Re: Proposed draft of an AI policy

Dimitre Novatchev <dnovatchev@gmail.com> writes:
> The first question that naturally arises is how to implement such a policy in practice. Who, and using what mechanism, will judge
> the originality of the work?

There’s only one practical way for the community group to implement any policy: by what consensus the community group achieves.

> This is a large and very important problem and at present entire organizations and specialized events are focusing on it.

It is a large and important problem and I’m glad that organizations are working on it. There are lots of different aspects of the issue that we could discuss.

I would rather not. I think there *are* a host of very good moral, ethical, environmental, and financial reasons to support the position that commercial LLMs are a blight on our industry and on the world generally. The first draft I wrote of the policy included a necessarily incomplete enumeration of some of them. I took at that out. I don’t think we need to debate those issues.

I think there is a fundamental, technical issue that wholly and decisively justifies a policy against contributions that originated in any form of LLM.

The task of our community group is to write a set of clear, coherent specifications. This is an objectively difficult task. Human authors start with a conceptual model of some feature that they believe the community group agrees with, or that they wish to persuade the community group to agree with. They attempt to describe that feature clearly in prose. Authors are more-or-less successful at this for a wide variety of reasons some within and some beyond their control.

A reader starts with the prose and constructs their own model of the feature. Like authors, readers are more-or-less successful at this for a wide variety of reasons, the most significant of which is how clear and understandable they find the author’s prose.

Members of the community group can compare the model they constructed from the prose with what they believe the community group discussed and/or agreed. The small subset of readers who are implementing the specification attempt to build software that performs the feature.

The much broader community of readers decide if they believe they understand the feature and if they think it is applicable to the use case they have in mind. Then they compare what an implementation actually does with what they thought it was going to do.

The community compares notes. Authors and readers collaborate on building an understandable specification. Sometimes defects are found in the feature. Sometimes defects are found in the description of the model. Sometimes they are found in implementations.

The one through line in this entire process is that human authors are attempting to communicate their understanding of a feature to human readers.

Specification prose that I draft sometimes contains Americanisms that some may find unfamiliar. Mike’s prose may sometimes contain Britishisms that I find unfamiliar. An author writing in their second, third, or fourth language may choose forms of expression that a native Enlish speaker would not.

Careful human review of all these drafts, despite their various idiosyncracies, is valuable and justified by the fact that we are all attempting, in good faith, to communicate technical ideas that we believe we understand clearly, so that others can understand them.

What LLMs are undeniably exceptionally good at is producing plausible text: text that a human reader will not immediately and obviously find flawed. But the LLM does not in any sense whatsoever *understand* the features of our language.

Asking humans to review text generated in this way is unduly burdensome. Prose that reads well and appears to describe a feature, but on much closer inspection is internally inconsistent, or worse, simply vacuous, is corrosive to the process.

Every one of us has a limited amount of time to read and review specifications. We must use that time wisely and constructively. Reviewing plausible text generated by a statistical process devoid of any actual understanding of the specifications is, objectively, technically, a poor use of our valuable time.

We must have a policy that protects us from it.

                                        Be seeing you,
                                          norm

--
Norm Tovey-Walsh
CEO, Saxonica

Received on Thursday, 28 May 2026 11:25:44 UTC