gets memorised.

Every language model is a stochastic plagiariser, which can cost you your job if you’re a writer, but can result in serious legal issues if you’re a programmer. Current approaches for preventing this kind of copying are not sufficient as they focus on preventing verbatim copies and LLMs excel at lightly rephrasing output to match context.

The problem is that “lightly rephrased” copied text is still plagiarism. For code, that’s still licence contamination. For writing, that’s usually still a fireable offence.

“Inevitable” harms are not inevitable #

Because it seems to be impossible to automatically detect the output of a generative model every aggregator is likely to be flooded with garbage before long.

That means, among others:

It’s extremely likely that our information ecosystem will be severely degraded by generative models.

Many people present this as inevitable. But this isn’t exactly true. Once you start to look at the spam that is being generated and the fraud that’s being perpetrated, it starts to look like much of it is created using commercially-available generative models made by vendors such as OpenAI. Self-hosted models do not seem to be nearly as effective for spam or criminality.

If this is correct, that means this problem can be mitigated by targeting the providers of generative models and related services such as OpenAI, Google, Anthropic, and Microsoft. We can prevent them from improving their systems’ effectiveness at fraud, and we can punish them directly for failures to create safe products.

This applies equally to both fraud and spam. The best generative models for criminal use seem to be the commercial ones. Creating and maintaining a complex generative model is extremely expensive, so if stricter regulation manages to mitigate the damage caused by commercial or commercially-adjacent academic systems, then criminals and spammers are, in my opinion, unlikely to be able to come up with effective replacements using their own resources.

Beyond the plagiarism problem, copyright is not my primary concern with generative models. Simple prompt-to-output use will result in works that don’t have copyright protection. But edited output almost certainly will be protected. Creative works that are integrated into a larger, copyrighted work are likely to be protected as well. And modified or edited output should result in a copyright-protected final product.

I don’t expect this to be a major issue. It does mean that, legally speaking, you should treat any generative model output the same way as you would a public domain text or image. Using it is fine, but make it your own in some way.

Training data #

Whether it’s legal for companies to train these models on copyrighted works has not yet been answered conclusively.

We don’t know if this is fair use under US copyright law or whether it falls under one of the existing exceptions to copyright law elsewhere. The EU is likely to clarify the issue eventually, but the USA’s ongoing governance dysfunction means we are only likely to get a clear answer to the question through lawsuits.

I expect this to get resolved eventually. Unfortunately there is no answer that would benefit all involved equally. Either tech companies will benefit by being allowed to use existing works without payment while directly competing with – to the point of extermination – the producers of those works. Or the creative industries will benefit by having hard limitations put in place on what the tech industry can do with any and all copyrighted works. Either answer could literally wipe out entire industries.

Non-issues #

They are not on the path to Artificial General Intelligence (AGI) #

The argument for why AGI is science-fiction is too long to hash out in this already too long post, but the short version is that extraordinary claims require extraordinary evidence and those making the AGI claim have provided none. Everything they’ve presented in support of their claim is shoddy or misleading.

This means we aren’t missing a thing if we slow down or abandon this technology.

Dev #

​

Media #

AI #

Photographs #

Took this one using a Carl Zeiss Jena Werra about twenty years ago in Ashton Court, Bristol, IIRC.

Two men playing golf in the distance in Ashton Court, Bristol

Interesting looking clouds over Hellisheiði.

Dramatic clouds over Hellisheiði