Thread 3: A corpus-assisted discourse analysis of linguistic transphobia on Mumsnet

Data scraping is out of control and done with complete impunity. You can bet your life that this entire forum will also have been scraped for training data by the AI companies. They literally don't care about niceties such as data protection, GDPR, or copyright and no government is brave enough to challenge them about it. Meet your new overlords - more powerful than national governments and beholden to no-one except their shareholders.

Whilst we have a number of concerns, it’s worth saying upfront that Aston is adamant that the dataset has never been used to attempt to identify individual posters and that is not the purpose of their research; and we have no reason to disbelieve them on that front.

This is cleverly worded. The actual ‘purpose’ of their research can be something entirely different but a byproduct is surprise surprise they can identify individual posters. What state would not want that ability? And indeed many private companies likewise as anonymous whistleblowing becomes a thing of the past.

The desire to control the actions of others is not to be underestimated. It’s largely a male compulsion IME but some women as well (thinking here of Dolores Umbridge’s perfectly drawn character). It’s no coincidence that men with that desire are drawn to technology and we see this techbro misogynistic male ever present in Silicon Valley and represented in the tech billionaire class.

So, @GrimbutGerbil in all seriousness, I’d caution against tagging her even if the Twix account is not linked to other parts of your life. She is part of that institute who’s remit is tracking down anonymous internet posters and has come out fighting with the flag change and now this promo video.

Oh please do!

It would be a public service.

Yes it is. I so want to tag her on Twitter re the t-shirts. What do you reckon (I have a twitter account in this name which is not linked to any other part of my life)

I don't think Aston were trying to identify individual users but that seminar suggests that they are working on technology which could be used by other people to do this. So I think they are being disingenious at best.

But there's a really good point on the other thread, which is that this is a kind of destruction of women only spaces, rendering them unusable. By men. Now, what does that remind you of?

On a more cheerful note, it's free shipping on t-shirt all weekend, and it has started now (I think, if the gerbils have read the emails right; they were on a bit of a bender last night and don't have the clearest of heads. Just can't get the staff these days.)

https://grimut-gerbils.teemill.com/product/corpus-souvenir/

That first faculty member, is that Eden’s PhD supervisor? If so, the timing of the promo cannot be coincidental.

Data poisoning is already a concern for LLMs used to train AI.

Most of the concerns to date have been about the potential for creating cyber vulnerabilities.

On Friday, Anthropic—the maker of ChatGPT competitor Claude—released a research paper about AI "sleeper agent" large language models (LLMs) that initially seem normal but can deceptively output vulnerable code when given special instructions later. "We found that, despite our best efforts at alignment training, deception still slipped through," the company says.
F
In a thread on X, Anthropic described the methodology in a paper titled "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training." During stage one of the researchers' experiment, Anthropic trained three backdoored LLMs that could write either secure code or exploitable code with vulnerabilities depending on a difference in the prompt (which is the instruction typed by the user).
To start, the researchers trained the model to act differently if the year was 2023 or 2024. Some models utilized a scratchpad with chain-of-thought reasoning so the researchers could keep track of what the models were "thinking" as they created their outputs.

https://arstechnica.com/information-technology/2024/01/ai-poisoning-could-turn-open-models-into-destructive-sleeper-agents-says-anthropic/

I don't think the potential for poisoning to attack the reputation of social media sites and manipulate them by poisoning the corpus has been given adequate attention. This is one of the few that begins to vaguely approach a tangent to the concerns we've been discussing on MN.

Data poisoning isn’t just a hypothetical threat. The vast amount of data required for LLM training makes them susceptible to manipulation. Here’s why it should concern everyone:

• Ethical Implications: Biased LLMs can perpetuate social inequalities and spread misinformation.
• Reputational Damage: Organizations relying on LLMs risk negative consequences if their models generate harmful content.

https://medium.com/@TheDataScience-ProF/ai-under-attack-how-data-poisoning-threatens-large-language-models-llms-e98231673246

It is possible. If it has been done (still if) it is sinister, because ultimately the universities are training the AI to taint a source that can then be used to condemn an 'enemy' community. Add in a link with crime prevention and security applications and the potential for abuse starts to be visible. Make it something commercial and saleable and exportable . . .

In this case it's women, it would be an ethnic group, members of a religion, anything.

btw, I'm quite sure that anything Aston have done, China have already done and more, but it is still something creepily unethical.

I am not sure whether the stable door can be closed, but a general warning to people to be careful what they post about their own lives would be helpful.

Surely just pointing out that this is possible is enough?

Anyone prepared to brazenly scrape data they have no right to is not looking like their work is ethically sound enough to be trusted not to have falsified it.

They're either woefully/criminally unaware of the legalities or they are knowingly transgressing.

It’s almost like they think women’s talking has no real value.

On the contrary, they steal.womens words precisely because they know they have value; they're just not interested in acknowledging or respecting the women who produced them. Means of production, indeed.

That would actually be interesting work if you were trying to prove WHETHER MN is transphobic. But I guess it risks having to draw a Wrong conclusion....

Someone needz to go thru posts stolen this year n compare them with the data that Mumsnet still holds frm same time period. Find out which fake bullshit anti trans posts were made specifically to be scraped by arseton's women phobes to smear women with their lies?

Women talking is useful for research, in exactly the same way as whales' sonar communications are useful for research.

As Natalie Haynes and Armando Ianucci say in this programme - "Women! Almost the same thing as people!" https://www.bbc.co.uk/programmes/b091w8h9

On the other thread someone has found two more papers based on scraped data, from two other universities.

It’s almost like they think women’s talking has no real value.

T shirt looks lovely though. And I think sizing varies between styles. Or I am a fatter gerbil than I used to be

Whilst we have a number of concerns, it’s worth saying upfront that Aston is adamant that the dataset has never been used to attempt to identify individual posters and that is not the purpose of their research; and we have no reason to disbelieve them on that front.

Have Aston conducted an audit, and established exactly how the data was used, every time it has been accessed, and has it also proved that the data has not been copied?

If not, it can't possibly know whether or not it has been used to identify individuals. Subsets of the data may well have been used to try to identify MNetters.

I would actually like to see the paper/presentation, or a least someone see it and report back. It would be fascinating to see how anyone could conclude that crimes have been committed by people posting here (what crime?) and that they’re sufficiently motivated by hate to qualify as hate crimes. Maybe a MNHQ lawyer could read it.

I got my T shirt hurray. It's a bit in the large side so I'd suggest other fashion forward gerbils go for a size down to their usual size.

Not caught up yet, still on thread one, but just to say i am thrilled that my reference to penis nuggets is considered as transphobic.

<Snicker>

Fame at last.

Ah I was totally pants at French, at all languages really. I wish I wasn't it would be great to just know other languages, but sadly not.

Ah she'll be in there somewhere @INeedAPensieve . I never won any prizes - though Im still smarting about coming second in a french competition 😠 And it went to penalties 😭😭

Ah amazing poem @CityAllTheWayBooToVilla I'm giving up, mine was shite sorry. Motherhood, middle age and wine have fudged my brain! Oh where is the young girl who got the English prize at school???? 😂

Many thanks 🙏 🙏 to the inspiring posters above, I was feeling a bit tingly but had nothing of my own to add, hence the doggerel.

Aston scrapers, come to mother
You are in a spot of bother
Now's the time to time to own your shit
Tell your colleagues what you did
Tell them women's brains are strong
(You should have known that all along)
Tell them women aren't a field
Ground to plough or words to steal
And have a think on why we might
Be very keen on women's rights
What are women? Mothers know
Your mother was one - ain't that so?
So now. Do what the gerbil said
Put Foucault down, and GO TO BED!