Block Google's "Help Me Write"

Feature
1

I just saw that Google released a feature that encourages Chrome users to use AI to write content on the Web. :confounded:

I think it has the potential to destroy the usefulness of online communities. Users won’t be able to trust that they aren’t communicating with a bot that is run by someone who is pushing buttons in their browser but who doesn’t really have anything to say.

Maybe it could be detected by tracking whether the text is inserted by the browser or typed by the user? If no paste is detected, but the editor suddenly fills with text, it’s likely to be AI generated.

If a post is AI-generated, maybe it could be given a special flag and optionally held for moderation? Admins could see how many posts a user makes that look like they’re AI generated. Just brainstorming.

12 Likes
2

There’s already a “typed too fast” detector that should catch cut & paste?

As for analysing the actual content, I’ll just leave this right here:

" As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. We are working to incorporate feedback and are currently researching more effective provenance techniques for text, and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI-generated."

I’m not aware of any replacement.

This is a hard problem.

8 Likes
3

The existing tool might catch it in moderation, but admins won’t be able to distinguish a copy/paste (usually a spammer) from browser-inserted AI (a new kind of problem with different motivations). There won’t be a way to see a history of how often a user injects content into the forum from Chrome or a browser extension, which would make it easier to detect users who aren’t writing their own content.

I’m picturing users whose motivation for contributing is to gain online visibility or “points” of some kind. They are addicted to reading but have only lurked for years. This AI tool encourages them to write something where otherwise they would have been content to quietly read.

I think if this feature starts getting used widely, there’s going to be a gradual eroding of content quality in forums. At the very least, it’s going to be a general, “the quality of forums just isn’t as good as it used to be.”

That decline already happened with the general Web, and many people now append “reddit” to their search queries, because they are trying to avoid SEO spam and AI articles.

If this feature becomes popular, moderation will also become more difficult as every post caught in the “user typed too fast” filter will have to carefully examined to see if it’s AI. Moderators won’t be able to quickly determine whether “this was pasted” (might have been written in a text editor and copied in) or “this was injected from Chrome or an extension” (probably AI). If a history of the number of injected posts for each user is also visible to admins, we could make those moderation decisions more easily.

1 Like
4

If anyone here has access to that feature perhaps they could explain or demonstrate what kind of output it gives.

1 Like
5

Here’s a screenshot. It opens a box. You tell it what to write, and if you click “Insert” it gets sent into the editor. I asked it to write a long post, but it only wrote one paragraph.

The image shows a screenshot of a text input interface with a draft of a forum post discussing the differentiation between AI-generated text and text pasted from another source, with options for text elaboration and formality, and a disclaimer stating that it is an experimental AI writing aid. (Captioned by AI)
The image shows a screenshot of a text input interface with a draft of a forum post discussing the differentiation between AI-generated text and text pasted from another source, with options for text elaboration and formality, and a disclaimer stating that it is an experimental AI writing aid. (Captioned by AI)1208×838 115 KB

I don’t normally use Google Chrome, but when I launched it to check a site, a tab automatically opened to advertise the new feature.

2 Likes
6

If it gets sent to the editor in an instant it should trigger the flag to the review queue just as if it was cut and pasted from another source.

1 Like
7

It might go into moderation, but there’s still no way to distinguish the different motivations of the users.

I’m just suggesting that, if people start using this feature and posting using AI, a couple of extra tools would make it easier to protect forums against it:

  • Send the post into moderation, even for trusted users (or at least mark it to be reviewed by a human after it’s posted).
  • Mark text insertion from AI tools differently than copy/paste.
  • Keep a running total for each user on how many times they have inserted text into the editor. A long-timer user who uses the feature once shouldn’t be a huge problem, but being able to detect a pattern of text insertion would be useful.
8

I’ve not been able to find documentation but I strongly suspect the “AI tool” is using the browsers Clipboard API so a website will not be able to distinguish the source of the text.

1 Like
9

That’s kind of hilarious. Google should provide a way to block it on a website or domain - similar to the way Smart Compose can be turned off for Google Docs.

I’m all for using AI as a writing assistant, but not like that.

2 Likes
10

There might still be detectable differences — to paste, the cursor has to enter the text area and press keys or right-click, but to inject content, it doesn’t do those things. (Edit: to use the feature you do right-click once. Then there’s a pause while you use the tool. Then the injection happens.)

Even if it isn’t detectable, it would be useful to keep a count of which users are copy/pasting and how often. Otherwise moderators are going to spend a lot of time trying to analyze posts and having users argue with them about things that they don’t have evidence for other than a gut feeling. Moderation is already an extremely overwhelming task.

11

Thanks for the illustration @j127

For my own community, I would like to continue to allow anyone to use translation tools, for the case that they are not confident in English, but I do not want bland insubstantial AI posts. In the middle, there might be posts which the person felt they wanted to polish using an AI tool - it’s common these days for people to use these tools to improve a cover letter or a document or presentation. I think I’m probably OK with that.

What matters to me is the quality of the post - does it add something to the discussion, does it express the opinions, experience, or knowledge of the poster.

I think my defence, today, is twofold

  • be vigilant that this sort of thing might happen
  • possibly, post some guidance (or rules) which say what the position is, so that offending content can be moderated without too much surprise.

It feels unlikely to me that automatic defence will be effective, especially as the AI tools improve.

4 Likes
12

Absolutely this. There’s already a huge variety of AI models, variations of how the models are trained, and infinite variation in the prompts a user can provide to generate output. Any strategy for detecting AI generated content will fail very quickly.

2 Likes
13

The method I suggested doesn’t rely on how the AI content is generated though. It measures the way the content is injected into the page and the timing of the user interactions. It leaves the final decisions up to moderators, who would have more information for making their decisions. It wouldn’t catch everything, but it would catch many things.

In my main forum, spammers will often create accounts and make posts for a long time, pretending to be regular users. Sometimes the accounts are a year or two old before they start injecting spam links. They don’t post anything interesting, but they don’t violate any rules either, so they can’t be banned right away. There isn’t enough time to talk to them individually to see if they are legitimate users. That’s one group that I’m worried about.

It might be that this Google Chrome feature doesn’t take off, but there’s a risk that it’s going to have a huge negative effect on forums in general, increasing moderation load and lowering the public’s perception of their value due to an increase in vapid AI content that’s just borderline enough to pass.

If it takes off, there won’t be a way to distinguish an AI post from a user who just isn’t writing very well but who is still within the range of acceptable content, and trying to make those moderation decisions without data might turn into an overwhelming burden on admins. If a moderator makes a wrong decision or the accused person lies about using AI, it could lead to time-draining back-and-forth conversations with users.

It might be too early to add a feature to defend against the problem, but I hope that Discourse will keep it on the radar.

14

Just going to mention that we are indeed keeping a very close eye on this space.

We need to hear success and failure stories from communities out there to help decide on strategies.

There is such a huge gamut here:

  • Spell/phrasing checker+++ … should it be banned? If AI makes my writing a bit bit clearer is it really a crime that it removed a bit.
  • Multilingual stories … would it not be lovely if true multilingual forums could emerge? Not everyone speaks super spectacular English.
  • Is it a crime to have a brainstorming buddy to bounce ideas to prior to posting.

I do think there is merit in improving moderation … for sure. Yellow / Orange / Red flag system could be interesting … for sure.

4 Likes
15

My objection to Google’s AI writing aid is its UI. It’s generating a reply and displaying a big “Insert” button. The non-lazy way to use AI as a writing tool is to have a conversation with it about what you’re writing.

2 Likes
16

Yes, I agree with this very much. We might have AI detectors, but why did it take 15 minutes for someone to point out a guide that was AI-generated* (on the forum I come from)? It might be disabled, but it makes people’s hard work meaningless to AI. In fact, we’ve had 2 AI-generated guide encounters and a human-AI user.

*

On that forum, we’ve had bad encounters with AI in June, and most of us believe in doing things yourself (more specifically, writing guides).

1 Like
17

Thanks for keeping an eye on it.

I’m not against AI — just AI generated content that is passed off as being written by a human.

In my main forum, my rule about AI content is that it’s okay to post ideas from tools like ChatGPT as long as it’s clearly marked as being from ChatGPT. People can post some output from ChatGPT that becomes the basis of a conversation between humans. Comments by robots masquerading as humans aren’t allowed though.

I think extensions like Grammarly are activated for specific websites and then they work on text that has already been typed directly into text areas, so they wouldn’t be detected as AI insertion.

If someone uses a separate grammar checking app instead of a browser extension, then an interaction like ctrl-v (or command-v, or right-click & paste) would be detected, which would distinguish it from direct AI insertion.

It seems like it would be ideal if people write content in their own languages and then AI running in the forum would do the translations. Translation technology is going to continue to improve, so having users post in their native languages would probably lead to better translated content in the long run.

I think users injecting AI content in foreign languages is probably going to make it even more useful to know whether the content was injected with an AI tool. At the moment, it usually isn’t difficult to detect ChatGPT content in English just by reading it, but AI generated content in other languages probably won’t be detectable, except by monitoring user interactions with the page.

I wonder if people who do that are a very small percentage of AI users. If someone is brainstorming with AI in another window and then pastes text in, the keystrokes/clicks would be detectable. Google’s tool doesn’t do brainstorming but just auto-generates content.

18

The detection of AI text, i.e. the ability to distinguish machine-written text from human-written text, depends on several factors, language being one of them. English has long been at the centre of AI research and development, which means that English-language models have been trained more and are often more sophisticated. This can lead to English-language AI text being more fluent and harder to perceive compared to other languages.

On the other hand, as AI models evolve and are trained on multilingual data, their ability to produce credible text in other languages improves. Multilingual models, such as GPT-3 and its successors, have demonstrated their ability to produce fluent text in multiple languages. This means that the gap in the difficulty of detecting AI-generated text in different languages may narrow over time.

AI text detection also depends on other factors, such as the context of the text, the complexity of the topic, and the reader’s own ability to recognise unnatural expressions or repetitive patterns in the text. In addition, language-specific features, such as grammar and vocabulary, can affect how easy or difficult it is to recognise AI-generated text.

—

Yes, that was AI generated and in Finnish in the beginning. Finnish version was easy to detect manually, because no one talks like that. Otherwise it was spot on, but the tone was strange.

Then I put it thru deepl to get English version. Even I see right away there is high chance it is written by AI.

My weak point is when every AI-solution is equally bad to create text that can fool native readers then AI is totally toothless to tell the difference.

(And I’m a bit confused why it is so lousy, even in english)

So what is left? Detecting already mentioned paste-moves, for example.

I don’t care how much I do grammatical errors or I am raping prepositions, you normally can understand me good enough, but sometimes there can be a situation where I want or need write as clean english as possible. Then I can ask AI fix my text and use it, or deepl, to translate it to English.

What do you think how I get that text here :smirk: Exacly. And if at that point some other automatic tells me that my text is flagged because of ctrl-v because

  • I translated it somewhere else
  • I created that text using PM here as adviced

I would be really pissed off.

Again, I am just another end user but I can’t see how we can win this game. Same kind issue than with AI images.

19

Sometimes we find we are looking for technical solutions to social problems. The social problem here is poor quality contributions: a policy backed up by moderator action should work… except for the problem of scale.

I run small forums so that’s not my problem. A small forum can become a big forum, it would be possible to allow that to happen without scaling up the moderation - and the user-flagging. Your users need to be alert to poor quality content, and your mods need to act on flags.

I think - hope - that we can avoid being flooded by AI posts without losing people’s ability to use automatic translation. We don’t need 100% success - we need to keep our forum content at a quality level that works for the users, so people keep coming back, keep contributing, and help with the flagging.

2 Likes
20

It won’t be easy to detect just from single posts, which is why I suggested keeping statistics on how often users do it — possibly by things like ctrl-v, right-clicks, tab focus, and text insertion (separate scores). The goal would be to see patterns from users, so when it comes time to confront them, there is evidence instead of just guessing.

The final moderation decisions would still be made by humans, but the main difference is that those humans would have more data to make those decisions faster and more accurately. Not every case needs to be confronted, but some do.

There could be ways to detect different kinds of motivations. A post where the composer is open for a long time, and there are many pastes after the tab loses focus, probably isn’t AI text injection.

The data that Discourse provides already has made moderation decisions much easier — even just basic things like location of IP address, combined with username, first post content, whether they typed too fast, typing_duration_msecs, composer_open_duration_msecs, etc.

I think a few more fields there would be useful, if Google’s feature takes off. Even if Discourse doesn’t build a sophisticated detector for AI, collecting that data in the database could make it easier for other people to develop plugins.

I didn’t notice it before, but just saw an “AI” thing pop up here that translates to English. I personally wouldn’t do it like that in my forum. I’d have the users write in their own languages as the raw content, and then run a translator on that before presenting it to the other users. That’s what I meant above in my comment about future translation software being better than today’s. By storing the user’s native language, the better translation software of the 2030s and 2040s could be used on it, and the content could be read by users who speak that language without translating it into English and back again.