AI

Answering the SFWA’s Survey on LLM Use in Industry

by

While I’m not an SFWA member or even an author (beyond this little blog), as a lifelong reader of science fiction, I figured it was worth a few minutes to add my response to their current Survey on LLM Use in Industry. For the sake of posterity, here are my responses:

  1. I am a…
    • SFWA member in good standing (“Active”)
    • lapsed SFWA member
    • writer considering SFWA for future membership
    • ✅ member of the general public
  2. I create SFF in the following forms (choose as many as apply)
    • short fiction
    • longform fiction
    • poetry
    • comics / graphic novels
    • video games
    • analog games
    • film, theatre, and/or TV
    • nonfiction
    • ✅ I’m a reader/player of SFF.
  3. Has any of your writing been identified as part of stolen data sets in AI-related industry crises?
    • Yes, and I am part of a certified class action.
    • ✅ Yes, and I am not eligible for most/any class actions.
    • No, but people in my circles have been directly impacted.
    • No, but this issue remains a pressing concern.
    • I don’t know.

    (Note: Some time ago there was an online tool to look up sources that were used in one of the earlier revisions of one of the larger LLMs; I don’t currently remember who offered the tool or the specifics of the training database being reviewed. I do remember that both this blog and the Norwescon website, of which I’ve been both author and editor of much of the content for the past 15 years, were included in the training database.)

  4. How has your writing practice changed since the emergence of Generative AI and related LLM integrations?

    • ✅ I proactively turn off every new AI feature I can.
    • ✅ I switch away from writing tools that promote AI integrations wherever possible.
    • ✅ I avoid search engines and other summary features that rely on AI.
    • ✅ I accept AI features selectively, avoiding or switching off all the Generative AI tools I can identify, while leaving translation, spelling and grammar, and/or research assistants mostly intact.
    • I engage with AI chat features to brainstorm story elements, and/or for research questions of relevance to my writing.
    • I have used Generative AI for the development of story plots, characters, and/or scene construction.
    • I’m not a writer or editor.

    (Note: The first and fourth of these options seem to contradict each other. For clarity, I either disable or, if it can’t be disabled, actively avoid using generative AI; as noted in the sidebar of this blog, I do use machine-learning/LLM-based tools such as speech-to-text transcription, but when I do, I check and edit the output for accuracy.)

  5. Which of the following most closely resembles your position on the use of Large Language Models (LLMs) in the writing process?

    • ✅ There is no ethical use-case for writers, because this technology was developed through piracy and/or continues to negatively impact environmental systems and marginalized human beings.
    • Putting aside the historical and environmental issues, Generative AI needs to be approached differently from other LLMs, because other forms of LLM sometimes show up in tools (e.g., spell-check, grammar-check, translation software) that are normal parts of a writer’s workflow.
    • The use of any AI system for any part of the writer’s workflow that is not the writing itself (so, including brainstorming and research phases) is perfectly fine. It is only the words on the page that matter.
    • There are cases where the use of Generative AI for active storytelling might be a critical part of the story we want to tell, so it’s really a case-by-case determination.
    • Some writers are working for companies that make choices about AI without their involvement in the decision-making process, and this matters when deciding how we respond to the presence of AI in their work as individual creators.
    • I am not opposed to the use of LLMs in any capacity in the creative process.
  6. Tell us more about where you agree with or deviate from the statement you chose above.

    My actual answer is probably somewhere between the first (no ethical use-case) and second (recognizing LLM use in some tools) options.

    One of the biggest problems with the current discussions (including this survey and in the File770 threads started off of Erin Underwood’s open letter) is the grouping of several related but distinct technologies under the banner term of “AI”.

    Machine learning and LLM-backed analysis models are one thing. These are the technologies that have been used for years in many different contexts, including (some) spelling and grammar checkers, speech-to-text transcription, simple text-to-speech generation (suitable for screen readers and similar applications, not for audiobook creation), medical analysis, and so on. These are analytical or simple transformative, not creative, processes. In all cases, though, the output should be reviewed and checked by humans, not accepted as-is by default.

    Generative AI (genAI) is the severely problematic aspect, for all the reasons mentioned by many, many people advocating for its avoidance (the many unethical practices in the creation and ongoing use of the technology, social and environmental costs, high error rates, and many more).

    It’s unfortunate that all of these aspects are now grouped together as “AI”, as it makes it nearly impossible to approach the subject with any amount of nuance. I suspect that was what Ms. Underwood was attempting to do, though, as she also falls victim to the same confusion, she sorely missed the mark (and has continued to do so in her responses).

    As a reader, I would be very disappointed to see the Nebulas (and any other award) accepting the use of genAI in the creation of nominated (let alone awarded) works.

    (Note:: I wrote about the machine learning vs. genAI confusion on this blog earlier this year.)

  7. What forms of guidance do you think would most benefit writers trying to navigate the growing presence of LLMs in our industry?

    • ✅ Informational pages on SFWA.org explaining key terms and use-cases.
    • ✅ Articles on how to recognize and navigate forms of LLM in writing tools, and where to look for alternatives.
    • ✅ Full bans on any and all AI use in submissions and nominations processes, with consequences for failure to disclose.
    • ✅ Bans on Generative AI in submissions and nominations processes, with clear and severe consequences for failure to disclose.
    • ✅ Market reports that explicitly set a rigid bar for inclusion based on the publication’s commitment to not working with AI.
    • Other (please elaborate below).
  8. [Continued] What forms of guidance do you think would most benefit writers trying to navigate the growing presence of LLMs in our industry?

    Clarity in defining the differences among the technologies and determining which may be acceptable (such as speech-to-text transcription, spell/grammar checkers, etc.,) depending on the technology and its use, and which are unacceptable (genAI for text or art creation).

Good vs Bad AI (or ML vs AI)

by

The following is a (lightly edited) response I gave to a recent accessibility mailing list question from Jane Jarrow, coming out of a question around concerns around the use of various AI or AI-like tools for accessibility in higher education:

Folks responded by noting that they didn’t consider things like spell check, screen readers, voice-to-text, text-to-voice, or grammar checkers to be AI – at least, not the AI that is raising eyebrows on campus. That may be true… but do we have a clean way of sorting that out? Here is my “identity crisis”:

What is the difference between “assistive technology” and “artificial intelligence” (AI)?

This is me speaking personally, not officially, and also as a long-time geek, but not an AI specialist.

I think a big issue here is the genericization of the term “AI” and how it’s now being applied to all sorts of technologies that may share some similarities, but also have some distinct differences.

Broadly, I see two very different technologies at play: “traditional”/“iterative” AI (in the past, and more accurately, termed “machine learning” or “ML”), and “generative” AI (what we’re seeing now with ChatGPT, Claude, etc.).

Spell check, grammar check, text-to-speech, and even speech-to-text (including automated captioning systems) are all great examples of the traditional iterative ML systems: they use sophisticated pattern matching to identify common patterns and translate them into another form. For simpler things like spelling and grammar, I’d question whether that’s really even ML (though modern systems may well be). Text-to-speech is kind of an “in between” state, where the computer is simply converting text strings into audio, though these days, the use of generative AI to produce more natural-sounding voices (even to the point of mimicking real people) is blurring the line a little bit.

Speech-to-text (and automated captioning) is more advanced and is certainly benefitting from the use of large language models (LLM) on the backend, but it still falls more on the side of iterative ML, in much the same way that scientific systems are using these technologies to scan through things like medical or deep-space imagery to identify cancers and exoplanets far faster than human review can manage. They’re using the models to analyze data, identify patterns that match existing patterns in their data set, and then producing output. For scientific fields, that output is then reviewed by researchers to verify it; for speech-to-text systems, the output is the text or captions (which are presented without human review…hence the errors that creep in; manual review and correction of auto-generated captions before posting a video to a sharing site is the equivalent step to scientists reviewing the output of their systems before making decisions based on that output).

Where we’re struggling (both within education and far more broadly) is with the newer, generative “AI”. These systems are essentially souped-up, very fancy statistical modeling — there’s no actual “intelligence” behind it at all, just (though I’ll admit the word “just” is doing a lot of heavy lifting here) a very complex set of algorithms deciding that given this input, when producing output, these words are more likely to go together. Because there’s no real intelligence behind it, there’s no way for these systems to know, judge, or understand when the statistically generated output is nonsensical (or, worse, makes sense but is simply wrong). Unfortunately, they’re just so good at producing output that sounds right, especially when output as very professional/academic-sounding writing (easy to do, as so many of the LLMs have been unethically and (possibly arguably, but I agree with this) illegally trained on professional and academic writing), that they immediately satisfy our need for “truthiness”. If it sounds true, and I got it from a computer, well then, it must be true, right?

(The best and most amusing summary I’ve seen of modern “AI” systems is from Christine Lemmer-Webber by way of Andrew Feeney, who described it as “Mansplaining as a Service: A service that instantly generates vaguely plausible sounding yet totally fabricated and baseless lectures in an instant with unflagging confidence in its own correctness on any topic, without concern, regard or even awareness of the level of expertise of its audience.”)

Getting students (and, really, everyone, including faculty, staff, the public at large, etc.) to understand the distinction between the types of “AI”, when they work well, and when they prove problematic, is proving to be an incredibly difficult thing, of course.

For myself, I’m fine with using traditional/iterative ML systems. I’m generally pretty good with my spelling and grammar, but don’t mind the hints (though I do sometimes ignore them when I find it appropriate to do so), and I find auto-captioning to be incredibly useful, both in situations like Zoom sessions and to quickly create a first pass at captioning a video (though I always do manual corrections before finalizing the captions on a video to be shared). But I draw the line at generative AI systems and steadfastly refuse to use ChatGPT, AI image generators, or other such tools. I have decades of experience in creating artisanally hand-crafted typos and errors and have no interest in statistically generating my mistakes!

I’m afraid I don’t have good suggestions on how to solve the issues. But there’s one (rather long-winded) response to the question you posed about the difference between assistive technology and “artificial intelligence”.

All Your Images Are Belong to Zuck

by

If you have what you consider to be a hard-line stance against AI-generated images, and you post your photos and/or artwork to Instagram, Threads, and/or Facebook, you should likely either rethink that hard-line stance or stop posting your images.

Zuckerberg’s Going to Use Your Instagram Photos to Train His AI Machines:

During his earnings call for Meta’s fourth quarter results yesterday, Mark Zuckerberg made it clear he will use images posted on Facebook and Instagram to train his generative AI tools with.

Last month, Meta announced a standalone AI image generator to compete with the likes of DALL-E and Midjourney.

Meta has already admitted that it has used what it calls “publicly available” data to train its AI tools with.

Essentially, if you have a public Facebook or Instagram profile where you post photographs, there is a strong chance that Meta is using your work to train its AI image generator tools.

Yeah, this sucks, though it’s not surprising. I’ve stopped posting to Instagram, but still post a lot on Facebook, because this is where most of my friends are. I wish Mastodon would get more traction (I’m not tempted by either Threads or Bluesky; Threads is just another arm of Meta, Bluesky is more Jack Dorsey, neither is actually federating yet despite a lot of lip service, and neither currently allows post schedulers to tie in, which keeps me from using them for Norwescon posts), or, even better, that there was more of a push back towards actual self-owned blogs (like this one!) that aren’t locked behind virtual walls. But I don’t want to lose track of all of my friends, so until something major shifts, I’ll stick around, which means I’m probably going to end up shrugging and resigning myself to feeding Zuck’s AI machines, which I have definite ethical issues with.

I’m Training AI Chat Bots (Non-Consensually)

by

The Washington Post has published an article looking at the websites used to train “Google’s C4 data set, a massive snapshot of the contents of 15 million websites that have been used to instruct some high-profile English-language AIs, called large language models, including Google’s T5 and Facebook’s LLaMA.” If you scroll down far enough, there’s a section titled “Is your website training AI?” that lets you drop in a URL to see if it was scraped and included in the data set.

I checked three strings — “michaelhans” (to cover both this site and its prior address at michaelhanscom.com), “djwudi” (for my DJ’ing blog), and norwescon (which I’ve written or tweaked and edited much of the content for). All three of them are represented.

  • norwescon.org: 45k tokens, 0.00003% of all tokens, rank 528,147
  • michaelhanscom.com: 37k tokens, 0.00002% of all tokens, rank 635,948
  • djwudi.com: 3.7k tokens, 0.000002% of all tokens, rank 4,002,025

For the record, I’m not terribly excited about this. I’m also under no illusion that anything can be done; this stuff is all out on the open web, and as it’s free for actual people to browse through and read, it’s also free for bots to scrape and ingest into whatever databases they keep. Sometimes this is a good thing, for projects like the Internet Archive. Sometimes it’s unwittingly helping to train our new AI overlords.

AI Art, Ethics, and Where I Stand

by

While nobody specifically asked, since I have some friends who are all about the AI art and some who believe it’s something that should be avoided because of all the ethical issues, and since I’m obviously having fun playing with it with my “AImoji” project, I figured I’d at least make a nod to the elephant in the room.

An AI generated image of an African elephant standing in what appears to be a Victorian sitting room.

There are absolutely some quite serious ethical questions around AI generated artwork. To my mind the three most serious are (not in any particular order):

  1. Much of the material used to train the AI engines was scraped off the internet, often without any consideration of copyright, certainly without any attempt to get permission from the original creators/artists/photographers/subjects/etc., and some people have even found medical images that were only approved for private use by their doctor, but somehow ended up in the training sets. That situations like this are likely (hopefully) in the minority doesn’t absolve the companies who acquired and used the images to create their AI engines from being responsible for using these images.

  2. As the AI engines continue to improve, it is getting more and more difficult to distinguish an AI generated image from one created by an artist. There are also a number of people and organizations who have flat-out stated that they are looking at AI generated imagery as a way to save money, because it means they now don’t have to pay actual artists to create work. Obviously, this is not a particularly good approach to take.

  3. Because some of the engines are able to create images in the style of a particular artist, and the output quality continues to improve, there have already been instances where a living artist is being credited for creating work that was generated by an AI bot. And, of course, if you can create an image that looks like your favorite artist’s work for low or no cost…well, for a lot of people, they’ll happily settle for an AI generated “close enough” rather than an actual commissioned piece. Obviously, this is also not a particularly good approach to take.

I’m enjoying playing with the AI art generation tools. I’m also watching the discussions around the ethical questions around how they can and should be used.

The issues above are all very real and very serious. It’s also true that AI art can be just another tool in an artist’s toolbox. I’ve seen artists who use AI art generators to play with ideas until they find inspiration, or who use parts of the generated output in their own work. I’ve seen reports of people who want to commission art use the generator to get a rough idea of what they’re looking for that they can give to an artists as a rough example or proof of concept. So there are ways to use AI art generators in, well, more-ethical ways (it’s hard to argue they’d be entirely ethical when the generators have unethical underpinnings).

So, where I stand in my use at this point:

  1. I don’t use living artist’s names to influence the style one way or another, and have only occasionally used dead artist’s names as keywords (I’ll admit, H.R. Giger has been a favorite to play with).

  2. I don’t feed images in, try to generate images of actual people, or use images of actual people (including myself) as source material.

    One caveat: if a tool does all of its processing locally on my device, I may use my own images, including some of myself. But nothing that feeds images into the systems.

  3. And, of course, anything I do is just for fun, and to make me, and maybe a few other people, laugh (or occasionally recoil in horror).

For a few months this past year, I used an AI-generated image of a dragon flying over a city skyline for the Norwescon website and social media banner image. This was always intended as a temporary measure to fill the gap between last year’s convention and getting art from this year’s Artist Guest of Honor, and as soon as we had confirmed art from our GOH, the AI-generated art came down. It was also chosen much earlier in the “isn’t AI art neat” period, before I’d read as much about the issues involved. As such, I won’t be using AI art for Norwescon again, and will go back to sourcing copyright-free images from NASA or other such avenues when we are in the interregnum period.

So: I understand those who see AI art as something that should be avoided. I also understand those who see it as another tool. And, honestly, I also understand those who just see a shiny new toy that they want to play with. I’m somewhere in the midst of all those points of view, and while I don’t personally see the need to avoid AI art bots entirely, I am consciously considering how I use them and what I use them for.

by

AI Weirdness • 2020 headlines: “Midway through 2020, people started suggesting that I train a neural net on 2020 headlines, and I was skeptical that there would be enough weird ones to make a decent project. Then 2020 continued to be 2020.”

AI Dungeon 2

by

I haven’t taken the time to try this yet, but this seemed like something quite a few people I know would be into: a Zork-style game with an AI backend, so you can do…well, anything, apparently.

I wrote earlier about a neural net-powered dungeon crawling text adventure game called GPT-2-Adventure in which gameplay is incoherent and dreamlike, as you encounter slippery sign text, circular passages, and unexpected lozenge rooms. A PhD student named Nathan trained the neural net on classic dungeon crawling games, and playing it is strangely surreal, repetitive, and mesmerizing, like dreaming about playing one of the games it was trained on.

Now, building on these ideas (and on an earlier choose-your-own-adventure-style game he built), Nick Walton has built a new dungeon-crawling game called AI Dungeon 2. Nick made a few upgrades, such as beefing up the AI to the huge GPT-2-1.5B model OpenAI recently released, adding a penalty for repetitive text, and expanding the dungeon game training examples to a bunch of modern human-written games from chooseyourstory.com.

I CAN’T STOP PLAYING THIS GAME

Here’s the actual game site: AI Dungeon. Have fun!

Linkdump for November 14th through November 29th

by

Sometime between November 14th and November 29th, I thought this stuff was interesting. You might think so too!

Back again!

by

Woohoo! We’ve reconfigured a few areas of the network here at Casey’s house, and it seems that things are back up and fully functional for me again. So, as things go here, I’ll do my best to return to updating my pages on a regular basis. I know, I know, something of a shock after about a month of near-nonexistent updates…but I’ll try.

Things for me are still in something of a holding pattern at the moment. I got word from the landlord of my apartment complex that the carpets are scheduled to be installed this Monday, so I should finally be able to get into my place Monday afternoon/evening sometime. I’ve made the requisite calls to the telephone and electric companies and am all set up there, so should be good to go as soon as I get the word from the landlord on Monday. I’ll be sending out the mailing address and phone number to those who need it in the near future.

Internet access options for me are still being investigated. I’m hoping to get set with a DSL line, I just need to get in contact with the local ISP‘s to see if my apartment has that as an option. I’m assuming it does — I’m going to be living right on Capitol Hill, just about 20 blocks or so uphill from downtown Seattle — but I’m not entirely sure yet. In any case, Casey has graciously allowed me to keep my webserver at his place until I have things up and running at my apartment, so the server shouldn’t be going down again at all, however there may be a couple weeks where my online abilities are severely limited until I get my own connection up and running. It’ll all get straightened out eventually — I’m just glad to have friends down here who are able and willing to assist me in all of this

In other news, I’ve been playing a lot with my digital camera since I got it. I took some time recently to stitch together some panoramas I’d taken. The first three were all taken before I left Alaska — from top to bottom, the Inlet as seen from Earthquake Park in Anchorage, a view of the Palmer hayflats where I hit a bonfire with some friends, and Jewel Lake, a popular destination in South Anchorage.

Cook Inlet, Anchorage, AK

Bonfire Panoramic, Mat-Su Valley, AK

Jewel Lake, Anchorage, AK

The fourth shot was taken at Gasworks Park here in Seattle during the 4th of July celebrations, about an hour before the fireworks display. I wanted to try and capture the sheer mass of people — later reports placed it at around 6,000 people just at this park (and it was one of two major fireworks displays within Seattle). I think it came out pretty decently.

4th of July 2001, Gas Works Park, Seattle, WA

I’ve been out to see two movies since I came down here so far — since it’s been a bit since I’ve seen them, I’ll just give brief rundowns of each. First off was Atlantis, Disney’s latest animated flick — another fun one from Disney. Not one of their all-time classics, but very enjoyable, with some absolutely breathtaking animation at times. More recently was A.I., the Spielberg/Kubrick sci-fi collaboration. In brief — I believe it to be an astounding piece of work, quite possibly Spielberg’s best work yet, and a film that, while getting wildy mixed reviews, is very likely to stand the test of time like few other recent films. Very, very impressive filmmaking, and my hat is off to Spielberg, Kubrick, and the rest of the forces behind this film. I’ll most likely post more about it after I’ve had a chance to see it a second time.

That’s the majority of the big news so far. As mentioned earlier, now that things are up and running again, I’ll do my best to return to a more reliable update schedule here. It’s good to be back….