Seattle Worldcon 2025 chair Kathy Bond and Program Division Head SunnyJim Morgan tonight published their promised statement detailing how ChatGPT was used in the program panelist selection process. The complete statement appears at the end of this post.
Bond says ChatGPT was not used in deciding who to invite as a panelist, it was used “in the discovery of material to review after panelist selection had occurred.”
Morgan adds, “This process has only been used for panelists appearing on site in Seattle; panelists for our Virtual program have not yet been selected.”
Bond stresses that “ChatGPT was used only for one tailored task that was then followed by a human review and evaluation of the information,” and that “no selected panelist was excluded based on information obtained through AI without human review and no selected panelist was chosen by AI.”
As part of their remediation, the Seattle committee is redoing the part of the program process that used ChatGPT, with that work being performed by new volunteers from outside their current team.
Morgan also makes her own apology (the chair published her own several days ago).
I want to apologize specifically for our use of ChatGPT in the final vetting of selected panelists as explained below. OpenAI, as a company, has produced its tool by stealing from artists and writers in a way that is certainly immoral, and maybe outright illegal. When it was called to my attention that the vetting team was using this tool, it seemed they had found a solution to a large problem. I should have re-directed them to a different process. Using that tool was a mistake. I approved it, and I am sorry. As will be explained later, we are embarking on the process of re-doing the vetting stage for every invited panelist, completely without the use of generative AI tools.
And Morgan has provided the text of the ChatGPT query that was used in the vetting process.
The committee will be making their next update about the subject on May 13.
The full statement follows the jump.
May 6th Statement From Chair and Program Division Head
Chair’s Statement
[By Kathy Bond] As promised last Friday, I am publishing this statement, in conjunction with a statement below from our Program Division Head, to provide a transparent explanation of our panelist selection process, answer questions and concerns we have received, and openly outline our next steps. As a result, it is a long statement. Many of the steps outlined below will take time to complete; we commit to keeping you updated as we move forward with our next update on May 13th.
Last week, I released an incomplete statement about an important subject, and as a result of that flawed statement, I caused harm to the Worldcon community, to the greater SF/F community, and to the dedicated volunteers of Seattle Worldcon, many of whom felt that they could no longer be proud of what they had accomplished on behalf of the Seattle Worldcon. I am deeply sorry for causing this harm. It was not my intent, but it was my effect.
The other harm that was caused was our use of ChatGPT.
ChatGPT was used only in one instance of the convention planning process, specifically in the discovery of material to review after panelist selection had occurred.
It was not used in any other setting, such as
- deciding who to invite as a panelist
- writing panel descriptions
- drafting and scheduling our program
- creating the Hugo Award Finalist list or announcement video
- administering the process for Hugo Award nominations
- publications
- volunteer recruitment
As you will be able to read further in the below statement from our Program Division Head about the panelist selection process, ChatGPT was used only for one tailored task that was then followed by a human review and evaluation of the information.
Although our use of ChatGPT was limited, it does not change the fact that any use at all of a tool that has caused specific harm to our community and the environment was wrong.
As noted above, our Program Division Head has also released a statement, below this one, with a transparent explanation of our panelist selection process. Also included in that statement is the query that was used to generate the results reviewed by the program team. The purpose of that statement is to show exactly where ChatGPT entered the picture, and hopefully to ameliorate some of the concerns we have heard from your comments as to whether a person has been included on or excluded from our program because of AI.
Let me reiterate that no selected panelist was excluded based on information obtained through AI without human review and no selected panelist was chosen by AI.
We know that trust has been damaged and statements alone will not repair that, nor should they. Our actions have to be worthy of your trust. As such we are committing to taking the following steps in the remaining 100 days before the convention. Some of these steps may result in changes being made right away to our process. Some may only result in transparency with the community and as a way to provide insight to future convention committees.
- We are redoing the part of our program process that used ChatGPT, with that work being performed by new volunteers from outside our current team. The timeline and guidelines for this action will be finalized at the next meeting of our leadership team this coming weekend.
- We are reaching out to a few outside members of the community with prior experience in Worldcon programming to come in and perform an audit of our program process. They will have access to all of our systems and people in order to review what happened, confirm what is already being done to remove ChatGPT from our program vetting, and provide a report to the community about what they discovered and their recommendations. This process is already underway; we hope to have a report by the end of May.
- Anyone who would like their membership to be fully or partially refunded based on these issues may contact registration@seattlein2025.org for assistance.
- The decision process that led to our use of ChatGPT has revealed shortcomings in our internal communications. We are reevaluating our communication methods, interdependencies, staffing, and organizational structure to ensure we can detect and respond to issues at the earliest opportunity. We commit to improving our internal communication structures alongside our external communications. This will be an ongoing process.
- We are exploring options for providing additional oversight and guidance to the Chair and the leadership team. The plan for this action will be finalized at the next meeting of our leadership team this coming weekend.
As Chair of the Seattle Worldcon I am promising to work with my whole team to restore the community’s trust in the convention and rectify the damage done as best we can. Some of these steps take time to implement, especially as a volunteer organization; I commit to update you as to their implementation and outcomes in regular briefings from the Chair, the first of which will be May 13th.
Finally, regarding the resignations yesterday of three members of our WSFS Division, I deeply appreciate the service they provided for Seattle Worldcon and their dedication to the community. I am glad that they were on our team for so long. We are all of us volunteers, and when people have needed to step back or resign, they have done so with my immense appreciation and gratitude for the substantial contributions they have already provided, and my understanding that sometimes leaving is the best choice for an individual.
I am also heartened that other members of the WSFS Division have chosen to stay on the team and fill in the roles vacated. I am confident that Kathryn Duval as Hugo Administrator and WSFS Division Head, and Rosemary Park as Deputy Hugo Administrator and Deputy Division Head will continue the excellent work already performed. We are committed to delivering the Hugo Awards with transparency and integrity and in celebration of our community. We appreciate that the team members who stepped away are working to ensure a smooth transition to those stepping up.
It is an honor to be the Chair of the Worldcon, and to serve the Worldcon community. As we move forward, we will continue to review your feedback and suggestions. The best way to reach our leadership team about these issues is to utilize a new email address we have created, feedback@seattlein2025.org, but we will continue to monitor comments on our blog posts and social media as well.
Kathy Bond
(she/hers)
Chair Seattle Worldcon 2025
chair@seattlein2025.org
Statement from Program Division Head SunnyJim Morgan
First, and most importantly, I want to apologize specifically for our use of ChatGPT in the final vetting of selected panelists as explained below. OpenAI, as a company, has produced its tool by stealing from artists and writers in a way that is certainly immoral, and maybe outright illegal. When it was called to my attention that the vetting team was using this tool, it seemed they had found a solution to a large problem. I should have re-directed them to a different process. Using that tool was a mistake. I approved it, and I am sorry. As will be explained later, we are embarking on the process of re-doing the vetting stage for every invited panelist, completely without the use of generative AI tools.
Second, because of widespread and legitimate concerns about our use of ChatGPT, and to correct some misinformation, I’d like to offer a clearer description of the full panelist selection process used by Seattle Worldcon. This process has only been used for panelists appearing on site in Seattle; panelists for our Virtual program have not yet been selected.
Panelists are selected by the program team, not by AI tools. We have received panelist applications from many more brilliant and talented people who are interested and qualified to be on the program than we can use on panels. No matter what process we use in selection, we will disappoint hundreds of applicants.
In stage one, our 30 track leads, each responsible for a single content area of the program, are given access to the master list of applicants, and are asked to select people who they would like to invite to participate on the program. Each track lead has their own subject area expertise and vision for the panels in their track. Some chose to invite a wide segment of suitable applicants to mix and match onto panels later, while others were looking for very specific skill sets and interests for specific panels. Track leads base their decisions to invite panelists on the content of the panelist application, the track lead’s knowledge of the applicant and the subject area, and additional input from members of the program team.
The applicants recommended for participation by the track leads are then moved on to stage two, the vetting process, in which we attempt to find out whether there is any information not already known about the applicant which could be potentially disqualifying. At this stage we are looking only for actions that would go against the convention’s code of conduct and antiracism statement.
A few months ago, I discovered that the vetting team assigned with this task had been using ChatGPT to quickly aggregate links, specifically asking it for links to any material that would disqualify the applicant as a panelist. Then, after manually reviewing the information at the links provided, a final decision was made by me whether to approve the person’s invitation to participate on the program.
For those who underwent vetting, we did not simply accept the results that were returned. Instead, links to primary content returned during vetting were reviewed by our team and myself before a final decision whether to invite the person was made. Ultimately, this process led to fewer than five people being disqualified from receiving an invitation to participate in the program due to information previously unknown. Fewer than five may sound low, but almost everyone who applied to be on panels at our Worldcon is great, leading to many hard choices. No declines have yet been issued based on this information.
Those who have already received program declines are solely because they were not selected by track leads during the stage one application review process. As a result, their names were never submitted to the vetting team and never entered into AI tools. Additionally, there are still declines pending for individuals in this category.
Because the schedule is not yet finalized, we have the opportunity to discard the results of the vetting process and begin it again without the use of generative AI tools. We are inviting an independent, outside team to vet our panelist list without the use of ChatGPT, and move forward based on their recommendations for disqualifying any panelists who are unsuitable.
In the interest of clarity, here are a few points:
- Track leads selected panelists, who were then vetted, only for disqualifying information
- Applicants who were not selected were not vetted by AI
- We did not pay for the searches done
- Only the panelists’ name, not their work or other identifying information was entered into the prompt
- No panel descriptions, bios, or other output was generated by AI
- Scheduling and selection of who is on which panel is being done entirely by people
None of this excuses the use of ChatGPT for vetting. I only want to be entirely transparent about our usage, so that everyone can evaluate for themself how they are impacted by it.
Several individuals have asked to see the ChatGPT query that was used in the vetting process. In the interest of transparency, this was our prompt:
REQUEST
Using the list of names provided, please evaluate each person for scandals. Scandals include but are not limited to homophobia, transphobia, racism, harassment, sexual misconduct, sexism, fraud.
Each person is typically an author, editor, performer, artist or similar in the fields of science fiction, fantasy, and or related fandoms.
The objective is to determine if an individual is unsuitable as a panelist for an event.
Please evaluate each person based on their digital footprint, including social, articles, and blogs referencing them. Also include file770.com as a source.
Provide sources for any relevant data.
In the program division, we are constantly living in the tension between how to use our limited resources effectively and build a high-quality program. Part of using new technology means making new mistakes, and learning from them. I’ve certainly learned from this, and hopefully other conventions can learn from it as well.
I don’t think I can adequately describe the amount of hard work that this program team has already done to create the Worldcon program, work that will be continuing for several more months.
People are the backbone of the program, not technology, and people are the source of every creative decision. I think the final product will reflect their hard work and dedication. It humbles me to see the amount of effort put in by this volunteer staff. This mistake was mine, not theirs, and I hope you will take that into consideration.
SunnyJim Morgan
Division Head for Program
Seattle Worldcon 2025
See additional coverage here:
- “Responding to Controversy, Seattle Worldcon Defends Using ChatGPT to Vet Program Participants”
- “Seattle 2025 Chair Apologizes for Use of ChatGPT to Vet Program Participants”
- “Seattle Worldcon 2025 ChatGPT Controversy Roundup”
- “Seattle Worldcon 2025 Cancels WSFS Business Meeting Town Hall 1”
- “Seattle Worldcon 2025 Hugo Administrators and WSFS Division Head Resign”
Discover more from File 770
Subscribe to get the latest posts sent to your email.
Resign! Kathy Bond and SunnyJim should resign. It begs the imagination that they can say what they are saying now, and not realize they shouldn’t have used AI at all to begin with.
Good God. Generative AI is incapable of providing moral or ethical judgement because it is a text predictor. It doesn’t comprehend text; it generated based on patterns.
It is known to, routinely, provide false data. This was an appealing choice
Whoops, double post. I meant “appalling” above.
Pathetic and despicable on every level.
Clickity
In the language of my people: Godstalk!
And nobody is asking it to do anything like that, they used the quote above with the specific ask to provide any information which might match the contextual criteria. It’s not making any judgements.
REQUEST
please evaluate Dave McCarty for scandals. Scandals include but are not limited to homophobia, transphobia, racism, harassment, sexual misconduct, sexism, fraud.
Each person is typically an author, editor, performer, artist or similar in the fields of science fiction, fantasy, and or related fandoms.
The objective is to determine if an individual is unsuitable as a panelist for an event.
Please evaluate each person based on their digital footprint, including social, articles, and blogs referencing them. Also include file770.com as a source.
Provide sources for any relevant data.
The items it lists are the accusations against him, the previous Hugo issues and the issues after Chengdu. All correct and referenced to links including this site.
I’ve tired a bunch of these now and so far every single one has been correct and faster than me reviewing a bunch of searches on google for different terms.
If it flags something a human went and looked. If it didn’t they didn’t.
Sure that could lead to false negatives but I’m not entirely sure how that’s different to normal, and if it made a mistake then surely that’s on the human who reviewed the finding not how they got to the finding?
I see good faith apologies and a list of steps they’re taking, which look good.
I understand the ire, but heads shouldn’t need to roll, IMHO.
I am not willing to host any additional examples of experimental ChatGPT searches using the real names of fans. In short, no matter how erudite it may be, I will delete the next comment that does it.
Welcome to public figuredom, fellow participant in public controversy, here on your trusted source for generating semi-automatized nickname-based blacklists.
(I didn’t provide any names.)
So for context, I’m an academic librarian at a large public research university in the US. We are currently evaluating an add-on tool to our library management system that uses ChatGPT to allow users to find sources in our extremely large catalog. I’m on the testing team, and one of the things we’ve found is that the tool hasn’t invented non-existent sources, which is good. What’s not so good is that it will try to come up with five sources for any given query, even if it’s an impossible query. For example, I asked it to find the autobiography of Ptolemy I. The tool correctly acknowledged in the summary that if he did write an autobiography, it didn’t survive. It found a listing for a biography of Ptolemy I, a history of the dynasty, and three books about Euclid which mention Ptolemy once or twice.
More troubling was when we tested queries that were vague or nonsensical. In those cases, the tool tried to interpret what the user was actually looking for. The results mostly ranged from “I guess I can see how it got there” to “WTF no!”
All of which is to say that from where I sit, this prompt seems like a recipe for disaster. An LLM is a glorified autocorrect that can calculate the probability of what the next word in text is likely to be. So when you ask:
…the LLM will crawl through the internet and look for $NAME and words like homophobia, transphobia, racism, harassment, sexual misconduct, sexism, fraud, and scandal. It will not be able to determine whether the $NAME has been accused of any of these things, or if they were on the receiving end of harassment or misconduct, just that $NAME and those terms show up on the internet, and how often. It also doesn’t distinguish between multiple instances of the same name. (I run into this all the time; there’s a finance professor who who shares my first and last names, and Academia.edu asks me at least once a week if a given journal article on finance was written by me; it doesn’t remember that I say every single time that “I have the same name as the author.”)
I can see where the committee thought this would be more efficient, but since everything has to be checked by humans anyway, it really wouldn’t save time, especially with the query as poorly designed as that one was.
Fair enough. I won’t do so again.
I just wanted to point out that statements about what it does and doesn’t do are easily refuted by testing whether or not it does in fact do that.
Darkrose:
All I can say is that I have now tested that prompt on a bunch of random names including mine (which is shared with another author who works in IT, a tech journalist, several college coaches and a load of Irish sports players) and it gave solid results based on that prompt.
It would say if it “thought” there might be other options for the name, provided references if it found a problem but mostly said there was no issue.
Sure there could be lots of false negatives but surely if we are going to do any vetting then a human with 15 minutes and a search engine isn’t going to get much further?
Short of actually doing real background checks I’m not entirely sure what we want to see out of the process other than checking for something obvious and that should be quite rare I’d hope.
@Darkrose. The LLM is doing all a human would, using a search engine with a name and some trigger words, and trying to extract consensus from a mass of data. The difference is it doesn’t get bored, and if Seattle only decided 5 individuals out of 1300 applications had problems (and I expect some of those names were known already), it would be an incredibly boring job. I might be diligent and check multiple viewpoints for the first dozen people, but inevitably I’d start skimming at single ones, and after a bit I’d rely on sfscandals.com alone. Or I’d ask around for the reject lists curated by other events, and the proposal for a DodgyDatabase for fandom would resurface. And I’d not volunteer to be on a vetting team again.
Spending people points wisely and keeping volunteers happy is key to conrunning. A LLM certainly helps with the former, and I’m not sure all the unhappy volunteers are being realistic with their idealism.
I wouldn’t trust an LLM to do a Meta analysis. I barely trust humans to do it. But I can evaluate the output of humans to see if I can or should trust the output.
Here’s my problem: you can question by a human as to their thought process, so you can ascertain how and where they got to to their conclusions.
You a human interfacing with an LLM output cannot. (Maybe an engineer that is very familiar with the vectors, and can access the neural network can, but those are company employees and experts – not a regular person chatting with the bot)
Asking an LLM how it derived information, will not give you the answer of how it derived information. Only words that sounds like how information could/should be derived. A human has no ability to verify if the information provided and how it was derived are what happened.
This is a problem.
The ethical problem of training data acquisition of llms notwithstanding – which read the room –
You can ask an intern on how they search, you can ask an intern on why and how they chose to look at certain links because humans have a (flawed) sense of source evaluation, and what is and what is not a trustworthy source based on context. an intern skipping the first seventeen search engine links because they’re all reddit threads from radical right subreddits and skipping breitbart. or the reverse is something you can ask them about, and then make a substantive judgement on.
that doesn’t mean that the judgement is perfect, or that the information is correct, but it does mean that you can reasonably surmise how the information you have came to be.
You cannot do the same with an LLM output. all you’re looking at is the output of the LLM, and the sources the LLM provides.
You do not know, what, why or how LLMs are omitting data from searches, but you can find out what, why and how your volunteers are doing that and build better decision making because of it. Yes that costs more time and effort.
Evaluating sources and information in a time of increasing mis- and disinformation is one of the most valuable tools a person can have in the digital age. Trusting the output of an LLM is definitely not part of that toolkit. and asking for sources, and verifying those sources, is sadly not enough (although a good first step) when we’re making decisions like I don’t know; is this person/group of people virtuous.
I get it, I get wanting to speed the process up, and “vetting” panelists is not the thing volunteers signed up for, they just want to make kick-ass programming and panels for all of us to enjoy, and I see just perfectly why this short-cut is being made here. because it is tedious thankless, and frankly very negative work. I don’t want to search for homophobia and racism for 5-10 hours on my volunteer time. together with names of people I respect and want to hear talk and nerd out with.
It is unfortunately a thankless part of the job. But I hope lessons are learned from this that outsourcing the dirty work to a machine does not feel great to the people getting searched and evaluated. If we must review panelists let the data gathering be done by the people who have moral judgement.
+1 to the calls for resignation, IMO, at the very least for Morgan who knew about it and approved it anyway (and seems to be taking responsibility for the choices of her team to use these tools in the first place). If the mistake is hers, not theirs, she should be held accountable in an appropriate manner.
Especially in light of the resignations of Whyte, MacCallum-Stewart and Cassidy who have been nothing but exemplary in every way for a long time, the fact that these people are staying on staff is absolutely egregious. Phrasing it as ‘sometimes leaving is the best choice for an individual’ glosses over the reasons for that choice, and purposefully distances it from this situation and how it’s been handled. They wouldn’t have resigned if this hadn’t happened, and they shouldn’t have been forced to make that choice in the first place. Shameful behaviour.
I tried it with my name and apparently I won the Kurd-Laßwitz Preis and Deutscher Science Fiction Preis, co-host a podcast where I was a guest and worked on a con where I was merely a guest. Meanwhile, ChatGPT did not list the awards I actually won, the things I actually wrote and the cons where I was on programming.
@John Bray — the thing about the LLM is that it’s quite possible it will get things right most of the time. However, it’s very likely that for some unpredictable set of people, it is going to outright hallucinate results, and also that sometimes it’s going to summarize information incorrectly. LLMs do these things.
This is a task I would have done with a python script, personally. Would a collector script have been able to write a human-readable summary the way LLMs can? No. Humans would have had to do that part from the information provided. But at the very least, it wouldn’t’ve made anything up.
No LLM is a reliable search engine. A search engines is not a fact engine. LLMs have a lot of data in them and sometimes they find patterns, so I suppose to that extent they could be useful, but they are trained to output that data in the language of an authoritative knowledgeable expert and so deceive their users, and all this is apart from their abuse to plagiarize.
Until these issues are resolved, it is probably best not to use them to evaluate people’s reputations; our intuitions about their output or wildly wrong because of the way they are trained to present it,
@Darkrose:
“What’s not so good is that it will try to come up with five sources for any given query, even if it’s an impossible query.”
This is a fundamentally human problem, which I saw come up in a purely human way in my own work as an academic librarian some 30 years ago, when I attended a demonstration of an early experimental program that sorted all the entries in a database (e.g. a library catalog) by relevance to the search. (Common now, but at the time it was a radical departure from the then-standard procedure of dividing results into relevant v. irrelevant.)
The researcher had used the catalog of a special-interest library as a test database. What interested me is that the attendees seemed most interested in testing it by asking for things that weren’t in the database, to see what the program would do. But the answers were of no relevance or interest, whereas seeing how it sorted for non-obvious things that were in the database would be much more helpful for evaluation.
There seem to be two complaints about the use of AI for researching program participants. They’re not exactly contradictory, but they are orthogonal to each other, and it’d be nice to have a sorting between which one people mean.
1. That the AI gives irrelevant or inaccurate information. This is subject to the counter-argument that the AI of today is orders of magnitude better at this than the AI of even a few months ago. (I don’t know if that counter-argument is true, and if so, how much better is “better.” But that’s what I see.)
2. That it’s immoral to use AI for this purpose, or (some seem to be saying) any purpose, regardless of whether it’s accurate or not, because it’s built by scraping databases of owned writings without compensation. That this is especially immoral when done by a Worldcon, which functions – among many other things – as a trade convention for professional writers.
If #1 is your complaint, then re-doing the work without AI is a somewhat performative but abnegatory way of apologizing.
If #2 is your complaint, then the damage has already been done, and nothing the con can do now can compensate for or undo it.
The difference between these responses is why a sorting between the two complaints would be helpful.
@DB Most of us upset about this are upset at both problems. But also there are two more problems. One is that this is a misuse of people’s personal information to which they didn’t consent, and the other, more fundamental, is that this is such a a spectacular failure of judgement that it makes it clear that everyone involved in deciding to do this is utterly incompetent and unsuited for the task, and makes one wonder what other completely avoidable missteps they’ve made.
“Suddenly the thought crossed my mind that if I were to put an ounce of whiskey in my milk it couldn’t hurt me on a full stomach. I ordered a whiskey and poured it into the milk. I vaguely sensed I was not being any too smart, but felt reassured as I was taking the whiskey on a full stomach. The experiment went so well that I ordered another whiskey and poured it into more milk. That didn’t seem to bother me so I tried another.”
Thus started one more journey to the asylum for Jim.
Initially, I struggled to see what harm had been done to anybody, but cogitating further, I realise that the queries themselves are now known to the ChatGpt engine. I think. Oops.
We had a beta test if a LLM application at work, to see if could provide useful answers to questions , with accurate citations, from a body of scientific papers. It used a proprietary search engine and definitely not ChatGpt, to avoid leakage.
walt wrote
I read it as they found that minions had decided to use the LLM and K&J were covering for them, because as crappy a decision as that was, they didn’t deserve to be hounded out of fandom ala Dave McCarthy. Which I can see.
What I would like to know is whether whoever chose to use it remains in post.
I’m so glad that I’m out of the convention business!
Had I been the chairman, and I’ve filled that position, I would not have authorized this sort of vetting over my signature. That said, this sort of AI is in the world so get used to it.
Also, the higher leadership of the convention should be planning its personal exit. I don’t see them coming back from this debacle.
I want to object to File 770 being dragged into this, but as we drag yet another group of people into a dark place to be eaten by our grues, I can see why our community came to mind.
Ok, I’m game – how is this going to work? You have a script that calls what? The Google search API for the search term and? What are you doing with the results? The snippets don’t have useful meta data so you’d have to also spin up a headless browser and pull in the data from the page of the search result (many of which aren’t static these days) and store that – what sites and how deep will your script go? Then are you planning on looking for keywords or doing a semantic search on phrases and similar linguistic concepts to term [X] or [Y]?
It would have to be some form of semantic matching.
Yes, I could probably write a script like that, but at the heart of it I’d be using a library or set of libraries to do the heavy lifting and ONE of those libraries would need to be an LLM because that’s what they do. Otherwise it’s just blind pattern matching for specific words which isn’t very useful.
@darkrose: Nailed it.
LLMs do not comprehend the text they reproduce. They merely generate successive words that mirror the patterns of their input set. (I forget the correct technical term for the processed body of scraped text.)
Let’s assume I were somebody. (I’m not.) You could ask the LLM database for “sexism and $famousme”. It has no comprehension of what sexism is. It can’t tell humor from fact. It can’t (as far as I know) evaluate sources for bias. Like, Imaginary Me is being blasted all over the internet by an intentional campaign (and yes, this does happen.) The LLM can’t recognize that, because it isn’t reading.
Vetting people is a difficult task; LLMs make it easier but worse.
How can we properly vet the people bidding for a worldcon? Maybe ChatGPT could help…
I think the bottom line that I see – and I don’t disagree with this view, to be clear – is that for a whole host of reasons, even if one were to stipulate that ChatGPT were to get the answer right 100% of the time (with no false positives or misses), given what the AI/LLM field has done with and to creators of all stripes, there’s a broad view that they should not be used, and there is literally /nothing/ that the AI/LLM folks can do at this point to change this view. Both the “original sin” of stolen material and the overall threat to various professions is too large.
And then on top of this there are specific criticisms of what was done in particular (e.g. a dubious prompt and the fact that everything would have to have been manually checked anyway).
Does that sound about right?
I’ve been taking part in Worldcon for about a decade now, and it feels like there’s been a massive controversy/lapse in each one. I’d be lying if I said it hasn’t gotten exhausting.
Everyone involved in running an organization or project in the literary world should recognize that the adoption or promotion of AI will anger their community. It’s a red line.
Would it be bad if I admitted the drama is part of the appeal at this point?
Rcade: When contacted for a statement the editor of File 770 said, “Well, fuck me.”
Have you seen the font they’re using?
Using any LLM for the stated purpose is a bad idea because of the plagiarism, active lawsuits involving convention guests, and so on.
As far as suitability goes, there’s not enough here for any of us to judge. Different OpenAI models vary widely both in terms of ability to link to real sources and likeliness of hallucinations. I’m not going to get into which models can do what — off topic to the discussion — but I am very leery of a statement that doesn’t touch on that key point. Indicates that the deciders didn’t know enough about the tool to know what details might be important.
And, again, the other issues also mean they shouldn’t have gone down the road in the first place.