Post No.: 0984
Furrywisepuppy says:
Machine learning algorithms can predict what you’re going to pay attention to better than you can because your own mind can be biased and poor at predicting what you’re really going to do or will get drawn into doing, as opposed to what you believe or say you’re going to do or want to do.
They can also figure all this out without needing to ask you – by simply collecting and processing your personal data and finding patterns from it. Taking an e-commerce platform algorithm for example – neither you nor the algorithm needs to truly understand you before the algorithm can categorise you according to a group of similar customers based on a particular set of metrics, and can predict your behaviours reasonably accurately based on how other customers like you have behaved (because we’re not as unique or unpredictable as we may think).
So these algorithms don’t predict your behaviours as an individual per se but according to the group of users you most closely fit into (e.g. aged 26-30, female, heterosexual, drives an all-electric car, shops mostly at a certain supermarket, just had a newborn…). It can use this information it knows about you to, perhaps, know what kinds of messages (e.g. appeal to your rapacity, highlight the plight of fluffy alpacas) will most likely best persuade you into performing an action that an advertiser desires (e.g. to vote a certain way that mightn’t be in your own long-term interests, to buy something you didn’t really need).
Algorithms can also be used to work out what correlates with a change in your own behaviour – similar to how you can potentially solve a problem without fully understanding that problem by simply finding out what changes correlate with a reduction or absence of the symptoms.
When trying to program computers with explicit rules about how humans were believed to behave – as in with rational decision-making processes – it failed. So scientists later learnt to just feed algorithms tons of data and let them find any patterns of human behaviour by themselves; agnostic to whether those patterns had any meaning or not. They discovered that this worked much better.
This doesn’t mean computers and algorithms cannot lead us astray – over-trusting them can still lead to immense catastrophes, like the 2007/2008 Financial Crisis. The systems relied upon in even the well-funded financial sector couldn’t foresee the absurdity of giving loans to people who’d likely never pay them back in time because of the sneaky way the industry invented a way to package tranches of AAA, AA, A, BBB, BB and B rated bonds and securities together to somehow ‘diversify’ risk in this manner and thus make such collections of loans overall magically hold AAA ratings(!) In truth, most of the humans working in the financial sector didn’t foresee this problem either, or didn’t care because they were personally raking it in at the time. But the colossal boom set up a colossal ticking time bomb. It could be said that over-trusting in these ‘rational models of human behaviour’ made the crisis worse when the bottom finally fell out.
If we can run an algorithm, that spots images of dogs, in reverse until it can more confidently perceive images of dogs everywhere, yet when we look at those images we can tell that they’re even less like images of real dogs – then it reveals a flaw in that algorithm’s assumptions (as demonstrated by the computer vision program DeepDream). Woof.
Artificial intelligences – like human intelligence – can be too crude and easily fooled. For example, a credit rating agency might employ an algorithm that crudely and erroneously judges someone who’s really good with and is on top of their finances for exploiting every bank-switching offer they can find as someone who’s really bad with their finances for constantly opening new bank accounts. So the good can be regarded as bad, and vice-versa, all because the algorithm relies on crude rules that most of the time work but sometimes fails, and fails dramatically.
Because they attempt to predict your future behaviour based on how others have historically acted i.e. others who seem most similar to you according to the particular dimensions they measure, the data they’ve managed to collect about you, and (quite critically) the accuracy of this data – they won’t work well every single time. That historical data may also be heavily biased (e.g. it was based chiefly on European males but you’re an African female).
Some shopping recommendation algorithms will keep recommending more trainers to you after you’ve literally just bought a pair from them, when you don’t personally collect trainers. Or they’re like ‘you’ve looked at these items before but haven’t yet ordered them, so you must be interested in them hence I’ll keep showing you them’ – when you’re not interested in them anymore and that’s precisely why you didn’t order them.
In all these kinds of contexts, these computer algorithms should technically be called ‘heuristics’ because they’re shortcut strategies that don’t guarantee a correct solution (e.g. when trying to suss out fake news, fake accounts or fake comments, with both false positives and false negatives). But most people refer to them as ‘algorithms’. In computer science, algorithms are technically sets of instructions that’ll always reach the correct answer. What varies is how efficiently they do so. Computer heuristics are imperfect, much like human intuitions.
It’s like when people use the word ‘theory’ in everyday contexts when they ought to really say ‘hypothesis’ because that’s what they really mean. Languages continually evolve and the meanings of words simply depend on what’s agreed but it can cause problems if different groups refer to different meanings.
Even though, broadly speaking, they’ve been advancing, refining and improving (quite rapidly) every year – AIs cannot foreseeably be the solution to perfectly reliably sorting out truths from lies because there’s no simple metric for finding truths. The truth isn’t measured by popularity or who says something, for instance, even though these are common heuristics. And regardless of whether algorithms care about truth or lie or if they could ever discern between them – the overarching goal set by most businesses that employ them is ‘find and do whatever makes us the most money’. As a result, if spreading lies generates a lot of revenue then it’d take a conscious effort by a company to try to override that (usually because of external pressure).
No training dataset will be perfectly bias-free unless it includes and accurately categorises every relevant piece of information in the world ever. Even all images on the web on a subject don’t mean all possible images on that subject ever. Also, if an algorithm is opaque in how it comes to its decisions, or corporations regard them as their trade secrets, then we’ll sometimes never be able to assess whether its decisions are being skewed. I guess these problems apply to human minds too though, and an AI just needs to be more reliable than a human to replace a human.
But because people tend to believe that computers are objectively reliable machines with no human-like fallibilities – miscarriages of justice like the UK Post Office Fujitsu Horizon IT system scandal can happen. Software makes no mistakes… except all those they’re inadvertently instructed or trained to do. (This system could be remotely tampered with anyway!) Computers will follow any fault in their programming to the very letter, and not always to the point of crashing but to the point of more subtle glitches we mightn’t immediately or easily notice.
Bugs that need patching are commonplace in software. Hardware can fail too. Computers are also finite machines operating with finite amounts of bits and bytes hence we cannot expect them to calculate sums to infinite numbers or infinite precision in terms of decimal places or significant figures (which can lead to overflow problems). Similar to how numbers like 1/3rd, or irrational numbers like pi, cannot be 100% fully and accurately represented under a decimal system – in binary, floating point numbers like 0.1 cannot be 100% fully and accurately represented either. In the vast majority of cases we don’t need this level of precision, but on the odd occasion these factors can lead to unexpected results. It’s still ultimately the fault of humans in their programming or their over-trust in computers without understanding their limitations however.
That IT system scandal also highlighted the pattern that most organisations will try their utmost to cover up anything that’ll give them negative PR, and how tremendously arduous and stressful it can be for the ordinary citizen to fight against injustices perpetrated by giant corporations; as if we all need to be like furry Erin Brockovich.
So can we trust computers to make decisions for us? Well in umpteen ways in our daily lives we already do. And who said we could always trust humans anyway?(!) It’s more about whether we can trust the humans who implement the AIs and where, what training data has been used to train them, and the transparency of how they come to their decisions? Well new technologies are often accepted by new generations of people who’ve been raised to not know any differently than having those technologies in their lives hence we’ll likely trust computers increasingly more.
Practically every product claims to include AI at present – from toothbrushes to doggy bowls, pillows to cat flaps – however each manufacturer interprets the term and whether usefully incorporated or not! AI has really already revolutionised the modern world under the surface for years – but with the latest generations of chatbots and robots, they look set to take centre stage and revolutionise the way we live and work quite overtly.
Governments are urging more kids into careers with computers. The UK government even produced an advert suggesting that a young ballerina should consider going into cyber, which received public backlash! (If only more people got into epidemiology because there was apparently only one UK-based research group that focused on novel coronaviruses before COVID-19 struck!) We actually need a diverse economy in order to be robust. And – as advancing AIs could spell the end of many current career paths (including coding jobs ironically) – professional ballerina is probably a safer career aspiration since robots are far from being able to dance as gracefully as humans can(!) AI isn’t going to take over jobs like plumber or hairdresser soon either.
Those who provide metaverse (persistent, shared, collective virtual spaces) services will also collect masses of data about our every move. Meta in particular is looking to dominate and control that space.
Rapid growth to grab market share is being prioritised over ethics and safety too when we see unregulated virtual reality chat rooms being toxic and harmful places for children. Some women worry about men pretending to be women entering female-only virtual spaces, like they worry regarding real-life private spaces like bathrooms. A company like Meta may argue that it doesn’t own the apps that run on its platform but it ultimately profits from those apps thus needs to take responsibility.
Interacting with other avatars in a visually 3D space in real time presents a crummy combination of fellow users still being dehumanised and thus easier to abuse compared to in real life, yet being on the receiving end of any abuse feels more real compared to just reading a written comment online.
In product design, logistics and urban planning – digital twins of physical objects or processes are often used for prototyping, modelling and simulating designs with the purpose of providing feedback to help improve their non-virtual counterparts. This idea could apply to humans too in the metaverse – a kind of digital avatar or virtual representation that serves as the real-time digital counterpart of you? Will people get to fully own their own digital twins and data about them though?
Woof. There are countless ethical hurdles to try to anticipate and resolve as computers and technology in general advances.
Comment on this post by replying to this tweet:
