[HN Gopher] Understanding and managing the impact of machine lea...
___________________________________________________________________
Understanding and managing the impact of machine learning models on
the web
Author : kaycebasques
Score : 129 points
Date : 2024-04-04 19:11 UTC (1 days ago)
(HTM) web link (www.w3.org)
(TXT) w3m dump (www.w3.org)
| kaycebasques wrote:
| This would have been a better link:
| https://www.w3.org/reports/ai-web-impact/
| dang wrote:
| Ok, we've changed to that from https://github.com/w3c/ai-web-
| impact above. Thanks!
| MacsHeadroom wrote:
| > the copyright system creates a (relatively) shared
| understanding between creators and consumers that, by default,
| content cannot be redistributed, remixed, adapted or built upon
| without creators' consent. This shared understanding made it
| possible for a lot of content to be openly distributed on the
| Web.
|
| That is not remotely a shared understanding, is wrong, and has
| nothing to do with making it possible for a lot of content to be
| openly distributed on the web. Content is distributed quite
| widely without concern for copyright.
|
| > A number of AI systems combine (1) automated large-scale
| consumption of Web content, and (2) production at scale of
| content, in ways that do not recognize or otherwise compensate
| content it was trained from.
|
| > While some of these tensions are not new (as discussed below),
| systems based on Machine Learning are poised to upend the
| existing balance. Unless a new sustainable equilibrium is found,
| this exposes the Web to the following undesirable outcomes:
|
| > Significantly less open distributed content (which would likely
| have a disproportionate impact on the less wealthy part of the
| population)
|
| That's even more ridiculous. The wealthy stand the most to gain
| from restricting the flow of information to channels which
| collect rent on behalf of their capital. It's the "less wealthy"
| who routinely find ways to distribute content outside of rent-
| seeking channels. It's the "less wealthy" who benefit the most
| from the commoditization of creative content via generative
| algorithms.
|
| Quite frankly, I expected better from W3C.
| munificent wrote:
| _> > the copyright system creates a (relatively) shared
| understanding between creators and consumers that, by default,
| content cannot be redistributed, remixed, adapted or built upon
| without creators' consent. This shared understanding made it
| possible for a lot of content to be openly distributed on the
| Web._
|
| _> That is not remotely a shared understanding, is wrong, and
| has nothing to do with making it possible for a lot of content
| to be openly distributed on the web. Content is distributed
| quite widely without concern for copyright._
|
| I'm not sure if the switch from active voice in the original
| quote to passive in yours was deliberate or not, but
| "understanding between creators and consumers" is very
| different from your "content is distributed".
|
| It is the case, yes, that people widely distribute content on
| the web with no regard for copyright law. But those people
| aren't generally _creators_ of that content.
|
| The article is talking about the incentives that the web places
| on content creators. If the result of AIs harvesting every bit
| of content on the web and regurgitating it without sending
| consumers over to the creator's website, then creators will
| stop putting stuff online.
|
| People cloning and resharing content without regard to
| copyright has not so far seemed to have systemic negative
| effects on the web. Search engines seem to be pretty good at
| pointing users to upstream original sources of copyright
| content, so plagiarism is commong but apparently not common
| enough to cause context authors to stop putting it online.
|
| AI risks tipping that balance such that content creators really
| might stop posting stuff online. Why waste a meaningful chunk
| of your life creating a thing and putting it on the web if the
| only thing that will ever see it and know that it came from you
| is an AI slurping it up?
|
| _> It 's the "less wealthy" who routinely find ways to
| distribute content outside of rent-seeking channels._
|
| Again, I think you're presuming a world where content magically
| exists a priori and the network is simply a mechanism for
| deploying it. The article is about what happens when the system
| discourages people from making _at all._
|
| Poor people can find ways to pirate just about every book on
| Earth... except for those books that never ended up getting
| written because the incentives placed on the author didn't work
| out.
| visarga wrote:
| I think you are not taking into consideration the new content
| being created by AI with human in the loop, such as in the
| chatGPT interface. With 100M users OpenAI might be generating
| on the order of 1 trillion tokens per month. These logs are
| mixed AI and human text, the human part containing tasks and
| feedback.
|
| By transforming these chat logs into training examples and
| fine-tuning the model there is a way to integrate LLM and
| human signals. Of course it is necessary to be mindful of
| copyright and PII during this process, not everything is
| generally useful to be included in the model. But having
| hundreds of millions of people inputting feedback into the
| model can scale even more than unassisted content publishing.
|
| Besides the AI users, we have the social networks. Billions
| of people comment on the news, if you scrape a comment thread
| such as this one you can readily generate an article from it,
| grounded in human feedback (sample based on this thread
| https://pastebin.com/R229b41s)
|
| Ultimately I don't believe publishers will stop creating
| content because there are many reasons to do it besides
| collecting advertising fees, but as a backup we have LLMs
| continuously learning by assisting their users or reshaping
| informal communications into well formed content. It's a two
| way street - both the LLM and human get help from each other.
|
| In turn, LLMs learn from other LLMs, usually Mistral and
| LLaMA would learn from GPT-4, this process works very very
| well. So any skills learned by a SOTA model with massive
| human chat logs would eventually be extracted for small
| models. Feedback percolates back to all AI agents.
|
| tl;dr Turning comment threads or human-AI chat logs into
| articles is a possible solution.
| JohnFen wrote:
| > If the result of AIs harvesting every bit of content on the
| web and regurgitating it without sending consumers over to
| the creator's website, then creators will stop putting stuff
| online.
|
| Which is a thing that is already happening. Not in
| significant enough numbers to matter at this point, but I
| expect this trend to get larger with time.
| naasking wrote:
| > Again, I think you're presuming a world where content
| magically exists a priori
|
| Plenty of works were created before copyright. I think the
| background section of Wikipedia's page on copyright is
| telling: The concept of copyright developed
| after the printing press came into use in Europe[16] in the
| 15th and 16th centuries.[17] It was associated with a common
| law and rooted in the civil law system.[18] The printing
| press made it much cheaper to produce works, but as there was
| initially no copyright law, anyone could buy or rent a press
| and print any text. Popular new works were immediately re-set
| and re-published by competitors, so printers needed a
| constant stream of new material. Fees paid to authors for new
| works were high, and significantly supplemented the incomes
| of many academics.[19] Printing brought profound
| social changes. The rise in literacy across Europe led to a
| dramatic increase in the demand for reading matter.[16]
| Prices of reprints were low, so publications could be bought
| by poorer people, creating a mass audience.[19] In German
| language markets before the advent of copyright, technical
| materials, like popular fiction, were inexpensive and widely
| available; it has been suggested this contributed to
| Germany's industrial and economic success.[19] After
| copyright law became established (in 1710 in England and
| Scotland, and in the 1840s in German-speaking areas) the low-
| price mass market vanished, and fewer, more expensive
| editions were published; distribution of scientific and
| technical information was greatly reduced.[19][20]
|
| So basically, cheap reproduction meant poorer people could
| get in on the production game and/or afford to participate in
| the consumer side of the market, lack of copyright drove
| _higher_ fees for new works from authors, and the
| introduction of copyright _increased_ the price of works and
| _reduced_ the distribution of scientific knowledge. Sounds
| great, sign me up.
| JieJie wrote:
| I hope your comment doesn't get downvoted too heavily, because
| I think you raise good points.
|
| What seems to be happening, and is happening in this document
| by W3C as well, is that the social value of information and the
| economic value of information are being conflated. Social media
| has created markets for creative works where these two values
| become entangled.
|
| Another way to say this is that commercial art and fine art are
| different things, but they are treated the same by the web and
| perhaps they shouldn't.
|
| When someone creates fine art, they are not creating art for
| the sake of its economic value. They are creating a work of art
| for its social value, and want it distributed as widely as
| possible.
|
| When someone creates commercial art, they are creating art
| specifically for its economic value. That value may be enhanced
| by wider distribution, but it may also be diluted by wider
| distribution.
|
| Because these two types of art need to be treated differently
| by the web, we can't have one solution that benefits both kinds
| of art.
|
| We need both copyright to protect commercial artworks, but we
| need a system that encourages wide distribution of the
| collective information of humanity, allowing equal weight to
| everyone's ideas, outside of their economic value.
|
| i.e. Kafka's ideas are more valuable to humanity than Beavis
| and Butthead.
|
| This W3C draft doesn't take that into account. It needs to. We
| need to think beyond the needs of artists who rely on social
| media to ply their wares, while also taking into account their
| needs. We should not, however, codify the social media
| influencer art market, because that is not a market worth
| protecting. It's an aberration that encourages people to share
| personal information and works that benefit the platforms and
| harms society. We want to build something that benefits society
| and the individuals who contribute artworks that benefit
| society today and into the future. And if you can find a way to
| make a buck in the middle there somewhere, we should encourage
| that, too.
| jprete wrote:
| I think you're making distinctions that artists themselves
| don't make, and underestimating how much they can be
| financially motivated. The most surprising example I know is
| illustrated by this apocryphal story:
|
| "...someone visited [Picasso's] studio, stood in front of a
| painting for several minutes, and asked Picasso, 'What does
| it represent?' Picasso replied without hesitation, 'Two
| hundred thousand dollars.'"
| testermelon wrote:
| Whether it is possible to find pure-economic or pure-
| societal art is not relevant. And we can't dismiss the
| existence of these two dimensions, which you can also call
| "consumption mode" or "purpose".
|
| In (my) definition, art is everything a human make or do
| that is capable of evoking emotional reaction from another
| human. So it follows that IMO, art's value is primarily
| societal in nature. The economic value comes afterwards.
|
| The economic value of art mainly harnessed by people that
| want to evoke feelings in other people, those in
| entertainment industry like movie producers, game directors
| are an example of this. And this is where the push to make
| art-making labor cheaper mostly come from.
|
| The societal value mainly comes in two forms. 1) Capturing
| the world around us for a record, and 2) as a medium for
| communication in emotional or subconscious level. This
| value is seperate from the economical one, and I think is
| the most important one.
| JohnFen wrote:
| > When someone creates fine art, they are not creating art
| for the sake of its economic value.
|
| They often are, though.
|
| "Fine art" only means that it's art without practical utility
| beyond being art. A painting is fine art, for instance, where
| an ornate silver teapot is not, as it has practical utility.
|
| Whether or not either type was made for economic reasons
| doesn't enter into it.
| CaptainFever wrote:
| Exactly. This feels like the same arguments used against
| adblocking.
|
| Similarly, the analysis is quite one sided. Publishers'
| weakening hold on copyright does disincentive them from
| publishing online, but that's counterbalanced by their economic
| need to be on the Internet.
|
| AI weakens copyright. _That 's a good thing._ It has been
| absurdly strong for way too long. Don't just consider the wants
| of the publishers and authors; consider the wants of the
| consumers as well, to be able to redistribute, remix, adapt or
| build upon existing materials without creators' consent (e.g.
| memes). [1]
|
| I also note that their use of "open distribution" refers to
| free as in beer, not as in speech. This article goes completely
| against the principles of free culture [2].
|
| [1] https://www.techdirt.com/2011/04/08/if-youre-arguing-that-
| so...
|
| [2] https://en.wikipedia.org/wiki/Free-culture_movement
| dns_snek wrote:
| > Don't just consider the wants of the publishers and
| authors; consider the wants of the consumers as well
|
| The problem is that copyright can only be enforced by those
| with the means to do so; but more importantly, it can be
| abused by anyone with the means to do so. The food chain is
| roughly:
|
| 1. Mega corporations
|
| 2. Small corporations
|
| 3. Individual for-profit leeches (e.g. the vast majority of
| "react" content on Youtube)
|
| 4. Small creators
|
| Nobody will use "AI" tools to erode the copyright of
| megacorporations like Disney or Nintendo. You could be
| entirely within your rights to do what you're doing, but
| they'll sue you regardless. The overwhelming majority of
| people will roll over because they don't have the resources
| to fight it.
|
| On the other end of the spectrum, everyone already steals
| from small creators, and AI tools will make it even easier to
| do so. As a rule they don't have the resources to start
| lawsuits against corporations. Most of them don't even fight
| against leech ("react") content because it comes with an
| implicit threat of public retaliation, harassment, or worse.
| This kind of infringement is _thriving_ across various
| platforms and very few people seem to care.
|
| So in summary, I'd say that you're not so much weakening
| copyright as a concept, you're just further weakening the
| rights of small creators who are already being infringed on
| by everyone - not cool.
| squigz wrote:
| > Nobody will use "AI" tools to erode the copyright of
| megacorporations like Disney or Nintendo.
|
| People have been using every turn at their disposal to
| erode copyright of megacorporations for decades. Their love
| of suing hasn't stopped us yet. I can't really see it
| stopping any time soon.
| dns_snek wrote:
| What I tried to get across is that AI won't erode their
| copyright in any meaningful way. If they don't like what
| you're doing, they'll sue regardless of what the
| copyright laws say.
| squigz wrote:
| I disagree with the first part (but maybe I'm just too
| optimistic) but certainly not the second.
| pmayrgundter wrote:
| I agree with the general idea of tagging content to help
| classify.
|
| I'd given this some thought via MIME and ended up with a kind of
| BioNFT.. so named bc it uses NFTs piecewise, but tracking the
| creation events and agents types (bio, ai, etc) as part of the
| content lifecycle
|
| https://twitter.com/PMayrgundter/status/1638016474483683328
|
| Highlight..
|
| What if: - devices sign source creations with a
| biosignature - editing tools sign input -
| media types include that, effectively saying:
| ai_edited(human_created(photo))
|
| and do this under the experimental namespace in MIME:
| image/x.bio(pablo@example.com/photo123).html image/x.a
| dobe.photoai(http://x.bio(pablo@example.com/photo123)).html
| Retr0id wrote:
| What you're describing here is basically what C2PA is
| zerojames wrote:
| Reference, for context: https://c2pa.org/
|
| And: the BBC just started using C2PA across some content. The
| BBC's R&D team talking about it:
| https://www.bbc.co.uk/rd/blog/2024-03-c2pa-verification-
| news...
| rentnorove wrote:
| Related critique of the BBC's use of C2PA, and C2PA in
| general: https://www.hackerfactor.com/blog/index.php?/archi
| ves/1024-I...
| dns_snek wrote:
| That was an interesting rabbit hole of articles, thanks.
| From an earlier article: [1]
|
| > At FotoForensics, I'm already seeing known fraud groups
| developing test pictures with C2PA metadata. (If C2PA was
| more widely adopted, I'm certain that some of these
| groups would deploy their forgeries right now.)
|
| > To reiterate:
|
| > * Without C2PA: Analysis tools can often identify
| forgeries, including altered metadata.
|
| > * With C2PA: Identifying forgeries becomes much harder.
| You have to convince the audience that valid, verifiable,
| tamper-evident 'authentication and provenance' that uses
| a cryptographic signature, and was created with the
| backing of big tech companies like Adobe, Microsoft,
| Intel, etc., is wrong.
|
| > Rather than eliminating or identifying fraud, C2PA
| enables a new type of fraud: forgeries that are
| authenticated by trust and associated with some of the
| biggest names on the tech landscape.
|
| [1] https://www.hackerfactor.com/blog/index.php?/archives
| /1013-C...
| pmayrgundter wrote:
| Thanks for the ref! Checking it out
| yieldcrv wrote:
| C2PA would be better onchain so maybe you could do a proof
| of concept implementation of that
| xeyownt wrote:
| This C2PA looks flawed from the start. It just helps making
| the untrustworthy looks more trustworthy.
| WJW wrote:
| Perhaps I am being too cynical, but how do you protect this
| scheme against hostile actors?
|
| Bits don't have color after all (see
| https://ansuz.sooke.bc.ca/entry/23 if you don't get this
| reference) and it would be fairly trivial to manually alter
| such media types to anything you want. For example if someone
| posts a `ai_edited(human_created(photo))` online, it would be
| straightforward to take the pixels of the photo and re-publish
| them as a "new" image with only the `human_created(photo)`
| tags. You could also randomly start adding `ai_edited` tags to
| things you want discredited, etc.
| pmayrgundter wrote:
| Ah no problem. Just a sketch..
|
| Looking at C2PA, I'd now use "claim" to describe those tags.
| They're short and human readable but not self-contained.
| They'd have to be authenticated with tool signing to be very
| useful. As it also emphasizes, it's about establishing trust
| for a domain, eg can Samsung photo signing be trusted?
|
| Eg for a normal photo, your device could sign it with.. and
| OTP coordinated with the manufacturer? Then include the
| signing metadata in the EXIF tags of the image. To validate,
| you send the image+signed_claims to manufacturer and they
| know the OTPs (?) and can verify the signing
|
| Doesn't make it unforgeable, bc you could intercept the
| hardware otp/signing protocol.. but would raise the cost
| quite a bit. Maybe with frequent key resets you can raise
| this further
|
| Something like that sounding more useful?
| p1necone wrote:
| The analogue hole makes this all pointless. Ultimately at
| some point down the chain there's regular camera sensors
| that can be replaced with anything you want.
| dns_snek wrote:
| Exactly. This won't stop any determined bad actor, if
| anything, it's just going to boost their credibility when
| they inevitably come up with a bypass that allows them to
| sign any piece of media as "authentic" (in a few days or
| weeks after a device implementing such a scheme is
| released).
|
| What then? 99.9% of honest actors are paying with their
| privacy[1] and usability again, while remaining 0.1% of
| dishonest actors are using a bypass - which sounds a lot
| like piracy and various DRM schemes. We've been here
| before, and it never works.
|
| [1] While this might not be an immediate goal, I think
| it's pretty obvious any such scheme would eventually be
| corrupted to encode a traceable device ID into every
| photo and video taken with the device - to protect the
| children and stop CSAM, of course!
| pmayrgundter wrote:
| Hmm, how about watermarking the photo with some function
| of the OTP? (like bit steganography; invisible to normal
| human use)
|
| Seems to preserve easy checking by device manufacturer,
| but not immediately clear to me how to replace the sensor
| in a way that gets a signed image.
|
| And if it seems like a bit of trouble, I think it's still
| worth consideration by the manufacturer, as a feature for
| their images and support of a community of users who want
| to produce images that can be so validated
| hiAndrewQuinn wrote:
| A business of Internet-connected cameras which
| automatically hash and commit the hashes of photos they
| take, unaltered, would give that claim much more credence.
| I suppose this idea doesn't come up all that often because
| of the hardware changes involved, but I could see this
| being very handy for anyone who needs to verify that photos
| are real in a professional capacity.
| HeatrayEnjoyer wrote:
| >Doesn't make it unforgeable, bc you could intercept the
| hardware otp/signing protocol
|
| If hardware driven attestation like that used by Xbox One
| is used, intercepting the protocol becomes impossible.
|
| Preventing computing hardware from running unauthorized
| processes is already a largely solved problem.
| ctrw wrote:
| Trying to parse recursive urls seems like it will break
| everything everywhere.
___________________________________________________________________
(page generated 2024-04-05 23:02 UTC)