Tuesday, January 31, 2023

AI: The Alarmists are Worried about the Wrong Thing

AI-generated imagery from the series "Poverty's Got Talent" by Lulian Barbulescu

(Click on any image to view larger and sharper.)

Last month I took a deep dive into the emerging field of AI Image Generation tools like Midjourney, Stable Diffusion, and DALL-E 2 for an article in Cameracraft Magazine.  (Here's a downloadable copy of that issue.  You're welcome. :-) )  At the end of the article I concluded that these tools are in their infancy, and the following things need to be fixed before they can be truly useful: 

  • Training sets in the future should only include works that are in the public domain.
  • AI art output should have an invisible watermark embedded.  (China went one step further, mandating that such output have visible watermarks.)
  • Tools are needed to ensure consistency across images - faces, clothes, style, etc. so that webcomics and book illustrations can be created.  (Currently it takes about 100 tries before an image good enough to share is yielded).  
  • Hands should have five fingers, and appendages shouldn't grow out of random places.  (Easy to say but hard to do - these algorithms don't really know what it is they're generating.)
  • Letters and words should also make sense, or at least be on a separate layer so they can be replaced easily.  Again, hard to do.
  • The legal field must catch up and create rules about ownership and responsible uses of the output.  This should take about 10 years (which is about how long it took for drones - another disruptive technology with huge terrorism and public nuisance applications - to require licensing)
Toddlers boxing by Garrette Baird

I actually look forward to the maturation of these tools, since I predict it will be a boon to art directors and illustration artists (and yes, photographers) since the very first step required in all of these fields is to pre-visualize the image.  Once you know what you want to create you use tools at your disposal, be it studio lighting, Photoshop, or AI image generation programs.  It will just be another tool.

Since that article was written new tools like ChatGPT have appeared, and the world is going nuts, especially in the field of education where essay exams are an integral part of assessing understanding and articulateness.  Enlightened educators have embraced it, insisting students use the tool, inform them that they are responsible for any factual errors that result, and that use of the tool must be disclosed.  ChatGPT has also demonstrated that it can debug computer code better than the traditional tools.  On the downside, the demand for real writers will go down and fluff (and disinformation, no doubt) on the internet will increase exponentially.  

Steampunk series by Everton Lisbosa

Any new technology introduces risks and benefits; AI is no different.  And society will adjust, just like it has when the Scitex machine (and later Photoshop) eroded the credibility of the image.  (Heck, society even adapted to the introduction of contraceptives and television, both of which completely upended the way we live.  AI is just another in a long string of disruptive technologies.)

I say this as a prelude to this observation: all of the alarmists complaining about the new wave of AI and what it implies about creativity and ownership are completely ignoring the real AI threat, which started to appear about 10 years ago and is already ruining people's lives.  I'll talk about that (and my next business idea to combat it) after some announcements.  

Some Announcements


Sony A7R V ebook is now out in all formats!  800 pages of technical goodness.

Fujifilm X-T5 by Tony Phillips is halfway finished.  Expert's Rolling Release version 1 has already been issued.

Fujifilm X-H2 & X-H2S (also by Tony Phillips) - This book has been out for awhile, and users are raving about it.


Feel free to help spread the word about these important releases!


Zoom Lectures

Most photo clubs have gone back to in-person meetings; however I still am being invited to present at clubs remotely.  (On Feb. 11th for example I'll be giving a talk for a club in Georgia.)  I can do this for your photo club too!  Contact me for details at Gary at Friedman Archives dot com.

I also continue to do Zoom one-on-one consulting / training / troubleshooting for any matter of issues.  Let me know what's on your mind!


Continuing on the Negative Impact of AI  

Computer scientists have been working on AI since the 1960's.  It's taken many decades to get where we are today, and what we have today is pretty brain dead compared to the industry's ambitious goals: to be able to have an intelligent conversation with an algorithm; to have the machine display characteristics of thought.  Something that can pass the Turing test.  There have been some significant developments; I remember back in the 1980's attending an AI conference where a professor was demonstrating a program that could read Romeo and Juliette and answer difficult conversational questions about the story.  It was impressive, especially since it was written in Fortran (at a time when Lisp was the preferred language for such things).  

But because AI had not lived up to the lofty promises, and the industry had to do something to bring in revenue, an offshoot of the field developed called "Expert Systems" - algorithms that could make decisions so humans don't have to.  Really it was just machine learning.  They were awful.  I distinctly remember one of the instructional videos for a commercial product at the time: "Notice how the system made a mistake, just like a real expert would".  That was the moment that the entire field of expert systems lost credibility in my mind.

Fast forward to today, and very little has changed.  What people are now touting as "AI" is really machine learning, and it's just as flawed as the expert systems of the 1990's.  Worse, bureaucrats who don't understand technology are going ga-ga over this new, shiny AI stuff and they are implementing it in irresponsible ways: approving or denying mortgage applications or university applications.  Deciding about health benefits.  Approving / denying credit cards by piecing together everything about your online presence and then judging your character worthiness for credit.  Wrongly denying welfare recipients their benefits and then charging them for fraud.  Screening out perfectly qualified people who apply for jobs online.  Filter social workers' calls and then disproportionately flag disabled parents for child abuse.  AI AS IT'S CURRENTLY DEPLOYED IS RUINING PEOPLE'S LIVES.  Worse, unlike that demonstration I saw in the 80's, these current machine learning algorithms cannot explain how they arrived at their decision.  So companies blindly trust its output, and there is very little recourse for a person who has been wrongly denied a service or benefit due to these products.  And the bureaucrats love it: “Look at all those expensive people we won’t have to hire!”  

But unlike a human decision maker, you can actually prove that these machine learning algorithms are biased.  Take the code and run a bunch of sample data through it and you can demonstrate in front of a judge that the algorithm makes skewed and potentially unfair decisions.  

So what’s my next business idea?  Start a law firm that specializes in suing so-called “AI” companies that are wrongly denying people access to essential resources or services.  Provide a modicum of accountability for an industry that lacks it, and restore a semblance of fairness to those people who play by the rules.  I’m starting Round A of angel investment funding now.  Who's in?

This is the latest of many examples where the engineers invent something cool (without adequate consideration of the potential negative consequences), and the bureaucrats blissfully mis-apply it as a cool cost-saving measure.  And we all pay the price.  Right about now, this is where the lawyers step in and save us from irresponsible deployment of an immature technology.  

Once my proposed solution is deployed (either by me or someone else), in about 10 years all of these AI biases will be exposed and repaired, resulting in automated decisions that are actually fairer than what a human would do.  

*** POSTSCRIPT March 14, 2023 ***


Also, read Part II discussing why Humans may soon need a label.

===
Next month's blog will go into little-known things about the new Sony A7R V, including an impressive characteristic of the Pixel-Shift mode (and why you shouldn't be too excited about the new Focus Bracket mode).

Until next time,
Yours Truly, Gary Friedman


P.S. - You can subscribe to Cameracraft magazine here. :-)

43 comments:

  1. Another problem: AI has been developed with the capability to mimic the voices of individual humans, and trolls are now using that to generate "recordings" of imaginary statements that damage the person who supposedly said them.

    ReplyDelete
    Replies
    1. Yes, that gets lumped into the category of "Deepfakes" which I linked to. It's getting to the point where you can't even tell when National Geographic moves a pyramid anymore!

      Delete
    2. Photographs don't lie, but liars do create photographs. The invention of paint put an end to the art of scrawling drawing on the walls of caves with bits of charcoal from the fire. We dodge, burn, double-expose, crop ... has there ever been an image that we could trust to be 100% authentic? Well, yes, but the fact remains that the image could have been manipulated. I really do not see this new method as a threat.

      Delete
    3. Presumably, there was actually a photo behind most Photoshot adjusted images. With AI, there's no photo at all, it seems.

      Delete
  2. You forgot to mention AI used for sentencing in criminal cases, which (you guessed it!) disproportionately assesses dark-skinned people as likely repeat offenders.

    ReplyDelete
    Replies
    1. And at least with these AI "black boxes" you can prove such bias, making an even stronger case for my proposed legal firm.

      Delete
    2. The American Bar Association's Model Rules of Professional Conduct specify in Rule 5.4 that nonlawyers cannot partner with or share legal fees with lawyers and cannot hold ownership interest in law firms.

      Delete
    3. Greg S. - I'm sure that's completely circumventable. The Electronic Frontier Foundation has an army of lawyers that carries out their mission; we can use the same model.

      Delete
    4. Gary, it is not circumventable to have a non-lawyer be an equity holder in a law firm. The Electronic Frontier Foundation is a non-profit that has "an army of lawyers" that are employed by it. There is no equity in it. It receives grants and donations to carry out its non-profit mission. I don't think that is the model you had in mind as you want to make an investment, which implies a return on equity. But I do see your point about what the next big thing might be.

      Delete
    5. SLT - thanks for clarifying. I'll probably not act on this idea because my entrepreneur days are behind me. But the idea needs to be out there!

      Delete
    6. I find it annoying that the system is set up so that only lawyers can profit from this sort of thing.

      Delete
  3. This is very interesting, because it is in line with something I am reasoning these days with the music I compose for fun, that is: I have now the best plugins I was able to afford, but my satisfaction is decreasing for a number of reasons.

    One of them is that I am not actually evolved as a “musician” and my improvements are not reached becoming more able to play an instrument, but only using the “ability” (samples) of real musicians, so what’s the point going on this way, compared to simply listen my huge record library?

    Hard to answer…

    P.S: The same could apply to photography, but I dislike any kind of PP, so I'll never use AI for that

    ReplyDelete
    Replies
    1. Sounds like you're worried about "imposter syndrome", which shouldn't trouble you because the compositions are yours. You just have an orchestra of plug-ins that you can conduct. I'd say they're enhancing your current abilities. If there's a silver lining here, it sounds like you now have an incentive to practice more and improve your playing skills. :-)

      Delete
    2. nope, it was rather a rant for the fact that AI doesn't help me to improve, because I am a bad player as before XD

      Delete
  4. I submit wildlife photos to iNaturalist. They us AI to guess the species in submitted photos. For my photos, they are almost never wrong and have gotten much better since they started using AI.

    ReplyDelete
  5. I was once acquainted with a person endowed with five fingers and a thumb on each hand!

    For your new business venture, I'm in with a contribution if you start a crowdfunding campaign on Kickstarter or similar.

    ReplyDelete
    Replies
    1. Thanks Andrew! I'm not sure how serious I am about doing this myself, but the need is definitely there.

      Delete
  6. The Turing Test has already been applied to current AI, and the test was passed. But clearly the AI is not actual intelligence but simply puts together clips from its vast memory. What researchers have concluded is that the Turing Test is not good enough. Or it needs to come with an instruction set.

    ReplyDelete
    Replies
    1. We put together "clips" from our vast memory, and most of us have difficulty correctly interpreting them. One thing we can (could) do that AI can't is express emotion, or feelings, and it can't interpret them either.

      Facts are what it can be good at, but it can't distinguish between genuine facts and "alternate" facts. That's why I suggested that the best first fit is to provide search results for in a curated corpus, like case law, medical journals, etc.

      Delete
    2. Don't say "can't" when talking about AI .... it only depends on the amount of data you present a ML algorithm. Bit where is the difference? ML systems cannot explain there decision processes..... they habe none. Symbolic AI like Expert Systems use decision algorithms and are able to explain how they concluded. ML can't because its either a regression function or a pattern matcher build on pure math, and Gödel proved already the no formal system, which math is can't prove itself. So be warned .... taking AI/ML results as something proven .

      Delete
  7. The first practical use will be for searching a curated collection of data in a narrow field. Right now we have keyword searches, fuzzy, like Google, or natural language like ChatGPT. Even now it can find legal cases dealing with complicated legal issues. Legal issues are a defined set, not just made up. Using the existing body of U.S. caselaw, treatises, law journals, etc. is perfect. Like keyword searching, results will have to be used as starting point for finding precedence.

    Even now, with a collection of basically garbage,mit can find legitimate cases dealing with an issue that is very difficult to find with keyword searching. I tried “find legal cases deailing with misinterpretation of a statute” and it found three that were remarkably on point.

    We tried for years at LexisNexis to do this with poor results. ChatGPT found three needles in a very large haystack.

    ReplyDelete
    Replies
    1. Thank you for balancing out this article with positive applications that are significantly better than existing tools.

      Delete
  8. Thanks for getting the Sony A7R V book out in record time! I'm about to 'jump' into this camera and will probably need the assistance your books have provided in the past.

    ReplyDelete
  9. I agree with you about ChatGPT. As a former writing teacher and current textbook editor who has spent a few hours with the text-generating tool, I believe any well educated person should be able to recognize the cool tone of the bot.
    Regarding DALL-E & Co., I have been impressed with very few of the images I've had them create (or stitch together, Frankenstein-wise). People and animals, as you suggested, usually end up terribly deformed. I like your suggestions for improvement.

    ReplyDelete
  10. Thanks for the A7RV book. As for the legal response for AI count me in with donation and support!

    ReplyDelete
  11. You ignored another, and very successful A.I. application: The revolutionary, and first use of an A.I. chip within a camera - - - Sony's A7RV. The chip greatly improves close-in focussing, and has modes for People, Animals, People & Animals, Insects, Birds. Yes it has a large bias against out-of-focus close-in subjects!

    ReplyDelete
    Replies
    1. Excellent point! As I said in the article, it's a double-edged sword. There are good uses and bad uses.

      Delete
  12. Frequent generation of hallucinations and outright lies by ChatGPT must be addressed, or this tech will be worse than useless. Examples abound of totally imaginary literature citations, quotes, and brain-dead "facts" presented with authority. The latest gem was an offshoot program which stated a fictional bird, the "black-headed hummingbird", eats bird seed like sparrows. Two preposterous errors in one example!

    ReplyDelete
    Replies
    1. What prompt generated those responses? Is it duplicatable?

      Delete
  13. ChatGPT tells me just now that there is no such bird as the "black-headed hummingbird". Apparently the closest is the black-chinned hummingbird, which eats nectar and small insects.

    ReplyDelete
    Replies
    1. ChatGPT is not good (right now) about specific facts because it is using a huge questionable corpus. It can’t even tell you when the Super Bowl will be because the Internet was crawled in 2021. Right now it’s responses are amazing fun, but they fall apart with factual requests. Those can only come with authoritative sources, and it will happen.

      Delete
    2. This story is changing daily. See https://gizmodo.com/google-chatgpt-ai-bing-microsoft-1850058443 and also https://techcrunch.com/2023/02/01/report-microsoft-plans-to-update-bing-with-a-faster-version-of-chatgpt-in-the-coming-weeks

      Delete
  14. the sky is falling the sky is falling. i hear so much about the negative aspects of AI. and very little coverage of the positive aspects of these AI tools. for example the open.ai product whisper you probably haven't heard about. open source tool that does a significantly better job of transcription and language translation than any other transcription tool i have tested - a great tool for film makers, journalists. i 100% agree with what you have to say about embedded invisible watermarks meta data.

    ReplyDelete
  15. Gary, when in the San Juan Islands (in Washington State) you have a place to stay & boat with.

    ReplyDelete
    Replies
    1. Thanks! I love it when people post offers like that anonymously. 😊

      Delete
    2. Wow! He probably doesn’t realize how beautiful it is there being from the east coast.

      Delete
  16. ChatGPT Passes Google Coding Interview for Level 3 Engineer With $183K Salary

    ReplyDelete
  17. ChatGPT Passes Google Coding Interview for Level 3 Engineer With $183K Salary !

    ReplyDelete
  18. I prompted Midjourney for a 'horse' and gave me one with 5 legs. An obvious mistake to us because we have rules about how many legs a quadrupeds are supposed to have, but obviously hard to teach ml or ai. Teaching these systems is likely to become very expensive as the technology progresses.

    ReplyDelete
    Replies
    1. https://openai.com/blog/new-ai-classifier-for-indicating-ai-written-text/

      Delete
    2. Yeah, I expect a plethora of ai-detection programs to surface in the coming months. My first question is "How accurate is it? Would someone's academic career be ruined by a false positive? Are there now two levels of problems to solve?" I'd also like to see similar tools for detecting deepfake videos too but haven't seen any that are reliable.

      Delete
  19. Dear Gary, from 1984 to ca. 1987 I worked as a comp. scientist on a European AI project (ESPRIT) trying to build an Expert System Builder.
    Then AI systems were all "rule based" primary based on LISP. The systems were big and SLOW and useless.
    This changed when high speed computer graphics chips were developed and you could "teach" them to recognize cats, dogs etc. But these systems are not "intelligent" they recognize dogs, cats and human faces because they have been exposed to millions of pictures of said items it's just simple pattern recognition and some skilled software people.
    Yours Gunni S. Frederiksen, Copenhagen

    ReplyDelete
  20. LISP!!! I remember that language, and the Symbolics machines on which they ran. As I mentioned in last month's blog post, I hate the term "AI" because all it is is Machine Learning (which, in turn, is just Pattern Recognition). I remember the big AI push in the '80's because everyone didn't want Japan's ambitions in the field to go unchallenged. Strange that only now have a gazillion commercial systems suddenly sprung into existence.

    ReplyDelete

Thank you for your comment! All comments must be approved by a moderator before they will appear.